July 3, 2025 • 57 mins
Ready for a great explanation of Agentic AI? For the last show at Build, Carl and Richard sit down with Seth Juarez to dig into what agentic AI really is - and how you can take advantage of it! Seth discusses the potential of MCP and NLWeb to enable agents to work with each other, as well as the challenges of managing these tools effectively. The conversation turns to what's happening under the hood of agentic AI software, including the limitations of its abilities. There is a need for governance and clear thinking with these new development tools!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
How'd you like to listen to dot net rocks with
no ads?
Speaker 2 (00:04):
Easy?
Speaker 1 (00:05):
Become a patron For just five dollars a month, you
get access to a private RSS feed where all the
shows have no ads. Twenty dollars a month will get
you that and a special dot net Rocks patron mug.
Sign up now at Patreon dot dot NetRocks dot com.
(00:34):
Hey guess what, it's dot net rocks from Bill. This
is the last show we were recording. I'm Carl Franklin.
We're here with Sef Juarez and we'll be talking to.
Speaker 2 (00:44):
Get that out of there. Get a little fam Sorry,
you gotta get it out. You know, we're a little punchy.
This is not only the last show of the day,
but it's the last show up. Oh, if you got
a punchy a show with a punchy dude, it's gonna
be great.
Speaker 1 (00:58):
It's gonna be You've been working hard, that's right. I
just want to start talking to Seth. Do we really
have to do the intro stuff? You should do the interesting?
Speaker 2 (01:06):
We gotta do it. You got to pay so you
know it's episode nineteen fifty seven. Wow, that's the year
I was born. It's not true. It could be in
a different universe very close. It could be in a
parallel universe. If you know quantum theory, the many worlds
as you know, Richard In posits that perhaps there's another
universe where I was born in nineteenfty seven. There's a
million universes. That happened, sixty seven year old set. I
(01:30):
just wanted to use the word posits in the thing.
And so now my job is I'm done. You must
have a master's degree.
Speaker 1 (01:37):
I have something. So do you want to do the
history first? You want to do I didn't do the history.
So what happened in nineteen fifty spot Nick one?
Speaker 2 (01:46):
Okay, the first the first artificial salarate over the world,
and then sput Nick two. They put a dog in
it and she died. No, actually no, she was in
the Guardians of the Galaxies. Right, wait, that's not real.
We want the historical documents and it's worse than any think.
(02:07):
She didn't she didn't die and re entry, she didn't suffocate.
She overheated. There wasn't temperature controls oh thanks.
Speaker 1 (02:15):
Sometime around here in the nineteen late nineteen fifties, electric
boat from General Dynamics was established in the Nautilus.
Speaker 2 (02:23):
That was no electric boat? Is that? Is that a
like a line dance? What is fifty five? Isn't it?
The electric boat? Is that like the electric slide? But no, no,
no waste? The nuclear submarine? Oh yeah, the first nuclear summer, theatlis.
Speaker 1 (02:37):
That's right, Okay, I'm sorry. I had a forward memory of.
Speaker 2 (02:42):
I also usually referenced a computer, and in nineteen fifty seven,
IBM released a compact computer, the six', ten only eight hundred,
pounds much smaller than the previous. Ones it really was
an improvement there you. GO i, mean if you want
to get, fit you get one of those commonract eight
hundred pounds. Computers also released the first compiler for. Fortrend hey, yo,
(03:03):
wow still running some code out. There there's as for
trend dot hold. On So i'm not. Kidding there are
some linear algebra operations that were written In fortrand that
people were, like let's just not do that. AGAIN i
think it's like some WEIRD langsos method for, computing like
singular value, decomposition that they're, like let's just not do that.
Again somebody did, that and we're, gonna we're, gonna we're
(03:26):
gonna honor them and we're just gonna leave that. Alone. Wow.
WOW i actually contributed this.
Speaker 1 (03:33):
SOMEHOW i actually remember taking A fortran. Class but it
was like in the. Eighties, yeah a little further, on yeah,
too it was like in THE i did fourtran.
Speaker 2 (03:44):
Three oh is there a sound? Effect weing? At BUT
i love? It all? Right, well, anyway anything else in
nineteen fifty there was a lot Of there was a
lot of. Things there's new clear. Test it's all kinds
of horrible stuff listed with, computers all, right and speaking
of atomic, fires what are we talking about? Today, well
(04:07):
we gotta finished the fish such a. Hurry i'm not
in a.
Speaker 1 (04:12):
Hurry, okay all, right let's do the better note of
framework roll the, music all. Right, so as you, Know
Jeff fritz AND I Jeff frits For microsoft And big
INTO ai And. Blazer we do This blazer puzzle.
Speaker 2 (04:34):
Show AND i was just.
Speaker 1 (04:35):
Talking to him about all of the cool THINGS i
saw at the, keynote including, you my friend oh hey
uh at the keynote it build and of course he's
hipped to all of this, stuff and we were kind
of talking about how and we talked about this with
some of our guests this. Week about how the prompts
are becoming code and instead of having code, libraries people
(04:57):
are going to start sharing prompt. Libraries Prompt so what
Did fritz point me? To he pointed me to sharp,
site which is all these prompts for Doing blazer applications
with the.
Speaker 2 (05:13):
Agent oh, interesting do they? Work?
Speaker 1 (05:15):
Yeah in good hub co pilot so code style and,
structure naming, conventions dot net specific, guidelines sho air handling and,
validation like all of these great prompts to help you
create Great blazer.
Speaker 2 (05:32):
Applications that's all the. Agent how about? THAT i think that's.
Great anything that we can do to further the work OF,
ai because you, Know i've BEEN i was like on
show like, SEVENTEEN i don't, know something like that eighteen
AND i was, like machine learning is the future AND
i was like What and we didn't understand. It then
we still, don't but at least it's. Working but it's,
true and you just still understand. It we just, believe you,
(05:55):
KNOW i really excited that people are, like let's use
this stuff right five seventy Six july of twenty. Ten
oh my, gosh aires on machines in eighteen years AND
i just get more. Beautiful i'm glad this is. True
in this, medium it's really hard to, know but it's for.
SURE i want you to imagine, it all.
Speaker 1 (06:14):
Right richard Save, us who's talking to us, today.
Speaker 2 (06:17):
Grabbed to comment off a show nineteen, nineteen which you
did in twenty twenty, four a little more recent In
october With Prascent boy here he's AN mvp DOING ai
related stuff that he wanted to talk, about and we
did talk About Copilot, studio WHICH i believe has now
been renamed because. Reasons, yeah And John tallon had this
great comments from a few months. Ago he, SAID i
enjoyed the. Show chat BASED ai apps are quick to stand,
(06:39):
up but tough to get right every time a man
rag dazzles that. First but yet it's limited without deep. Engineering.
Ooh in real, conversations not everything is a, question and
bought led chats are no. Different we must parse the user's,
intent evaluate the quality of their, input then apply a
chain of erication to fact checkbox. Responses i'm a practice
director in data AND II, ai And i've worked in Gen,
(07:01):
Ai AZURE Ai studio and Co Pilot, studio and despite
the Marketing Copilot studio isn't for your Quote jim From.
Accounts it's for seasoned consultants Like brashant who know how
to use the. Platform using it for complex, interactions like
a rocket launcher on a. Fly, yeah it, works but
it's a little over. Complicated rocket launcher on a. Fly it's,
(07:21):
like use. It, yeah let me say that. Again using
it for complex interactions is like using a rocket launcher
on a.
Speaker 1 (07:26):
Fly it's kind of taken a nod from my term
swatting a fly with a.
Speaker 2 (07:31):
Buick we had a term swatting a fly with a,
shotgun and my algorithms professor was, like we have an
algorithm for sorting that's faster than en log, in but
only with giant data. Sets it's like swatting a fly
with a. Shotgun, yes, yeah. Nice we should only explose
enough complexity to get the job. Done letting use His
harner's AI's potential without feeling. Overwhelmed after, all the goal
(07:53):
is to DEMOCRATIZE, ai not. Intimidate what year was? That
that was last? Week, yeah this is like a few months.
Ago the answer is you just, wait wait till next.
Years you. Know one of the things that CAME i
felt came out of build was this sense that nobody's
dominating in the space at. All, yeah there's lots of different,
models there's lots of different. Tools the fact that everybody
jumped ON mcp is. Clear nobody has the confidence to, go,
(08:15):
oh we'll go it. Alone we're all trying to find
standards in inter interapt with each. Other. Yeah BUT i
mean even EVEN mcp has. Limitations AND i have thoughts on,
that And i'm tired And i'm just gonna say WHAT i.
Think Oh, NO mtcp definitely has some. Issues there's no
two ways about. That but let me wrap up With john. Here,
john thanks you so much for your, comment and the
copy of Music coba is on its way to. You
and if you'd like a copy Of Music, COBE i
(08:36):
write a comment on the website at dot NetRocks dot
com or on the. Facebooks we publish every show, there
and if you comment there and every read it on the,
show we'll send you a copy Of Music. Koba And music.
Speaker 1 (08:44):
To Code by is still going. Strong twenty two tracks
to help you get in the zone and focus twenty
five minutes each to coincide with The pomodor, method and
they will help.
Speaker 2 (08:54):
YOU i actually bought the first. Ones yeah a long
times years.
Speaker 1 (08:57):
Ago, yeah it was only three AND i did A,
KICKSTARTER i did.
Speaker 2 (09:02):
THREE i got the. THREE i still have THE mp
three is somewhere in my junk drawer of one.
Speaker 1 (09:08):
Drive, well now you can get the whole collection IN
mp three flak en. Wave so you can go to
music To code by dot net for that or you,
know you could just send us a comment or leave a.
Speaker 2 (09:18):
Comment, yeah, absolutely and reading on the show you'll get
a copy of MUSIC o.
Speaker 1 (09:21):
Bye, yeah all, Right So i'm going to introduce.
Speaker 2 (09:23):
You for real areas here we go That i'm introducing.
Speaker 1 (09:27):
It, yeah that voice that you've been listening to the
can who it is Is Seph wares and he works
in THE ai platform group At. Microsoft he received his bachelor's.
Degree that's, it that's your. Title Now i'm going to
talk about your education because that's the part of your.
Bio we see his bachelors and computer science at you
n L v saying place that Guy fieri got his.
Speaker 2 (09:49):
Degree we were. Partners he's a he's a machine learning master.
Speaker 1 (09:54):
Absolutely with a minor in mathematics and complete his master's
great at The university Of utah and the field of
computer science And. SETH i have a picture that BECAUSE
i was in the front row at the build AND
i took a picture of.
Speaker 2 (10:09):
You oh what WAS i? Doing? Well you you opened the.
Speaker 1 (10:12):
Show you came out and, SAID i Am seth and you,
know with with somebody, else AND i DON'T i didn't recognize.
Her BUT i took a picture of you two and
here it. Is oh my, gosh That's. Kadasha, yeah she
is delightful human. Beings SO i sent this To, kelly my,
wife you, know, yes of course she actually met you
BEFORE i met.
Speaker 2 (10:30):
You, yes she did, so and, she.
Speaker 1 (10:33):
You, know she says back to me and she, says
that's a great. Picture but who is that black lady
With Ricky?
Speaker 2 (10:39):
Martin but, WOW i do get more. Beautiful AND i, said,
again this is the medium for me to say these
kinds of. THINGS i am getting more. Beautiful look at the.
Speaker 1 (10:53):
Picture your name is not on the, screen it's it's her. Last,
yeah so she wouldn't know and she didn't recognize.
Speaker 2 (11:00):
You that is so. Funny then WHEN.
Speaker 1 (11:02):
I told, her she, says, oh, yes that is.
Speaker 2 (11:05):
YEAH i actually owe her a. Call she called me
the other day AND i haven't called. Back, oh, okay she's,
delightful she. Called she checks up on. Me she does. That, yeah,
yeah so. Nice well you do need. Supervision it is.
True my wife has had me now for twenty two ish,
years and so during THE covid she may be. Tired
during THE.
Speaker 1 (11:24):
Covid, times she reached out to friends of mine and
hers that she hadn't talked to in years and, Said,
hey you know, What i'm just calling BECAUSE i want
to check.
Speaker 2 (11:32):
In it was.
Speaker 1 (11:34):
GREAT i don't want, anything you, KNOW i just wanted to, Say, Hi.
Speaker 2 (11:37):
GOSH i wish the niceness would rub off on other. People,
yeah it's not working. Right all.
Speaker 1 (11:45):
Right so we're talking about agentic, hey AND i got
to tell, You i'm blown away by WHAT i saw.
HERE i, MEAN i know that there have been other
agents and products that work with agency and all that,
stuff and the guys From manet have been into, it,
right but they all said the same. Thing it's, Like,
Wow microsoft is kind of catching up to what these
(12:07):
guys have been. Doing AND i hope they dominate because
they have the, power, Right microsoft has the power Of
azure and all of these. Things and WHEN i learned ABOUT,
mcp we've been talking about anybody who's listened to the
last six, shows oh, yeah knows about this about the,
(12:28):
okay is it.
Speaker 2 (12:29):
Model model context context.
Speaker 1 (12:31):
Protocol so this is the protocol by which agents can
talk to your resources to do things like talk to
YOUR sql, server talk To azure resources and all that.
Stuff and you can just tell it to go do, things,
yes and it will figure them.
Speaker 2 (12:44):
Out.
Speaker 1 (12:44):
Yes the combination of AN llm like get Hub copilot
AND mcp connections to other resources just blows my. Mind
it changes the.
Speaker 2 (12:54):
Game, Yeah i'd love to get more low level, though
so folks have an understanding of what an agent actually
is because there is This what happened was and you
were there with me for fifteen. Years i've been in
this space for fifteen. Years nobody, cared and then we
flip from nobody cared to this stuff is like jungle,
(13:17):
magic or this stuff is like going to throw you
into outer space or.
Speaker 1 (13:20):
So we didn't have to create the. Models, yeah that
was the hard.
Speaker 2 (13:23):
Thing it's. True but but what happened is we kind
of missed the in between part Of, hey let's explain
what these things are doing and why it's actually. Useful
because there's a THERE'S i don't want. To as you,
know we had SEVERAL ai winters because of the over,
hype AND i AM i am concerned that if we
(13:43):
overhype too much that there's going to be a correction
when there really doesn't need to be because it actually
this time does do qualitatively cool. Stuff, yes and that's
What i'm worried. ABOUT i, mean the winters were mostly
induced by the fact that there was a limited utility
what was being made, exactly and there was a limited
of computational power to actually make the model the thing
(14:05):
we wanted them. To hinton's famous papers actually the paper
talking about backpropagation SAYING i can't do with this equipment
in like nineteen eighty four mm. Hmmm and then In
suskovar does off that paper in two thousand and nine
and they start working on the image net. Problem, yes
and those models are quite simple as compared to the
ones that are now FOR. LMS i, MEAN i can
(14:25):
get into the architecture and stuff of. It they're just much.
Bigger and the thing about the thing about neural networks
of that size Is i've spoken To Turing award winners about,
like how do you shape these actual? Models to do
the things you, Want and the reality is we don't
really have a good understanding or. Explanation but what happens
(14:47):
is they're innovations that come out like a transformer model
with attention that was really designed to do translation that,
somehow when you made it, bigger learned language, well presented
language in a form that you could anthropomorphize. Correct and
that's that's the Part richard does not like at. All
(15:10):
the thing about it is is WHEN i picture these
things in my, head it's like the matrix for, me
Where i'm picturing like there's, weights multiplying vectors and then
going over some kind of smoothing function to, remove you,
know the linearities so that these scenes can predict complex.
Things but in the, end all these things are doing
(15:30):
is you're getting a string of words that are broken
up into vectors that get passed in and multiplied by
the transformer model to figure out how these things are,
related pushed through a giant neural network multiple, times and
the output is a giant vector of the entire space
with a probably distribution that says which token should you
(15:52):
put out? Next so when it looks like it's, typing
it isn't it's just going back through the, model and
that's what these things are. Doing and as, programmers the
only thing we can control is what goes in the
front the. Prompt and that's the beefy, bit right that
people don't. Understand at that, point when we first started
(16:13):
doing this, stuff it was magical for. Chat, Yeah but
then early, on about maybe like six months into, it
folks were starting to use these things for something that
we liked to THAT i like to call intent. Mapping
Cassie brevy AND i worked together for a couple of
years and that's what we called, it where we would
make THE lm like look at a block of text and,
(16:35):
say is this question a support or a sales? Question
and it was at that point that these, things because
we allowed them to decide is it a support or
a sales? Question when the code, branched those things all
of a sudden were imbued with this notion OF i
can make a CHOICE lm on what this thing. Is
(16:57):
and at that point those things became, agents, yes and
and and because we trusted them to answer our. Questions. Yeah.
Yeah but but but nottice that WHEN i was doing this,
ENGINEERING i narrowed the task to very specific you will
say either support or, sales and that's. It and what
(17:19):
that did is it allowed me to create a prompt
that made it so that any user input that we
put in, it it would tell me support or sales every.
Time and my code had an if statement if support do,
this if sales do, this, right and that at that,
point that is the most simplistic way of thinking about an.
Agent when AN lm has the has the ability to
(17:42):
choose reason overset of data and then act on the
reasoning you've. Generated and effectively that became a squishy programming
control structure and that's what an agent is. Fundamentally but you're,
like but it's got to be more than. That seth
what about the tools? Stuff at that people are, like,
well let's do structured output so that it could output adjacent.
(18:04):
Thing and if we can output adjacent, thing why can't
we output adjacent Thing that says this function with these.
Parameters and at that point those things became tools and
now instead of supporter, sales it, said run this function
with these. Parameters and that is the extent of all
(18:25):
of the, magic sauce of all of anything.
Speaker 1 (18:28):
Agentic NOW i want to say something Here, seth you
are among some of the most brilliant people That i've ever.
Speaker 2 (18:35):
Met, oh thank.
Speaker 1 (18:36):
You richard is one of. Them you are one of.
Them but the other people are like dev express. People,
Yeah Mark miller And Richard morris like these guys and. You,
NOW i want to tell the listeners if you don't
know Who seth. Is, seth you had an open SOURCE
AI mL library Before microsoft hired. You, yeah before they
(18:57):
Had AZURE. Ai you had this open source. Library we
did a thing about it on D NR tv that
showed it to.
Speaker 2 (19:03):
ME i didn't understand. It, yeah of COURSE i kind of.
Did but but you were doing.
Speaker 1 (19:08):
That you've been doing this longer Than microsoft has been doing.
Speaker 2 (19:12):
It way before it was. COOL i was doing before
it was. Cool the secret, sauce, though is machine learning
has been happening since nineteen fifty. Six Frank rosenblatt The,
perceptron which is the basic foundation OF.
Speaker 1 (19:25):
Ski so that was the last, show righteen Fifty.
Speaker 2 (19:28):
Oh, hey, yeah let's call back To Yesterye, frank we've missed.
Now but the reality is That microsoft's been DOING ai
in the form of machine learning for a very long.
Time WHAT i was doing IS i was trying to
bring it to the dot net community BECAUSE i recognize
that there is a space for algorithms which are learned
(19:49):
instead of, programmed AND i thought that IF i could
do a. Thing OBVIOUSLY i haven't kept up on that
for a while because it was just. Me BUT i
implement a whole set of stuff to make that. Easier
but my goal has been for the last you, know
fifteen twenty, YEARS i want people to know exactly what
these things are, doing so that with, confidence you can
make them do even cooler.
Speaker 1 (20:09):
Things, right, yeah, Yeah and so now you're At, microsoft
you're working on THE ai.
Speaker 2 (20:15):
Stuff that's The azure. Stuff that's. Right The Italians channel,
Nine oh, yeah, No channel nine was. Great you had
such a good. TIME i really. GOOD i love. That
and then somebody figured out actually knew them all, stuff
and it was. Important it's, like you come with, me you,
know CAN i be? HONEST i ACTUALLY I actually it
was different than that because AS i started doing The
channel nine, stuff the reason WHY i knew what questions
(20:36):
asked is BECAUSE i already knew their content AND i
knew how to make them talk about and make things.
Sing but people just started saying stuff to me, Like,
seth don't you ever get tired of all this technical?
Stuff And i'm, like oh, man MAYBE i should move
into something more. Technical maybe. Maybe so THEN i moved
into DevRel FOR aiml and Then John, montgomery who was
(20:57):
A cvp Of young he was, Like, seth come work
for me in THE ai core.
Speaker 1 (21:03):
Group so that's what you obviously knew your, pedigree, right
they knew they.
Speaker 2 (21:07):
Did nobody, knew LIKE i THINK i think people record
like see me and like Because i'm pretty jovial and
and BUT I i.
Speaker 1 (21:16):
Think you don't come across as AN ai.
Speaker 2 (21:18):
Scientist, NO i KNOW i do. Not and but if
you looked at, me obviously we've said this many, Times
i'm very beautiful in, appearance and people would not Like mark.
Exactly people wouldn't, know BUT I i really have been
doing machine learning. Forever and, again my primary goal is
for people to understand at a profound level what these things.
Are and to, me they are basically a new programming
(21:41):
control structure which can Take english and point you to
what to run. Next, yeah fairly still Constrained, english, yes, well,
no even Regular english with AN, lm because if you
can you can type into AN ll m like Bad.
English and for, EXAMPLE i did a demo at my
talk two days ago WHERE i showed him the supporter
(22:01):
sales thing AND i, said my tent is, broken, yo
AND i spelled it wrong and it said support AND
i was, like what's the borst tent in the, world?
Yo and it was like burst and not b e
r S t BECAUSE i fat fingered it. Right and
then it was, like, uh. Sales notice that because it's
(22:21):
a large language, model it's able to reason over the
text and then create the output that we. Want and
if that output is a, function then it tells you
what to run. Next so it.
Speaker 1 (22:33):
Didn't understand, bursts so it figured that you really meant worst.
Speaker 2 (22:36):
A best, Best, yeah, okay best b e r S.
T and it was, like even if it was, worst
it's still a sales and it still was able to
figure it. Out and and that was with me, again
and THEN i. CAN'T i don't know how to emphasize
this is. Enough, basically you get a control structure that
gives You english and tells you what things run. Next
(22:59):
WHAT mcp does and whether it does it well or
not is neither here nor. THERE i don't think it's
Necessary what it does is it allows you to query
what functions can you? Call and what it does is
it maps those into the tool calls that are available
to the L, m and then it can call. It
(23:19):
but you can do that With swagger AND. Rest, sure,
yeah but it's.
Speaker 1 (23:23):
Not it doesn't have the the almost it does.
Speaker 2 (23:27):
Reasoning it, does but it does because the reasoning happens
in THE. Lm and what it outputs is it outputs
what functions Are swagger AND rest don't have that they,
don't but you don't need it BECAUSE, mcp if you
look at, it is basically Like. Swagger it tells you
what tools are, available and it's LIKE, rest and it
tells you how to call. Them, Right and the thing
THAT i like about like those kind of and OPEN,
(23:47):
api which has been NOW i think the Next swagger is, that,
RIGHT i don't. Know but with OPEN api and WITH
rest you're able to do the same things but with
the benefit of like twenty years of security, handling with,
headers WITH, ooth with token, negotiation et. Cetera and so that's,
why like IF mcp gets you, going like WHEN i
(24:10):
USE vb, six, right it got me. GOING i loved
it because it was able to. Get but then when
you were, LIKE i need to be more memory, CONSCIOUS
i needed to be more. Careful then you can skip
down a layer if you if you want. To but,
AGAIN i love THAT mcp has gotten people thinking about.
It but the reasoning doesn't happen WITH. Mcp, no it's,
(24:30):
yeah it just tells you what function are. Next but
the magic that makes it look good is once it
runs the, function there's a thread that has like what
the QUESTION i? Asked it asks for a Function i'm
like pointing like people can. See and then when it executes,
out it puts that function return call back on the
thread and THE lm is called again with that on its,
thread and then it comes with the. Answer and so
(24:51):
that actually going out from THE ai space is effectively
bridging the gap FROM ai reasoning to perative or functional. Code.
Right how we MAKE sky? IT i get? It, NO
i mean it's NOT i don't even think IT'S.
Speaker 1 (25:05):
Sky and it may not be the best security wise right,
now but it's clearly where everybody wants to.
Speaker 2 (25:11):
Go, yeah it's because it simplifies. It but my sense
is that as you see, This i'm the prognostica, here,
right BECAUSE i like doing. That as you see this thing,
maturing it's going to be indistinguishable from rest right and discovery.
Protocols that it's the way the embracement OF mcp says to,
me everybody wants to interrupt with everybody. ELSE i fact
that we could come up with a better way to do,
(25:32):
it and in fact already. Have we could just be
publishing APIs for crying. Out, Yeah And i'm telling, you
like we we had a phase At microsoft and bless
Our hearts where we went full. Wisdole, yes and.
Speaker 1 (25:45):
That's what The southern people say about the town drunk
bless his.
Speaker 2 (25:48):
Heart, yeah bless our. Heart But wisdole was the same,
thing the same kind of desire to to abstract over.
Primitives and then we went back to primitive still being
WSD l Or Web Service Discovery language. Wisdom does that sound? Familiar,
friends done done? Dumb and so What i'm telling you
(26:09):
is that and some of the things you saw with
N a Web the reason why this is such a cool.
Speaker 1 (26:13):
Space if we haven't talked about AN a web on the,
show oh, really yeah this WEEK i haven't.
Speaker 2 (26:18):
EITHER i started looking at. It but the intuition here
is if you can language do stuff and then call
stuff and then language do stuff on the, Out like
what you if you hook this up to the in
language tool, stuff, Right you're? Fired, yeah you're my problem
WITH nl. WEBS i keep, thinking why are you talking About?
Dutch people's confused The dutch. Web but hopefully that helps
(26:41):
BECAUSE i really want to emphasize that once you know
that that's what it's doing at a low level and
you play with it that, way it changes the way
you think about These. SURE i had a journalist talk
to me and, say, Hey, SETH i get a lot
of questions about SHOULD i trust these? Things and my
response surprised. HIM i, SAID i don't trust these. Things,
no don't be. Silly that's WHY i program them the
(27:02):
WAY i. DO i restrict what they have to. DO
i restrict the tool calls that they, have and then
when the tool calls are, CALLED i write the code
that the tool that what the tool responds. With because
all these things are effectively becoming part of the prompt
for the next tokens to. Generate and that is when
(27:23):
you get that, Intuition you're, like, oh that MEANS i
can do this ANYWAY i want to ye.
Speaker 1 (27:30):
And that's that's WHERE i think we've been talking about
this this week, too that the attributes of the program
or the future are going to be imaginations over technical
probsolutely imagination mean and we may be held back a
little bit by what we think we can do and
what we can't. Do but somebody who doesn't have that
(27:52):
kind of you, know, limitation self, limitation will be able
to do amazing.
Speaker 2 (27:58):
Things. Yeah like if you watch the keynote on day,
two you Saw Amanda foster And elijah on stage talking
to this thing that it was executing a bunch of
stuff on their. BEHALF i think it was. Software it
was literally software BECAUSE i KNOW i, WROTE i wrote
THE i wrote the website and the back. End amanda
worked on the, agent And elijah worked on the on the.
(28:19):
Evaluations but the thing about it is it was actually running.
Live that thing was actually happening. Live and one of
the things that we had is once you get into
a single, agent it makes, sense but when you get
into multi, agent you wonder, like, well what is the
interface between one agent and? Another and the surprising thing
is that the interface is actually language, weird which is, weird,
(28:42):
Right so for, Example i'll give you, example the voice
agent that we, made and if you watch the, keynote
you'll see her talking to. It the voice agent is
using THE gpt four oh Real TIME, api which has
the also has the ability to issue function. Calls, Okay
and so what we did is we collected the list
of azure Ai foundry agents that we made Andree foundry
AND i just loaded them up and then she picked
(29:04):
the ones she wanted because she was working on. Them
and what it did is it mapped those other agents
as tool calls interesting to the voice. Agent but you're, like,
well what did you pass to the? Agent, well the
agents themselves when you run, them have this property called additional,
instructions which means you can augment their system prompt and
(29:24):
then that's. It that's all you can. Do so WHAT
i did IS i mapped all of those agents as tool,
calls gave it a function. Name you're back to sales
and support, again that's all this, IS i. Know and
THEN i, said you passed voice agent additional instructions and
ask it for what you wanted to. Do so the
voice agent then is taking the initial requests and assessing
(29:47):
it to go which agent to go, to back to
your sales and. Support that's. Right then taking the portion
of the requests that makes sense for that, agent plus
the additional, instruction let's send it on. Down but it's
even more than. That it actually asks, it can you
help me with? This and we were me And amanda
because it's like in the two in the, morning we're
looking at this, stuff like because we're on this stuff all.
Night we looked at the law because we Ad jason
(30:09):
printed out everything and we could, see, uh image agent system,
instructions you are to generate a blah blah blah. Question
can you generate an? Image it was Actual english and
we were, like holy, cow this is this is amazing
and the software being polite to. Software, yes it.
Speaker 1 (30:26):
Was really cook but but it's nothing. New if you
look at the S MT smtp protocol when you connect
When i'm When gmail connects to an S mvp, server it,
says you, know, hello And I'm Carl, Franklin carl frank
and THE smtp server actually writes, back, Hello Carl, franklin
(30:47):
nice to meet.
Speaker 2 (30:48):
You, yes but we're not talking part of the. Protocol
it's Not eliza in This, NO i, KNOW.
Speaker 1 (30:53):
I, know but but you know that those kind of
friendly things have been.
Speaker 2 (30:57):
Protocols for a long. Time but the difference is that
the only things are being generated by the through the
agency of AN. LLM i get. It it's, big big.
CHANGE i do question if if if the politeness affects
the prompt. Results, yes it absolutely. Does And i'll tell.
You i'll tell you what happens And i'll tell. You
(31:19):
so there was a part of the demo Where amanda
generates an image for a hackathon and they AND i
don't think they understood because someone backstage is, like, hey
can you change the image to something more? Happy and we're, like,
no it generates it every. Time it's it's doing it
on its. Own and a man is, like but we
can change it by telling, it please make the image more,
(31:42):
happy and like we made that change in The english
and the and not in the voice controller because that's
the thing we talked, to but in the literal little
agent the, customer and it called, it called. It and
then it was, like and now everyone's happy at the.
Hackathon and and because everyone thought we were using some
like pre baked image and, no it's not every, time
(32:05):
and then we did, that and then it was really cool.
Speaker 1 (32:07):
All, RIGHT i think we got to take a, break
so we will be right back after these very important
messages don't go. Away did you know there's a dot
net on aws. Community follow the social media, blogs YouTube
influencers and open source projects and add your own. Voice
get plugged into the dot net on aws community at
(32:29):
aws Dot amazon dot, com slash dot, net.
Speaker 2 (32:36):
And we're.
Speaker 1 (32:36):
Back it's dot net. Rocks I'm Carl, Franklin I'm Richard,
campbell And Seth warez is. Here we are JUST i
feel like we could be talking for, hours hours an,
hour so much to talk, about but we're talking about
agentic coding and. PROGRAMMING i Took microsoft Agent mode In
Visual studio for a spin yesterday to go.
Speaker 2 (32:57):
Good, okay it went really.
Speaker 1 (32:58):
WELL i gave it some tasked to create some stuff
THAT i knew how to do BECAUSE i wanted to.
See but there was one very tricky thing that it
was intuitive for it to, do but it didn't. Work
so it's not like agent mode is going to compile
your code.
Speaker 2 (33:17):
And see if it.
Speaker 1 (33:18):
Works it's just going to do what it thinks is. Right, yes,
SORRY i didn't meet us the ord. Thing it's going
to use what it assesses to be the correct. Answer
so but WHEN i went and told, it, no that's.
Wrong this doesn't work on this particular. Object you're going
to have to subclass, this it, said oh. Okay AND
i didn't tell it what to. Do ALL i had
(33:38):
to say was you're going to have to subclass this,
subject and it figured it.
Speaker 2 (33:42):
Out it's pretty cool. Stuff one of the things that
you're going to find is that if you ever have
an agentic thing that is not quite doing what you
want in, general from an engineering, perceptive for, me it
means that they've overstuffed what they want a single agent to.
Do so the actual engineering problem becomes agent specificity versus agent.
(34:06):
Count and that's an engineering, issue, right because if you
have too many, agents it might cost, more or it
might it might work. Slower if you have too few,
agents the recall isn't going to be as, good, Right
and so the actual engineering problem is that it turns
out the agent you use was. Able once you specify
down a little, bit it made things. Better but if
you think you're going to make a single agent with
(34:28):
like a two page prompt with fifteen, tools it's not
going to do the right. Thing and that's the trade
off from an engineering pler because now that you know
you have to make it, specific you have to limit
its tool because it's an interesting switch from what build
up lms the first, place which is making them bigger and.
Bigger now we're, going, hey if you wanted me to respond.
Better but see that's the thing, small that's a. Thing
(34:49):
it's not THE lms that are getting, right it's the
tasks that you're assigning THE. Lm so it's THE. Lm
let's just fix THE lm for. One let's just SAY gbd. Four,
oh like that models, Huge but like you are gonna
narrow the agent's task to your job is to send
an email to the right. Place and then here are
the tools send email or delete email or. Whatever those
(35:12):
are the only. Things that thing is gonna be very
good at sending, emails and it's gonna be very good
to be, like, HEY i need to send an emails
to whom please type or tell me the. Email and
then it's gonna be, like please validate the. Email what
content do you? Want because it's trying to fill out
the parametery of the function call because that's what the
prompt wants it to. Do and that's what's. Happening so
(35:35):
for your, case LIKE i worry like even in some
of the tools that we, make maybe At, microsoft maybe
they're too broad and that's where you find this notion of.
Hallucination now people aren't gonna like, this but the elem's
job is to actually make up. Tokens if you give
it too broad a, thing it's gonna make stuff. Up
if you get it to if you give it a narrow,
(35:56):
thing it's gonna make the right stuff. Up BUT i
thought that's.
Speaker 1 (35:59):
What the point of the agent THAT i access in
visual studio, is and it's NOT gpd four oh with
this broad uh set of knowledge or. Whatever i'm, Sorry,
RICHARD i didn't mean to use the. KNOWLEDGE i have
to apologize every TIME i answerpomorphize than LL.
Speaker 2 (36:14):
M but he is. Right thank you keep him on. Track, yeah,
No i'm. SORRY i apologize now ALREADY i. Apologize please
be with the pigmies In New. Guinea. Yeah.
Speaker 1 (36:23):
Amen but it narrows the focus to truss the things
that are in. Code so theoretically it should know everything
about the Language i'm using the, markup you, know If
i'm Using blazer or. Whatever so it's got to UNDERSTAND
html and JavaScript and all Those you're.
Speaker 2 (36:39):
Gonna love this. Answer it knows. Nothing it knows only
what You i'm trained on. It, well, yeah so there
there is a, yeah so that's WHAT i. Mean, Yeah
so there's certain models that are actually Like, codex for,
example is perfectly trained on, code but it doesn't know
like in my, head it doesn't know the. Languages it
knows that if you give it this sequence of programming or,
(37:00):
questions it's going to return this. Sequence let me just
add this to the.
Speaker 1 (37:04):
EQUATION i did the same problem with chatchipt AND i
did get the right. Answer, yes, so but the thing
that was The Microsoft agent mode In Visual studio did
not give me the right answer.
Speaker 2 (37:17):
Because it's a different.
Speaker 1 (37:17):
Model, Right but wouldn't it be narrowed down to just
the programming stuff like why did CHAT gpt with vast
programming and learning that it did on all this data
provide a better answer.
Speaker 2 (37:30):
Because models are stochastic, processes and the thing about one
model versus another is one will do better with the
same prompt or. Not and so that's when he gets
the college. Words that's five dollars five dollar. Word it's
a stochastic process because the output vector gives you the
probability distribution and samples according to that probably. Distribution so some,
(37:55):
models depending on let's just say that use clawed in that,
one or they are going to do different things and
it all so the promptly makes a, difference and the
tools that it has really makes a. Difference and that's
where it sounds like an excuse to. Me it does
and here's what they should have. Done just like we
have unit tests for, code we should have quote unit
(38:15):
tests for. Agents and those things are called. Evaluations and
so An azree foundery we have a bunch of evaluations for,
example is it follow the right? Task is it? Coherent
is it? Relevant is it? Grounded we also have these
other agents whose job is to attack our. Agents they're
called the red teaming, agent and its job is to
(38:37):
literally send garbage at it to see how it. Does
and so that is the unit. Test and the difference
between a unit tests and an evaluation is that an
evaluation you need to give it enough examples to be
satisfied that you're going to be. Okay it needs to
be probabilistically, relevant, Yes and then once you're okay with,
it when you put it in, production you monitor it
(38:59):
with the same evails and send alerts when it goes.
Off the rails diverts because when you're looking, at for,
example regulated, industries regular, industries like if you make a,
mistake you got to report, It but they also want
to know how do you fix these? Things and that's
probably more, important and that's the part of regulated. Industries
when you use this s uff you can see where
it went wrong and you fix it and then you go.
Speaker 1 (39:20):
Back SO i think the real answer, is, hey we
just launched this, shit give us some time.
Speaker 2 (39:26):
Time i'm not part of that. Team they mean they
could have made it, better.
Speaker 1 (39:29):
Maybe BUT i mean it literally just came, out like you,
know so and.
Speaker 2 (39:34):
They'll learn a lot, because, like for, example the question
you asked it maybe something they never. Anticipated that's, Right,
yes so there's Learn so there's some alert that probably
fired during THE evls that's, saying, Oh carl did not
get this thing that he, wanted and so they update
the prompt or a new tools AND i expect.
Speaker 1 (39:50):
That. Yeah, so BUT i love the, idea and you,
know talking To Scott hunter and some other, people the
possibilities of things That i'm to do with it in
the real world from my customers is going to blow their.
Speaker 2 (40:06):
Minds, Yes and in terms of like the stuff you can,
do it's the ceiling is very high because the biggest
barrier to programming isn't like the new algorithms or the
new it's like literally people using this stuff and having
the like like someone Putting bob in the age. Field you,
(40:26):
know that's the kind of dumb stuff that we write
most of our code. For right ll ms together in
an agentic, mode are going to soften that so that
we don't have like why do we need a web form?
ANYMORE i can do like my name Is seth and
you can use my current location to find the.
Speaker 1 (40:43):
Address, now you're absolutely, right and YOU i is going to,
change yeah from you. Know but but users are trained on,
forms aren't. They, yes you're going to have to train
a user to all. Right so here's here's a. Nightmarebox nightmare.
Scenario a textbox to tell the application what you, want
and somebody comes in and, says, hey you want A turkey.
(41:05):
Club i'm going to The. Delhi, Yeah i'll have A turkey.
Club and that's Writing turkey, club you know in the
in the. Thing you know WHAT i mean, like there's there's.
You now you have to train the users as to
what to say or what to, type and that's the.
Speaker 2 (41:19):
Beauty. THOUGH i think given a properly built, agent that
turkey sandwich is just gonna be noise to. It like
it's gonna look for the things that will force the
linear algebra and the calculus to produce the right next.
Speaker 1 (41:36):
Time, yeah and then there's gotta be validation and. Verification
did you really mean to say? This is this what
you really?
Speaker 2 (41:41):
Want you? Know and ALSO i think the software is
fairly good at recognizing a contact shift and SAYING i
don't THINK i was supposed to listen to, that, yeah.
Speaker 1 (41:49):
Right and Also i'm assuming that you're speaking to it
right when you really need to be, Typing like you
don't Want bob From engineering to pop his head in and,
say delete all records from, customers you, know.
Speaker 2 (42:01):
DUDE i saw AN mcp server that was attached to
a file, system AND i was, like maybe let's not do. That,
yeah it seems like about on mind's. IDEA i remember
when the you, know the First amazon devices showed, up
and right away it was, like send my porn collection
to my mother.
Speaker 1 (42:15):
Joke oh my, gosh send one hundred pounds of cement
To Richard.
Speaker 2 (42:19):
Campbell oh my, gosh so. Dumb but, yes BUT i
think LIKE i, Think LIKE i said. It it was
surprising to me how the the controller agent in this
case was communicating with other. Agents, yes that. Was and
then you, Know i'm not surprised by this stuff BECAUSE
i know how it. Works, Right but that was the
(42:41):
part that was, Like, okay maybe there's something. Here, well
and we get back to this whole like The battle
Of VISIBLE. Ai, yeah, right and we still haven't really
fixed in there all that. Problem, no there's no way
we can fix. That, well and that the Whole i've
read the anthropic paper on the reasoning models and what's
actually going on. There it's calling itself right couple. Times,
well first it gets the, results and then it peaks
(43:03):
pieces of the results and feeds it back into it
to say why WOULD i have said? This so it
literally is in reverse. Reasoning i've already decided what the answer,
is but let me break it down into pieces so
it looks Like i'm thinking about. It, yeah that doesn't
seem to be a good way to do, it but
it makes people, happy, Right, well it doesn't make me,
happy well until you know what's going. On and it's
LIKE i just.
Speaker 1 (43:20):
Read something in the news today that said THE ceo
Of anthropic admitted that he has no idea how this
stuff for but.
Speaker 2 (43:28):
You, know maybe like in his, defense like there's certain
things that CEOs do THAT i just don't want to
know how to. Do, yeah you know What i'm, Saying
like there's only there's only so MANY i don't know
if this is, true but there's only so much space
in my brain for things, nice, Right and it's LIKE
i have, seen like BECAUSE i work with a lot
of our EVPs and CEOs and and all that, stuff
(43:50):
and like And i've worked And i've talked to CEOs
and cvps and it's just like MAYBE i don't want
to do. That, Yeah and and the fact that he
doesn't know exactly how it works is, okay focus on
the right. Things. Yeah and the other thing that for,
me LIKE i mathematically know how these things, work BUT
(44:10):
i don't expect other people to just to. Me it's
just LIKE i love looking at like the beauty of
the mathematical structure of these things and trying to think
about why does this actually?
Speaker 1 (44:21):
Work but you know, WHAT i feel better knowing that
as long AS sef.
Speaker 2 (44:25):
Knows BUT i think this idea that that neural net
is is essentially creating a jpeg of, language lossy compression
of language one hundred. Percent that is exactly. Right but
when we, say and the whole thing, is don't put
any too many pictures in, It, yeah you're just gonna
get a GREAT FUZD i. Know and then but but,
(44:47):
then but, then because these things are really good at,
sequences think of it as any. Sequence so we have
models And andrea found with that do protein. Sequencing, yeah,
well deep. Fault and if you think about, it you're, like,
well how could that? Work, well because again you're giving
it a sequence of, things and it's telling you the
next likely. Thing and then it's doing it in a
(45:10):
loop right while not end of line token or end
of thing. Token, right that's what it's doing and get
next probable. Words. Yeah and the thing about it is
people will talk about the, temperature whether it makes it
more creative or. Whatever it's not making it more. Creative
it's basically squashing the output distribution so other things are more.
(45:32):
Likely and that's all it's. Doing and SO i hear
it's making it more. Creative, no it's. Not it's changing
the hand of the output. Distribution it'll make it, faster
but you'll get. Crap it's not even. Faster it's exactly
the same. Speed it's just the probability distribution is warped
so that if, like for, sure it's this, one if
temper is, one it's like the probably is what it
it'll pick the, right it'll. Almost but if it's, different
(45:52):
there's other things that become more equally. Likely and so
it's chain of like this, token now this token it
Becomes they say it's more, creative AND i hate. That
why DO i hate? That? Richard it's more, exactly it's
literally just changing the output.
Speaker 1 (46:09):
Distigue but, okay So joe, programmer, though do you think
they need to answer pomorphize this stuff in order to
talk about it or think about it in a in
a way that they can use it correctly because everybody
falls into this. TRAP i mean we're, Humans.
Speaker 2 (46:23):
Yeah like humans are funny. Creatures we will we will
pick up a rock and draw a smiley face on,
it and so. Happy it's such a happy. Rock it
must have had a really good Dayne and now. Imagine
but isn't THIS i WANT i have to think this
is what makes humans? Humans first, off paradolia is a survival.
(46:44):
Strategy the tendency to see, faces whether they're there or,
not means you were the first one to, run and
so the tiger probably didn't get. You so that's an evolved.
Trait absolutely, right it's. Paradolia but what if this whole
practice event memorphizing is the mechanism that led to us
being able to, strategize like how you have to think
(47:07):
like the animal to hunt the animal successfully with your limited.
RESOURCE i, See and so that tendency again becomes a competitive.
Strategy you're more likely to survive because your hunting is.
Speaker 1 (47:17):
Morphout not to mention, mythology, right, well in, mythos in
religion and all of these, things it's all about. Anthropomorphizing,
sure and not just.
Speaker 2 (47:25):
Why is the sky making loud noises?
Speaker 1 (47:27):
Shaming right As thor is sending lightning, thing that's because
he's angry with.
Speaker 2 (47:31):
US i feel like we got to go the final
mile and somehow get into The roman spears because we
all think About roman, period, RIGHT i don't, Know but,
well actually that project the name for the project that
you Saw amanda And elijah Do we called it, sustinnao
which Is latin for, SUPPORT i support, nice, Okay and
(47:53):
SO i told him that and let's lie it for.
Sales way to close the loop that was that was.
Amazing this is a quality. Show my job as co
host Of. Professionals the joke that's but, Yeah so again the,
key the key insight is think of these things as
(48:16):
things that take. Language and if you have a prompt
that's narrow enough and a collection of tools that support the,
task it will tell you what consistent. Results, yeah, yeah
That's these are the dimensions of a tool set for
a group of programmers to be successful with correct they
just had to learn the dimensions of. It and the
joke THAT i, said is it like a switch statement
(48:38):
and if statement but. Huge So i'm going to try
to make it a thing and call it the giant swift.
Statement it's the swift for those that are. Swifties swift all.
Speaker 1 (48:49):
RIGHT i figured out The latin for sales is.
Speaker 2 (48:55):
Isn't That italian full Of? Chicago oh my, gosh that's.
Hilarious but, yeah and so that that inside couple with
the ability to evaluate them as unit with the unit
with the agentic. Unit tests should bring this all back
to quality instead of the schlop that people put. Out
(49:15):
sure they, are and THAT'S i mean the battle here
is how do you measure? Quality, oh there's a way
to measure? It, yeah BECAUSE i keep being told this
THIS lm is better than THAT llm is, like how
how would you measure? That what makes that? True so
what they do is they give it a bunch of,
stuff like, like they feed it stuff and then they.
Measure i'll give you there's a couple of. Emails so
(49:35):
there's something Like blue And, rouge which is like how
close is it to the right? Answer? Right those are
supervised tests where you're, like you give it a thing
and you're, like here's the OUTPUT i wanted to. Have
we've talked about this.
Speaker 1 (49:45):
Before, Yeah blue And rouges is like those like supervised
and supervised.
Speaker 2 (49:50):
Tests. Right but then there's one supervised tests and you're, like,
well how do those? Work, Well i'll give you an.
Example there's a there's a judge BASED evl called, groundedness
and what groundedness. Measures it measures how rounded a response
from AN lm is in the facts and context that
you gave. It, Yeah and so for, example the three
pairing that you give it for the unit test is sas,
asked how much does your you, know outdoor tent? Cost
(50:13):
the context is we give it all of like the tent,
information and the cost is in, there, right it's just
like the slop of it and the outputs like the
tent is ten. Dollars. Right so given given that a judge.
Speaker 1 (50:25):
Base before the, tariffs now there are a lot, more.
Speaker 2 (50:27):
That's, Right and now they're gonna be a lot. More
but that particular three two, pole it's not, twopole but
the three pole three. Yeah of the, question the context
and the. Answer, yeah it's given to a judge to
judge on a scale of one to. Five how grounded is?
It and that judge is also AN Lllm, yeah okay
(50:49):
another one, yeah and it. Works and that's why that's
the whole idea of, Agency, yeah is that you have
agents that are specialized in their different, things just like
you have you on your. Staff there special that's. Right
the people of my staff Question, mark it's just, me all, Right.
Speaker 1 (51:05):
But Joe developer who's listening is, thinking, oh this is.
Great you, know after listening to these shows for the,
last you, know the week that we've been doing, These
i'm thinking now that you, Know i'm going to get
rid Of bob and engineering And i'm going to replace no,
agent And i'm going to set up this little staff
of programmers under. Me agent programmers are going to help
(51:27):
me do my.
Speaker 2 (51:28):
JOB i don't think that's ever going to. Happen WHAT
i think is going to happen AS i already. HAPPENED
i think we're going to solve. WELL i mean those
that think are they're going to solve it that way
are going to come in Like i've seen vibe coders
like put out stuff and be, like oh why is
MY aws bill so high.
Speaker 1 (51:45):
About vibe Coders and There's we've been arguing about whether
that's pejorative or, not because the ryan it's.
Speaker 2 (51:52):
Not, okay is it a pejorative?
Speaker 1 (51:55):
Thing some people see, it some people see it, Goodness,
Yeah But i'm talking about a professional co that maybe
works by themselves and hires out other specialists to do
that kind of. Stuff, interesting and NOW i can maybe
spin off some agents to help me with that kind
of stuff AND i not depend on them as.
Speaker 2 (52:11):
MUCH i think that the agents are going to be
able to do with a low powered brain work that
we as humans do and and but the actual creative
move forward and always human right and not and not
(52:32):
be stuck in the drudgery of. That so you actually
have more creative energy to do more experimentation. Exactly because
then that's.
Speaker 1 (52:38):
Why imagination is going to be the number one uh
feature of the Rock star. Developers, oh there you, go
there you. Go that's WHY i Think i've never been
a rock star. Though, dude you are the more like
a flute, player saying like You're Eddie eddie Van halen
jazz PLAYER ai At Microsoft Cool.
Speaker 2 (52:57):
Beings well look, again AND i just want to emphasize
this is all accessible to. Everybody you can start playing
with it. Today make it do a raw function call
and see the output so you can see that What
i'm saying is. True don't have it do the function,
call just call it and tell it to do. Return
do a tool call and you'll see it literally Returns.
(53:18):
Jason that, says run this function with these, parameters.
Speaker 1 (53:23):
And be as careful with your prompts with your agents
as you are with your ten year old when they
tell them to go to the store and get used xy.
Speaker 2 (53:30):
Eggs or your programmer.
Speaker 1 (53:31):
Husband, yes that's, right that's you, know the one about the.
Programmer your husband and the wife, said go to the
store and get a gallon of. Milk if they have,
eggs get a. Dozen he came back with a dozen.
Gas they had.
Speaker 2 (53:46):
Eggs they had. Eggs but that's just how we have
the whole world turned into paper. Clips but if you
need to be, precise but if you think about, it
that's exactly how you should think about these. Things that's.
Right you have to be very because you want to.
Listen you want to maximize the probability of it returning
the right tokens based upon the ones you gave.
Speaker 1 (54:05):
It so we're all in for a new world of
discovery and refinement of search and refinement of. Prompts, yeah
and the whole new.
Speaker 2 (54:15):
World and if you go about retrieval augmented, generation it's
literally the process of you get a, question you programmatically
search for an, answer you put it into the, prompt
and then you get the. Answer that's. It, yeah that's
all it. Is that's How, yeah we just gave it
a fancy. Name, Yeah WELL i don't like the TERM.
Rag it sounds it sounds kind Of but, YEAH i
(54:35):
do a demo WHERE i do augmented, generation WHERE i
literally go and copy From, wikipedia BUT i think the
right answer is put it in AND i show it
AND i was, like how does it know the right? Answer
and it was, like, well because you put it. In,
okay now you understand what augmented generation. Is the only
process of retrieval was artisanal in my case BECAUSE i
went and did. It it was. Artisanal, yeah it's like
(54:59):
a fine, cheese, Right but we's That and then if
you think about we're too. Generation by putting the right
context in and then coupled with tool, calling now you
have a powerful. Combo and that's WHERE i, think if
you think about it like, this you're really going to
change the way things. Happen sure, WELL i also love
how much you dilute the pretense of all this. Technology,
(55:24):
OH i know it's the. PRETENSE i said that in
a very pretentious. WAY i. Know it's kind of.
Speaker 1 (55:30):
Amazing, actually, yeah.
Speaker 2 (55:33):
It was like a two fur three For, yeah there you.
Go but like That, latin but it's a plural of. Slezolo.
Right but if you think about, it this has been
What i've been doing for the last twenty. Years, Sure,
yes that's. True i've been trying to make this stuff
(55:55):
lose all of its pretentiousness and when you, do it
empowers you to do really. Things absolutely. Sure does the
tools only get? Better that's, Right Brave New.
Speaker 1 (56:04):
World and that's the last word from. BUILD i, Suppose
WE'RE i, Love we're. Done it's TIME uh to hit the.
Road thank you for. Listening thank, You.
Speaker 2 (56:12):
Seth it's always a, pleasure always a.
Speaker 1 (56:13):
Pleasure and we'll talk to you next time on dot net.
Rocks dot Net rocks is brought to you By Franklin's
(56:41):
net and produced By Pop, studios a full service, audio
video and post production facility located physically In New, London,
connecticut and of course in the cloud online at pwop dot.
Com visit our website at D O T N E
T R o c K s dot com FOR R
S s, feeds, downloads mobile, apps, comments and access to
(57:04):
the full archives going back to show number, one recorded
In september two thousand and. Two and make sure you
check out our. Sponsors they keep us in. Business now
go write some, code see you next. Time you Got
Jent Middle vans
How'd you like to listen to dot net rocks with
no ads?
Speaker 2 (00:04):
Easy?
Speaker 1 (00:05):
Become a patron For just five dollars a month, you
get access to a private RSS feed where all the
shows have no ads. Twenty dollars a month will get
you that and a special dot net Rocks patron mug.
Sign up now at Patreon dot dot NetRocks dot com.
(00:34):
Hey guess what, it's dot net rocks from Bill. This
is the last show we were recording. I'm Carl Franklin.
We're here with Sef Juarez and we'll be talking to.
Speaker 2 (00:44):
Get that out of there. Get a little fam Sorry,
you gotta get it out. You know, we're a little punchy.
This is not only the last show of the day,
but it's the last show up. Oh, if you got
a punchy a show with a punchy dude, it's gonna
be great.
Speaker 1 (00:58):
It's gonna be You've been working hard, that's right. I
just want to start talking to Seth. Do we really
have to do the intro stuff? You should do the interesting?
Speaker 2 (01:06):
We gotta do it. You got to pay so you
know it's episode nineteen fifty seven. Wow, that's the year
I was born. It's not true. It could be in
a different universe very close. It could be in a
parallel universe. If you know quantum theory, the many worlds
as you know, Richard In posits that perhaps there's another
universe where I was born in nineteenfty seven. There's a
million universes. That happened, sixty seven year old set. I
(01:30):
just wanted to use the word posits in the thing.
And so now my job is I'm done. You must
have a master's degree.
Speaker 1 (01:37):
I have something. So do you want to do the
history first? You want to do I didn't do the history.
So what happened in nineteen fifty spot Nick one?
Speaker 2 (01:46):
Okay, the first the first artificial salarate over the world,
and then sput Nick two. They put a dog in
it and she died. No, actually no, she was in
the Guardians of the Galaxies. Right, wait, that's not real.
We want the historical documents and it's worse than any think.
(02:07):
She didn't she didn't die and re entry, she didn't suffocate.
She overheated. There wasn't temperature controls oh thanks.
Speaker 1 (02:15):
Sometime around here in the nineteen late nineteen fifties, electric
boat from General Dynamics was established in the Nautilus.
Speaker 2 (02:23):
That was no electric boat? Is that? Is that a
like a line dance? What is fifty five? Isn't it?
The electric boat? Is that like the electric slide? But no, no,
no waste? The nuclear submarine? Oh yeah, the first nuclear summer, theatlis.
Speaker 1 (02:37):
That's right, Okay, I'm sorry. I had a forward memory of.
Speaker 2 (02:42):
I also usually referenced a computer, and in nineteen fifty seven,
IBM released a compact computer, the six', ten only eight hundred,
pounds much smaller than the previous. Ones it really was
an improvement there you. GO i, mean if you want
to get, fit you get one of those commonract eight
hundred pounds. Computers also released the first compiler for. Fortrend hey, yo,
(03:03):
wow still running some code out. There there's as for
trend dot hold. On So i'm not. Kidding there are
some linear algebra operations that were written In fortrand that
people were, like let's just not do that. AGAIN i
think it's like some WEIRD langsos method for, computing like
singular value, decomposition that they're, like let's just not do that.
Again somebody did, that and we're, gonna we're, gonna we're
(03:26):
gonna honor them and we're just gonna leave that. Alone. Wow.
WOW i actually contributed this.
Speaker 1 (03:33):
SOMEHOW i actually remember taking A fortran. Class but it
was like in the. Eighties, yeah a little further, on yeah,
too it was like in THE i did fourtran.
Speaker 2 (03:44):
Three oh is there a sound? Effect weing? At BUT
i love? It all? Right, well, anyway anything else in
nineteen fifty there was a lot Of there was a
lot of. Things there's new clear. Test it's all kinds
of horrible stuff listed with, computers all, right and speaking
of atomic, fires what are we talking about? Today, well
(04:07):
we gotta finished the fish such a. Hurry i'm not
in a.
Speaker 1 (04:12):
Hurry, okay all, right let's do the better note of
framework roll the, music all. Right, so as you, Know
Jeff fritz AND I Jeff frits For microsoft And big
INTO ai And. Blazer we do This blazer puzzle.
Speaker 2 (04:34):
Show AND i was just.
Speaker 1 (04:35):
Talking to him about all of the cool THINGS i
saw at the, keynote including, you my friend oh hey
uh at the keynote it build and of course he's
hipped to all of this, stuff and we were kind
of talking about how and we talked about this with
some of our guests this. Week about how the prompts
are becoming code and instead of having code, libraries people
(04:57):
are going to start sharing prompt. Libraries Prompt so what
Did fritz point me? To he pointed me to sharp,
site which is all these prompts for Doing blazer applications
with the.
Speaker 2 (05:13):
Agent oh, interesting do they? Work?
Speaker 1 (05:15):
Yeah in good hub co pilot so code style and,
structure naming, conventions dot net specific, guidelines sho air handling and,
validation like all of these great prompts to help you
create Great blazer.
Speaker 2 (05:32):
Applications that's all the. Agent how about? THAT i think that's.
Great anything that we can do to further the work OF,
ai because you, Know i've BEEN i was like on
show like, SEVENTEEN i don't, know something like that eighteen
AND i was, like machine learning is the future AND
i was like What and we didn't understand. It then
we still, don't but at least it's. Working but it's,
true and you just still understand. It we just, believe you,
(05:55):
KNOW i really excited that people are, like let's use
this stuff right five seventy Six july of twenty. Ten
oh my, gosh aires on machines in eighteen years AND
i just get more. Beautiful i'm glad this is. True
in this, medium it's really hard to, know but it's for.
SURE i want you to imagine, it all.
Speaker 1 (06:14):
Right richard Save, us who's talking to us, today.
Speaker 2 (06:17):
Grabbed to comment off a show nineteen, nineteen which you
did in twenty twenty, four a little more recent In
october With Prascent boy here he's AN mvp DOING ai
related stuff that he wanted to talk, about and we
did talk About Copilot, studio WHICH i believe has now
been renamed because. Reasons, yeah And John tallon had this
great comments from a few months. Ago he, SAID i
enjoyed the. Show chat BASED ai apps are quick to stand,
(06:39):
up but tough to get right every time a man
rag dazzles that. First but yet it's limited without deep. Engineering.
Ooh in real, conversations not everything is a, question and
bought led chats are no. Different we must parse the user's,
intent evaluate the quality of their, input then apply a
chain of erication to fact checkbox. Responses i'm a practice
director in data AND II, ai And i've worked in Gen,
(07:01):
Ai AZURE Ai studio and Co Pilot, studio and despite
the Marketing Copilot studio isn't for your Quote jim From.
Accounts it's for seasoned consultants Like brashant who know how
to use the. Platform using it for complex, interactions like
a rocket launcher on a. Fly, yeah it, works but
it's a little over. Complicated rocket launcher on a. Fly it's,
(07:21):
like use. It, yeah let me say that. Again using
it for complex interactions is like using a rocket launcher
on a.
Speaker 1 (07:26):
Fly it's kind of taken a nod from my term
swatting a fly with a.
Speaker 2 (07:31):
Buick we had a term swatting a fly with a,
shotgun and my algorithms professor was, like we have an
algorithm for sorting that's faster than en log, in but
only with giant data. Sets it's like swatting a fly
with a. Shotgun, yes, yeah. Nice we should only explose
enough complexity to get the job. Done letting use His
harner's AI's potential without feeling. Overwhelmed after, all the goal
(07:53):
is to DEMOCRATIZE, ai not. Intimidate what year was? That
that was last? Week, yeah this is like a few months.
Ago the answer is you just, wait wait till next.
Years you. Know one of the things that CAME i
felt came out of build was this sense that nobody's
dominating in the space at. All, yeah there's lots of different,
models there's lots of different. Tools the fact that everybody
jumped ON mcp is. Clear nobody has the confidence to, go,
(08:15):
oh we'll go it. Alone we're all trying to find
standards in inter interapt with each. Other. Yeah BUT i
mean even EVEN mcp has. Limitations AND i have thoughts on,
that And i'm tired And i'm just gonna say WHAT i.
Think Oh, NO mtcp definitely has some. Issues there's no
two ways about. That but let me wrap up With john. Here,
john thanks you so much for your, comment and the
copy of Music coba is on its way to. You
and if you'd like a copy Of Music, COBE i
(08:36):
write a comment on the website at dot NetRocks dot
com or on the. Facebooks we publish every show, there
and if you comment there and every read it on the,
show we'll send you a copy Of Music. Koba And music.
Speaker 1 (08:44):
To Code by is still going. Strong twenty two tracks
to help you get in the zone and focus twenty
five minutes each to coincide with The pomodor, method and
they will help.
Speaker 2 (08:54):
YOU i actually bought the first. Ones yeah a long
times years.
Speaker 1 (08:57):
Ago, yeah it was only three AND i did A,
KICKSTARTER i did.
Speaker 2 (09:02):
THREE i got the. THREE i still have THE mp
three is somewhere in my junk drawer of one.
Speaker 1 (09:08):
Drive, well now you can get the whole collection IN
mp three flak en. Wave so you can go to
music To code by dot net for that or you,
know you could just send us a comment or leave a.
Speaker 2 (09:18):
Comment, yeah, absolutely and reading on the show you'll get
a copy of MUSIC o.
Speaker 1 (09:21):
Bye, yeah all, Right So i'm going to introduce.
Speaker 2 (09:23):
You for real areas here we go That i'm introducing.
Speaker 1 (09:27):
It, yeah that voice that you've been listening to the
can who it is Is Seph wares and he works
in THE ai platform group At. Microsoft he received his bachelor's.
Degree that's, it that's your. Title Now i'm going to
talk about your education because that's the part of your.
Bio we see his bachelors and computer science at you
n L v saying place that Guy fieri got his.
Speaker 2 (09:49):
Degree we were. Partners he's a he's a machine learning master.
Speaker 1 (09:54):
Absolutely with a minor in mathematics and complete his master's
great at The university Of utah and the field of
computer science And. SETH i have a picture that BECAUSE
i was in the front row at the build AND
i took a picture of.
Speaker 2 (10:09):
You oh what WAS i? Doing? Well you you opened the.
Speaker 1 (10:12):
Show you came out and, SAID i Am seth and you,
know with with somebody, else AND i DON'T i didn't recognize.
Her BUT i took a picture of you two and
here it. Is oh my, gosh That's. Kadasha, yeah she
is delightful human. Beings SO i sent this To, kelly my,
wife you, know, yes of course she actually met you
BEFORE i met.
Speaker 2 (10:30):
You, yes she did, so and, she.
Speaker 1 (10:33):
You, know she says back to me and she, says
that's a great. Picture but who is that black lady
With Ricky?
Speaker 2 (10:39):
Martin but, WOW i do get more. Beautiful AND i, said,
again this is the medium for me to say these
kinds of. THINGS i am getting more. Beautiful look at the.
Speaker 1 (10:53):
Picture your name is not on the, screen it's it's her. Last,
yeah so she wouldn't know and she didn't recognize.
Speaker 2 (11:00):
You that is so. Funny then WHEN.
Speaker 1 (11:02):
I told, her she, says, oh, yes that is.
Speaker 2 (11:05):
YEAH i actually owe her a. Call she called me
the other day AND i haven't called. Back, oh, okay she's,
delightful she. Called she checks up on. Me she does. That, yeah,
yeah so. Nice well you do need. Supervision it is.
True my wife has had me now for twenty two ish,
years and so during THE covid she may be. Tired
during THE.
Speaker 1 (11:24):
Covid, times she reached out to friends of mine and
hers that she hadn't talked to in years and, Said,
hey you know, What i'm just calling BECAUSE i want
to check.
Speaker 2 (11:32):
In it was.
Speaker 1 (11:34):
GREAT i don't want, anything you, KNOW i just wanted to, Say, Hi.
Speaker 2 (11:37):
GOSH i wish the niceness would rub off on other. People,
yeah it's not working. Right all.
Speaker 1 (11:45):
Right so we're talking about agentic, hey AND i got
to tell, You i'm blown away by WHAT i saw.
HERE i, MEAN i know that there have been other
agents and products that work with agency and all that,
stuff and the guys From manet have been into, it,
right but they all said the same. Thing it's, Like,
Wow microsoft is kind of catching up to what these
(12:07):
guys have been. Doing AND i hope they dominate because
they have the, power, Right microsoft has the power Of
azure and all of these. Things and WHEN i learned ABOUT,
mcp we've been talking about anybody who's listened to the
last six, shows oh, yeah knows about this about the,
(12:28):
okay is it.
Speaker 2 (12:29):
Model model context context.
Speaker 1 (12:31):
Protocol so this is the protocol by which agents can
talk to your resources to do things like talk to
YOUR sql, server talk To azure resources and all that.
Stuff and you can just tell it to go do, things,
yes and it will figure them.
Speaker 2 (12:44):
Out.
Speaker 1 (12:44):
Yes the combination of AN llm like get Hub copilot
AND mcp connections to other resources just blows my. Mind
it changes the.
Speaker 2 (12:54):
Game, Yeah i'd love to get more low level, though
so folks have an understanding of what an agent actually
is because there is This what happened was and you
were there with me for fifteen. Years i've been in
this space for fifteen. Years nobody, cared and then we
flip from nobody cared to this stuff is like jungle,
(13:17):
magic or this stuff is like going to throw you
into outer space or.
Speaker 1 (13:20):
So we didn't have to create the. Models, yeah that
was the hard.
Speaker 2 (13:23):
Thing it's. True but but what happened is we kind
of missed the in between part Of, hey let's explain
what these things are doing and why it's actually. Useful
because there's a THERE'S i don't want. To as you,
know we had SEVERAL ai winters because of the over,
hype AND i AM i am concerned that if we
(13:43):
overhype too much that there's going to be a correction
when there really doesn't need to be because it actually
this time does do qualitatively cool. Stuff, yes and that's
What i'm worried. ABOUT i, mean the winters were mostly
induced by the fact that there was a limited utility
what was being made, exactly and there was a limited
of computational power to actually make the model the thing
(14:05):
we wanted them. To hinton's famous papers actually the paper
talking about backpropagation SAYING i can't do with this equipment
in like nineteen eighty four mm. Hmmm and then In
suskovar does off that paper in two thousand and nine
and they start working on the image net. Problem, yes
and those models are quite simple as compared to the
ones that are now FOR. LMS i, MEAN i can
(14:25):
get into the architecture and stuff of. It they're just much.
Bigger and the thing about the thing about neural networks
of that size Is i've spoken To Turing award winners about,
like how do you shape these actual? Models to do
the things you, Want and the reality is we don't
really have a good understanding or. Explanation but what happens
(14:47):
is they're innovations that come out like a transformer model
with attention that was really designed to do translation that,
somehow when you made it, bigger learned language, well presented
language in a form that you could anthropomorphize. Correct and
that's that's the Part richard does not like at. All
(15:10):
the thing about it is is WHEN i picture these
things in my, head it's like the matrix for, me
Where i'm picturing like there's, weights multiplying vectors and then
going over some kind of smoothing function to, remove you,
know the linearities so that these scenes can predict complex.
Things but in the, end all these things are doing
(15:30):
is you're getting a string of words that are broken
up into vectors that get passed in and multiplied by
the transformer model to figure out how these things are,
related pushed through a giant neural network multiple, times and
the output is a giant vector of the entire space
with a probably distribution that says which token should you
(15:52):
put out? Next so when it looks like it's, typing
it isn't it's just going back through the, model and
that's what these things are. Doing and as, programmers the
only thing we can control is what goes in the
front the. Prompt and that's the beefy, bit right that
people don't. Understand at that, point when we first started
(16:13):
doing this, stuff it was magical for. Chat, Yeah but
then early, on about maybe like six months into, it
folks were starting to use these things for something that
we liked to THAT i like to call intent. Mapping
Cassie brevy AND i worked together for a couple of
years and that's what we called, it where we would
make THE lm like look at a block of text and,
(16:35):
say is this question a support or a sales? Question
and it was at that point that these, things because
we allowed them to decide is it a support or
a sales? Question when the code, branched those things all
of a sudden were imbued with this notion OF i
can make a CHOICE lm on what this thing. Is
(16:57):
and at that point those things became, agents, yes and
and and because we trusted them to answer our. Questions. Yeah.
Yeah but but but nottice that WHEN i was doing this,
ENGINEERING i narrowed the task to very specific you will
say either support or, sales and that's. It and what
(17:19):
that did is it allowed me to create a prompt
that made it so that any user input that we
put in, it it would tell me support or sales every.
Time and my code had an if statement if support do,
this if sales do, this, right and that at that,
point that is the most simplistic way of thinking about an.
Agent when AN lm has the has the ability to
(17:42):
choose reason overset of data and then act on the
reasoning you've. Generated and effectively that became a squishy programming
control structure and that's what an agent is. Fundamentally but you're,
like but it's got to be more than. That seth
what about the tools? Stuff at that people are, like,
well let's do structured output so that it could output adjacent.
(18:04):
Thing and if we can output adjacent, thing why can't
we output adjacent Thing that says this function with these.
Parameters and at that point those things became tools and
now instead of supporter, sales it, said run this function
with these. Parameters and that is the extent of all
(18:25):
of the, magic sauce of all of anything.
Speaker 1 (18:28):
Agentic NOW i want to say something Here, seth you
are among some of the most brilliant people That i've ever.
Speaker 2 (18:35):
Met, oh thank.
Speaker 1 (18:36):
You richard is one of. Them you are one of.
Them but the other people are like dev express. People,
Yeah Mark miller And Richard morris like these guys and. You,
NOW i want to tell the listeners if you don't
know Who seth. Is, seth you had an open SOURCE
AI mL library Before microsoft hired. You, yeah before they
(18:57):
Had AZURE. Ai you had this open source. Library we
did a thing about it on D NR tv that
showed it to.
Speaker 2 (19:03):
ME i didn't understand. It, yeah of COURSE i kind of.
Did but but you were doing.
Speaker 1 (19:08):
That you've been doing this longer Than microsoft has been doing.
Speaker 2 (19:12):
It way before it was. COOL i was doing before
it was. Cool the secret, sauce, though is machine learning
has been happening since nineteen fifty. Six Frank rosenblatt The,
perceptron which is the basic foundation OF.
Speaker 1 (19:25):
Ski so that was the last, show righteen Fifty.
Speaker 2 (19:28):
Oh, hey, yeah let's call back To Yesterye, frank we've missed.
Now but the reality is That microsoft's been DOING ai
in the form of machine learning for a very long.
Time WHAT i was doing IS i was trying to
bring it to the dot net community BECAUSE i recognize
that there is a space for algorithms which are learned
(19:49):
instead of, programmed AND i thought that IF i could
do a. Thing OBVIOUSLY i haven't kept up on that
for a while because it was just. Me BUT i
implement a whole set of stuff to make that. Easier
but my goal has been for the last you, know
fifteen twenty, YEARS i want people to know exactly what
these things are, doing so that with, confidence you can
make them do even cooler.
Speaker 1 (20:09):
Things, right, yeah, Yeah and so now you're At, microsoft
you're working on THE ai.
Speaker 2 (20:15):
Stuff that's The azure. Stuff that's. Right The Italians channel,
Nine oh, yeah, No channel nine was. Great you had
such a good. TIME i really. GOOD i love. That
and then somebody figured out actually knew them all, stuff
and it was. Important it's, like you come with, me you,
know CAN i be? HONEST i ACTUALLY I actually it
was different than that because AS i started doing The
channel nine, stuff the reason WHY i knew what questions
(20:36):
asked is BECAUSE i already knew their content AND i
knew how to make them talk about and make things.
Sing but people just started saying stuff to me, Like,
seth don't you ever get tired of all this technical?
Stuff And i'm, like oh, man MAYBE i should move
into something more. Technical maybe. Maybe so THEN i moved
into DevRel FOR aiml and Then John, montgomery who was
(20:57):
A cvp Of young he was, Like, seth come work
for me in THE ai core.
Speaker 1 (21:03):
Group so that's what you obviously knew your, pedigree, right
they knew they.
Speaker 2 (21:07):
Did nobody, knew LIKE i THINK i think people record
like see me and like Because i'm pretty jovial and
and BUT I i.
Speaker 1 (21:16):
Think you don't come across as AN ai.
Speaker 2 (21:18):
Scientist, NO i KNOW i do. Not and but if
you looked at, me obviously we've said this many, Times
i'm very beautiful in, appearance and people would not Like mark.
Exactly people wouldn't, know BUT I i really have been
doing machine learning. Forever and, again my primary goal is
for people to understand at a profound level what these things.
Are and to, me they are basically a new programming
(21:41):
control structure which can Take english and point you to
what to run. Next, yeah fairly still Constrained, english, yes, well,
no even Regular english with AN, lm because if you
can you can type into AN ll m like Bad.
English and for, EXAMPLE i did a demo at my
talk two days ago WHERE i showed him the supporter
(22:01):
sales thing AND i, said my tent is, broken, yo
AND i spelled it wrong and it said support AND
i was, like what's the borst tent in the, world?
Yo and it was like burst and not b e
r S t BECAUSE i fat fingered it. Right and
then it was, like, uh. Sales notice that because it's
(22:21):
a large language, model it's able to reason over the
text and then create the output that we. Want and
if that output is a, function then it tells you
what to run. Next so it.
Speaker 1 (22:33):
Didn't understand, bursts so it figured that you really meant worst.
Speaker 2 (22:36):
A best, Best, yeah, okay best b e r S.
T and it was, like even if it was, worst
it's still a sales and it still was able to
figure it. Out and and that was with me, again
and THEN i. CAN'T i don't know how to emphasize
this is. Enough, basically you get a control structure that
gives You english and tells you what things run. Next
(22:59):
WHAT mcp does and whether it does it well or
not is neither here nor. THERE i don't think it's
Necessary what it does is it allows you to query
what functions can you? Call and what it does is
it maps those into the tool calls that are available
to the L, m and then it can call. It
(23:19):
but you can do that With swagger AND. Rest, sure,
yeah but it's.
Speaker 1 (23:23):
Not it doesn't have the the almost it does.
Speaker 2 (23:27):
Reasoning it, does but it does because the reasoning happens
in THE. Lm and what it outputs is it outputs
what functions Are swagger AND rest don't have that they,
don't but you don't need it BECAUSE, mcp if you
look at, it is basically Like. Swagger it tells you
what tools are, available and it's LIKE, rest and it
tells you how to call. Them, Right and the thing
THAT i like about like those kind of and OPEN,
(23:47):
api which has been NOW i think the Next swagger is, that,
RIGHT i don't. Know but with OPEN api and WITH
rest you're able to do the same things but with
the benefit of like twenty years of security, handling with,
headers WITH, ooth with token, negotiation et. Cetera and so that's,
why like IF mcp gets you, going like WHEN i
(24:10):
USE vb, six, right it got me. GOING i loved
it because it was able to. Get but then when
you were, LIKE i need to be more memory, CONSCIOUS
i needed to be more. Careful then you can skip
down a layer if you if you want. To but,
AGAIN i love THAT mcp has gotten people thinking about.
It but the reasoning doesn't happen WITH. Mcp, no it's,
(24:30):
yeah it just tells you what function are. Next but
the magic that makes it look good is once it
runs the, function there's a thread that has like what
the QUESTION i? Asked it asks for a Function i'm
like pointing like people can. See and then when it executes,
out it puts that function return call back on the
thread and THE lm is called again with that on its,
thread and then it comes with the. Answer and so
(24:51):
that actually going out from THE ai space is effectively
bridging the gap FROM ai reasoning to perative or functional. Code.
Right how we MAKE sky? IT i get? It, NO
i mean it's NOT i don't even think IT'S.
Speaker 1 (25:05):
Sky and it may not be the best security wise right,
now but it's clearly where everybody wants to.
Speaker 2 (25:11):
Go, yeah it's because it simplifies. It but my sense
is that as you see, This i'm the prognostica, here,
right BECAUSE i like doing. That as you see this thing,
maturing it's going to be indistinguishable from rest right and discovery.
Protocols that it's the way the embracement OF mcp says to,
me everybody wants to interrupt with everybody. ELSE i fact
that we could come up with a better way to do,
(25:32):
it and in fact already. Have we could just be
publishing APIs for crying. Out, Yeah And i'm telling, you
like we we had a phase At microsoft and bless
Our hearts where we went full. Wisdole, yes and.
Speaker 1 (25:45):
That's what The southern people say about the town drunk
bless his.
Speaker 2 (25:48):
Heart, yeah bless our. Heart But wisdole was the same,
thing the same kind of desire to to abstract over.
Primitives and then we went back to primitive still being
WSD l Or Web Service Discovery language. Wisdom does that sound? Familiar,
friends done done? Dumb and so What i'm telling you
(26:09):
is that and some of the things you saw with
N a Web the reason why this is such a cool.
Speaker 1 (26:13):
Space if we haven't talked about AN a web on the,
show oh, really yeah this WEEK i haven't.
Speaker 2 (26:18):
EITHER i started looking at. It but the intuition here
is if you can language do stuff and then call
stuff and then language do stuff on the, Out like
what you if you hook this up to the in
language tool, stuff, Right you're? Fired, yeah you're my problem
WITH nl. WEBS i keep, thinking why are you talking About?
Dutch people's confused The dutch. Web but hopefully that helps
(26:41):
BECAUSE i really want to emphasize that once you know
that that's what it's doing at a low level and
you play with it that, way it changes the way
you think about These. SURE i had a journalist talk
to me and, say, Hey, SETH i get a lot
of questions about SHOULD i trust these? Things and my
response surprised. HIM i, SAID i don't trust these. Things,
no don't be. Silly that's WHY i program them the
(27:02):
WAY i. DO i restrict what they have to. DO
i restrict the tool calls that they, have and then
when the tool calls are, CALLED i write the code
that the tool that what the tool responds. With because
all these things are effectively becoming part of the prompt
for the next tokens to. Generate and that is when
(27:23):
you get that, Intuition you're, like, oh that MEANS i
can do this ANYWAY i want to ye.
Speaker 1 (27:30):
And that's that's WHERE i think we've been talking about
this this week, too that the attributes of the program
or the future are going to be imaginations over technical
probsolutely imagination mean and we may be held back a
little bit by what we think we can do and
what we can't. Do but somebody who doesn't have that
(27:52):
kind of you, know, limitation self, limitation will be able
to do amazing.
Speaker 2 (27:58):
Things. Yeah like if you watch the keynote on day,
two you Saw Amanda foster And elijah on stage talking
to this thing that it was executing a bunch of
stuff on their. BEHALF i think it was. Software it
was literally software BECAUSE i KNOW i, WROTE i wrote
THE i wrote the website and the back. End amanda
worked on the, agent And elijah worked on the on the.
(28:19):
Evaluations but the thing about it is it was actually running.
Live that thing was actually happening. Live and one of
the things that we had is once you get into
a single, agent it makes, sense but when you get
into multi, agent you wonder, like, well what is the
interface between one agent and? Another and the surprising thing
is that the interface is actually language, weird which is, weird,
(28:42):
Right so for, Example i'll give you, example the voice
agent that we, made and if you watch the, keynote
you'll see her talking to. It the voice agent is
using THE gpt four oh Real TIME, api which has
the also has the ability to issue function. Calls, Okay
and so what we did is we collected the list
of azure Ai foundry agents that we made Andree foundry
AND i just loaded them up and then she picked
(29:04):
the ones she wanted because she was working on. Them
and what it did is it mapped those other agents
as tool calls interesting to the voice. Agent but you're, like,
well what did you pass to the? Agent, well the
agents themselves when you run, them have this property called additional,
instructions which means you can augment their system prompt and
(29:24):
then that's. It that's all you can. Do so WHAT
i did IS i mapped all of those agents as tool,
calls gave it a function. Name you're back to sales
and support, again that's all this, IS i. Know and
THEN i, said you passed voice agent additional instructions and
ask it for what you wanted to. Do so the
voice agent then is taking the initial requests and assessing
(29:47):
it to go which agent to go, to back to
your sales and. Support that's. Right then taking the portion
of the requests that makes sense for that, agent plus
the additional, instruction let's send it on. Down but it's
even more than. That it actually asks, it can you
help me with? This and we were me And amanda
because it's like in the two in the, morning we're
looking at this, stuff like because we're on this stuff all.
Night we looked at the law because we Ad jason
(30:09):
printed out everything and we could, see, uh image agent system,
instructions you are to generate a blah blah blah. Question
can you generate an? Image it was Actual english and
we were, like holy, cow this is this is amazing
and the software being polite to. Software, yes it.
Speaker 1 (30:26):
Was really cook but but it's nothing. New if you
look at the S MT smtp protocol when you connect
When i'm When gmail connects to an S mvp, server it,
says you, know, hello And I'm Carl, Franklin carl frank
and THE smtp server actually writes, back, Hello Carl, franklin
(30:47):
nice to meet.
Speaker 2 (30:48):
You, yes but we're not talking part of the. Protocol
it's Not eliza in This, NO i, KNOW.
Speaker 1 (30:53):
I, know but but you know that those kind of
friendly things have been.
Speaker 2 (30:57):
Protocols for a long. Time but the difference is that
the only things are being generated by the through the
agency of AN. LLM i get. It it's, big big.
CHANGE i do question if if if the politeness affects
the prompt. Results, yes it absolutely. Does And i'll tell.
You i'll tell you what happens And i'll tell. You
(31:19):
so there was a part of the demo Where amanda
generates an image for a hackathon and they AND i
don't think they understood because someone backstage is, like, hey
can you change the image to something more? Happy and we're, like,
no it generates it every. Time it's it's doing it
on its. Own and a man is, like but we
can change it by telling, it please make the image more,
(31:42):
happy and like we made that change in The english
and the and not in the voice controller because that's
the thing we talked, to but in the literal little
agent the, customer and it called, it called. It and
then it was, like and now everyone's happy at the.
Hackathon and and because everyone thought we were using some
like pre baked image and, no it's not every, time
(32:05):
and then we did, that and then it was really cool.
Speaker 1 (32:07):
All, RIGHT i think we got to take a, break
so we will be right back after these very important
messages don't go. Away did you know there's a dot
net on aws. Community follow the social media, blogs YouTube
influencers and open source projects and add your own. Voice
get plugged into the dot net on aws community at
(32:29):
aws Dot amazon dot, com slash dot, net.
Speaker 2 (32:36):
And we're.
Speaker 1 (32:36):
Back it's dot net. Rocks I'm Carl, Franklin I'm Richard,
campbell And Seth warez is. Here we are JUST i
feel like we could be talking for, hours hours an,
hour so much to talk, about but we're talking about
agentic coding and. PROGRAMMING i Took microsoft Agent mode In
Visual studio for a spin yesterday to go.
Speaker 2 (32:57):
Good, okay it went really.
Speaker 1 (32:58):
WELL i gave it some tasked to create some stuff
THAT i knew how to do BECAUSE i wanted to.
See but there was one very tricky thing that it
was intuitive for it to, do but it didn't. Work
so it's not like agent mode is going to compile
your code.
Speaker 2 (33:17):
And see if it.
Speaker 1 (33:18):
Works it's just going to do what it thinks is. Right, yes,
SORRY i didn't meet us the ord. Thing it's going
to use what it assesses to be the correct. Answer
so but WHEN i went and told, it, no that's.
Wrong this doesn't work on this particular. Object you're going
to have to subclass, this it, said oh. Okay AND
i didn't tell it what to. Do ALL i had
(33:38):
to say was you're going to have to subclass this,
subject and it figured it.
Speaker 2 (33:42):
Out it's pretty cool. Stuff one of the things that
you're going to find is that if you ever have
an agentic thing that is not quite doing what you
want in, general from an engineering, perceptive for, me it
means that they've overstuffed what they want a single agent to.
Do so the actual engineering problem becomes agent specificity versus agent.
(34:06):
Count and that's an engineering, issue, right because if you
have too many, agents it might cost, more or it
might it might work. Slower if you have too few,
agents the recall isn't going to be as, good, Right
and so the actual engineering problem is that it turns
out the agent you use was. Able once you specify
down a little, bit it made things. Better but if
you think you're going to make a single agent with
(34:28):
like a two page prompt with fifteen, tools it's not
going to do the right. Thing and that's the trade
off from an engineering pler because now that you know
you have to make it, specific you have to limit
its tool because it's an interesting switch from what build
up lms the first, place which is making them bigger and.
Bigger now we're, going, hey if you wanted me to respond.
Better but see that's the thing, small that's a. Thing
(34:49):
it's not THE lms that are getting, right it's the
tasks that you're assigning THE. Lm so it's THE. Lm
let's just fix THE lm for. One let's just SAY gbd. Four,
oh like that models, Huge but like you are gonna
narrow the agent's task to your job is to send
an email to the right. Place and then here are
the tools send email or delete email or. Whatever those
(35:12):
are the only. Things that thing is gonna be very
good at sending, emails and it's gonna be very good
to be, like, HEY i need to send an emails
to whom please type or tell me the. Email and
then it's gonna be, like please validate the. Email what
content do you? Want because it's trying to fill out
the parametery of the function call because that's what the
prompt wants it to. Do and that's what's. Happening so
(35:35):
for your, case LIKE i worry like even in some
of the tools that we, make maybe At, microsoft maybe
they're too broad and that's where you find this notion of.
Hallucination now people aren't gonna like, this but the elem's
job is to actually make up. Tokens if you give
it too broad a, thing it's gonna make stuff. Up
if you get it to if you give it a narrow,
(35:56):
thing it's gonna make the right stuff. Up BUT i
thought that's.
Speaker 1 (35:59):
What the point of the agent THAT i access in
visual studio, is and it's NOT gpd four oh with
this broad uh set of knowledge or. Whatever i'm, Sorry,
RICHARD i didn't mean to use the. KNOWLEDGE i have
to apologize every TIME i answerpomorphize than LL.
Speaker 2 (36:14):
M but he is. Right thank you keep him on. Track, yeah,
No i'm. SORRY i apologize now ALREADY i. Apologize please
be with the pigmies In New. Guinea. Yeah.
Speaker 1 (36:23):
Amen but it narrows the focus to truss the things
that are in. Code so theoretically it should know everything
about the Language i'm using the, markup you, know If
i'm Using blazer or. Whatever so it's got to UNDERSTAND
html and JavaScript and all Those you're.
Speaker 2 (36:39):
Gonna love this. Answer it knows. Nothing it knows only
what You i'm trained on. It, well, yeah so there
there is a, yeah so that's WHAT i. Mean, Yeah
so there's certain models that are actually Like, codex for,
example is perfectly trained on, code but it doesn't know
like in my, head it doesn't know the. Languages it
knows that if you give it this sequence of programming or,
(37:00):
questions it's going to return this. Sequence let me just
add this to the.
Speaker 1 (37:04):
EQUATION i did the same problem with chatchipt AND i
did get the right. Answer, yes, so but the thing
that was The Microsoft agent mode In Visual studio did
not give me the right answer.
Speaker 2 (37:17):
Because it's a different.
Speaker 1 (37:17):
Model, Right but wouldn't it be narrowed down to just
the programming stuff like why did CHAT gpt with vast
programming and learning that it did on all this data
provide a better answer.
Speaker 2 (37:30):
Because models are stochastic, processes and the thing about one
model versus another is one will do better with the
same prompt or. Not and so that's when he gets
the college. Words that's five dollars five dollar. Word it's
a stochastic process because the output vector gives you the
probability distribution and samples according to that probably. Distribution so some,
(37:55):
models depending on let's just say that use clawed in that,
one or they are going to do different things and
it all so the promptly makes a, difference and the
tools that it has really makes a. Difference and that's
where it sounds like an excuse to. Me it does
and here's what they should have. Done just like we
have unit tests for, code we should have quote unit
(38:15):
tests for. Agents and those things are called. Evaluations and
so An azree foundery we have a bunch of evaluations for,
example is it follow the right? Task is it? Coherent
is it? Relevant is it? Grounded we also have these
other agents whose job is to attack our. Agents they're
called the red teaming, agent and its job is to
(38:37):
literally send garbage at it to see how it. Does
and so that is the unit. Test and the difference
between a unit tests and an evaluation is that an
evaluation you need to give it enough examples to be
satisfied that you're going to be. Okay it needs to
be probabilistically, relevant, Yes and then once you're okay with,
it when you put it in, production you monitor it
(38:59):
with the same evails and send alerts when it goes.
Off the rails diverts because when you're looking, at for,
example regulated, industries regular, industries like if you make a,
mistake you got to report, It but they also want
to know how do you fix these? Things and that's
probably more, important and that's the part of regulated. Industries
when you use this s uff you can see where
it went wrong and you fix it and then you go.
Speaker 1 (39:20):
Back SO i think the real answer, is, hey we
just launched this, shit give us some time.
Speaker 2 (39:26):
Time i'm not part of that. Team they mean they
could have made it, better.
Speaker 1 (39:29):
Maybe BUT i mean it literally just came, out like you,
know so and.
Speaker 2 (39:34):
They'll learn a lot, because, like for, example the question
you asked it maybe something they never. Anticipated that's, Right,
yes so there's Learn so there's some alert that probably
fired during THE evls that's, saying, Oh carl did not
get this thing that he, wanted and so they update
the prompt or a new tools AND i expect.
Speaker 1 (39:50):
That. Yeah, so BUT i love the, idea and you,
know talking To Scott hunter and some other, people the
possibilities of things That i'm to do with it in
the real world from my customers is going to blow their.
Speaker 2 (40:06):
Minds, Yes and in terms of like the stuff you can,
do it's the ceiling is very high because the biggest
barrier to programming isn't like the new algorithms or the
new it's like literally people using this stuff and having
the like like someone Putting bob in the age. Field you,
(40:26):
know that's the kind of dumb stuff that we write
most of our code. For right ll ms together in
an agentic, mode are going to soften that so that
we don't have like why do we need a web form?
ANYMORE i can do like my name Is seth and
you can use my current location to find the.
Speaker 1 (40:43):
Address, now you're absolutely, right and YOU i is going to,
change yeah from you. Know but but users are trained on,
forms aren't. They, yes you're going to have to train
a user to all. Right so here's here's a. Nightmarebox nightmare.
Scenario a textbox to tell the application what you, want
and somebody comes in and, says, hey you want A turkey.
(41:05):
Club i'm going to The. Delhi, Yeah i'll have A turkey.
Club and that's Writing turkey, club you know in the
in the. Thing you know WHAT i mean, like there's there's.
You now you have to train the users as to
what to say or what to, type and that's the.
Speaker 2 (41:19):
Beauty. THOUGH i think given a properly built, agent that
turkey sandwich is just gonna be noise to. It like
it's gonna look for the things that will force the
linear algebra and the calculus to produce the right next.
Speaker 1 (41:36):
Time, yeah and then there's gotta be validation and. Verification
did you really mean to say? This is this what
you really?
Speaker 2 (41:41):
Want you? Know and ALSO i think the software is
fairly good at recognizing a contact shift and SAYING i
don't THINK i was supposed to listen to, that, yeah.
Speaker 1 (41:49):
Right and Also i'm assuming that you're speaking to it
right when you really need to be, Typing like you
don't Want bob From engineering to pop his head in and,
say delete all records from, customers you, know.
Speaker 2 (42:01):
DUDE i saw AN mcp server that was attached to
a file, system AND i was, like maybe let's not do. That,
yeah it seems like about on mind's. IDEA i remember
when the you, know the First amazon devices showed, up
and right away it was, like send my porn collection
to my mother.
Speaker 1 (42:15):
Joke oh my, gosh send one hundred pounds of cement
To Richard.
Speaker 2 (42:19):
Campbell oh my, gosh so. Dumb but, yes BUT i
think LIKE i, Think LIKE i said. It it was
surprising to me how the the controller agent in this
case was communicating with other. Agents, yes that. Was and
then you, Know i'm not surprised by this stuff BECAUSE
i know how it. Works, Right but that was the
(42:41):
part that was, Like, okay maybe there's something. Here, well
and we get back to this whole like The battle
Of VISIBLE. Ai, yeah, right and we still haven't really
fixed in there all that. Problem, no there's no way
we can fix. That, well and that the Whole i've
read the anthropic paper on the reasoning models and what's
actually going on. There it's calling itself right couple. Times,
well first it gets the, results and then it peaks
(43:03):
pieces of the results and feeds it back into it
to say why WOULD i have said? This so it
literally is in reverse. Reasoning i've already decided what the answer,
is but let me break it down into pieces so
it looks Like i'm thinking about. It, yeah that doesn't
seem to be a good way to do, it but
it makes people, happy, Right, well it doesn't make me,
happy well until you know what's going. On and it's
LIKE i just.
Speaker 1 (43:20):
Read something in the news today that said THE ceo
Of anthropic admitted that he has no idea how this
stuff for but.
Speaker 2 (43:28):
You, know maybe like in his, defense like there's certain
things that CEOs do THAT i just don't want to
know how to. Do, yeah you know What i'm, Saying
like there's only there's only so MANY i don't know
if this is, true but there's only so much space
in my brain for things, nice, Right and it's LIKE
i have, seen like BECAUSE i work with a lot
of our EVPs and CEOs and and all that, stuff
(43:50):
and like And i've worked And i've talked to CEOs
and cvps and it's just like MAYBE i don't want
to do. That, Yeah and and the fact that he
doesn't know exactly how it works is, okay focus on
the right. Things. Yeah and the other thing that for,
me LIKE i mathematically know how these things, work BUT
(44:10):
i don't expect other people to just to. Me it's
just LIKE i love looking at like the beauty of
the mathematical structure of these things and trying to think
about why does this actually?
Speaker 1 (44:21):
Work but you know, WHAT i feel better knowing that
as long AS sef.
Speaker 2 (44:25):
Knows BUT i think this idea that that neural net
is is essentially creating a jpeg of, language lossy compression
of language one hundred. Percent that is exactly. Right but
when we, say and the whole thing, is don't put
any too many pictures in, It, yeah you're just gonna
get a GREAT FUZD i. Know and then but but,
(44:47):
then but, then because these things are really good at,
sequences think of it as any. Sequence so we have
models And andrea found with that do protein. Sequencing, yeah,
well deep. Fault and if you think about, it you're, like,
well how could that? Work, well because again you're giving
it a sequence of, things and it's telling you the
next likely. Thing and then it's doing it in a
(45:10):
loop right while not end of line token or end
of thing. Token, right that's what it's doing and get
next probable. Words. Yeah and the thing about it is
people will talk about the, temperature whether it makes it
more creative or. Whatever it's not making it more. Creative
it's basically squashing the output distribution so other things are more.
(45:32):
Likely and that's all it's. Doing and SO i hear
it's making it more. Creative, no it's. Not it's changing
the hand of the output. Distribution it'll make it, faster
but you'll get. Crap it's not even. Faster it's exactly
the same. Speed it's just the probability distribution is warped
so that if, like for, sure it's this, one if
temper is, one it's like the probably is what it
it'll pick the, right it'll. Almost but if it's, different
(45:52):
there's other things that become more equally. Likely and so
it's chain of like this, token now this token it
Becomes they say it's more, creative AND i hate. That
why DO i hate? That? Richard it's more, exactly it's
literally just changing the output.
Speaker 1 (46:09):
Distigue but, okay So joe, programmer, though do you think
they need to answer pomorphize this stuff in order to
talk about it or think about it in a in
a way that they can use it correctly because everybody
falls into this. TRAP i mean we're, Humans.
Speaker 2 (46:23):
Yeah like humans are funny. Creatures we will we will
pick up a rock and draw a smiley face on,
it and so. Happy it's such a happy. Rock it
must have had a really good Dayne and now. Imagine
but isn't THIS i WANT i have to think this
is what makes humans? Humans first, off paradolia is a survival.
(46:44):
Strategy the tendency to see, faces whether they're there or,
not means you were the first one to, run and
so the tiger probably didn't get. You so that's an evolved.
Trait absolutely, right it's. Paradolia but what if this whole
practice event memorphizing is the mechanism that led to us
being able to, strategize like how you have to think
(47:07):
like the animal to hunt the animal successfully with your limited.
RESOURCE i, See and so that tendency again becomes a competitive.
Strategy you're more likely to survive because your hunting is.
Speaker 1 (47:17):
Morphout not to mention, mythology, right, well in, mythos in
religion and all of these, things it's all about. Anthropomorphizing,
sure and not just.
Speaker 2 (47:25):
Why is the sky making loud noises?
Speaker 1 (47:27):
Shaming right As thor is sending lightning, thing that's because
he's angry with.
Speaker 2 (47:31):
US i feel like we got to go the final
mile and somehow get into The roman spears because we
all think About roman, period, RIGHT i don't, Know but,
well actually that project the name for the project that
you Saw amanda And elijah Do we called it, sustinnao
which Is latin for, SUPPORT i support, nice, Okay and
(47:53):
SO i told him that and let's lie it for.
Sales way to close the loop that was that was.
Amazing this is a quality. Show my job as co
host Of. Professionals the joke that's but, Yeah so again the,
key the key insight is think of these things as
(48:16):
things that take. Language and if you have a prompt
that's narrow enough and a collection of tools that support the,
task it will tell you what consistent. Results, yeah, yeah
That's these are the dimensions of a tool set for
a group of programmers to be successful with correct they
just had to learn the dimensions of. It and the
joke THAT i, said is it like a switch statement
(48:38):
and if statement but. Huge So i'm going to try
to make it a thing and call it the giant swift.
Statement it's the swift for those that are. Swifties swift all.
Speaker 1 (48:49):
RIGHT i figured out The latin for sales is.
Speaker 2 (48:55):
Isn't That italian full Of? Chicago oh my, gosh that's.
Hilarious but, yeah and so that that inside couple with
the ability to evaluate them as unit with the unit
with the agentic. Unit tests should bring this all back
to quality instead of the schlop that people put. Out
(49:15):
sure they, are and THAT'S i mean the battle here
is how do you measure? Quality, oh there's a way
to measure? It, yeah BECAUSE i keep being told this
THIS lm is better than THAT llm is, like how
how would you measure? That what makes that? True so
what they do is they give it a bunch of,
stuff like, like they feed it stuff and then they.
Measure i'll give you there's a couple of. Emails so
(49:35):
there's something Like blue And, rouge which is like how
close is it to the right? Answer? Right those are
supervised tests where you're, like you give it a thing
and you're, like here's the OUTPUT i wanted to. Have
we've talked about this.
Speaker 1 (49:45):
Before, Yeah blue And rouges is like those like supervised
and supervised.
Speaker 2 (49:50):
Tests. Right but then there's one supervised tests and you're, like,
well how do those? Work, Well i'll give you an.
Example there's a there's a judge BASED evl called, groundedness
and what groundedness. Measures it measures how rounded a response
from AN lm is in the facts and context that
you gave. It, Yeah and so for, example the three
pairing that you give it for the unit test is sas,
asked how much does your you, know outdoor tent? Cost
(50:13):
the context is we give it all of like the tent,
information and the cost is in, there, right it's just
like the slop of it and the outputs like the
tent is ten. Dollars. Right so given given that a judge.
Speaker 1 (50:25):
Base before the, tariffs now there are a lot, more.
Speaker 2 (50:27):
That's, Right and now they're gonna be a lot. More
but that particular three two, pole it's not, twopole but
the three pole three. Yeah of the, question the context
and the. Answer, yeah it's given to a judge to
judge on a scale of one to. Five how grounded is?
It and that judge is also AN Lllm, yeah okay
(50:49):
another one, yeah and it. Works and that's why that's
the whole idea of, Agency, yeah is that you have
agents that are specialized in their different, things just like
you have you on your. Staff there special that's. Right
the people of my staff Question, mark it's just, me all, Right.
Speaker 1 (51:05):
But Joe developer who's listening is, thinking, oh this is.
Great you, know after listening to these shows for the,
last you, know the week that we've been doing, These
i'm thinking now that you, Know i'm going to get
rid Of bob and engineering And i'm going to replace no,
agent And i'm going to set up this little staff
of programmers under. Me agent programmers are going to help
(51:27):
me do my.
Speaker 2 (51:28):
JOB i don't think that's ever going to. Happen WHAT
i think is going to happen AS i already. HAPPENED
i think we're going to solve. WELL i mean those
that think are they're going to solve it that way
are going to come in Like i've seen vibe coders
like put out stuff and be, like oh why is
MY aws bill so high.
Speaker 1 (51:45):
About vibe Coders and There's we've been arguing about whether
that's pejorative or, not because the ryan it's.
Speaker 2 (51:52):
Not, okay is it a pejorative?
Speaker 1 (51:55):
Thing some people see, it some people see it, Goodness,
Yeah But i'm talking about a professional co that maybe
works by themselves and hires out other specialists to do
that kind of. Stuff, interesting and NOW i can maybe
spin off some agents to help me with that kind
of stuff AND i not depend on them as.
Speaker 2 (52:11):
MUCH i think that the agents are going to be
able to do with a low powered brain work that
we as humans do and and but the actual creative
move forward and always human right and not and not
(52:32):
be stuck in the drudgery of. That so you actually
have more creative energy to do more experimentation. Exactly because
then that's.
Speaker 1 (52:38):
Why imagination is going to be the number one uh
feature of the Rock star. Developers, oh there you, go
there you. Go that's WHY i Think i've never been
a rock star. Though, dude you are the more like
a flute, player saying like You're Eddie eddie Van halen
jazz PLAYER ai At Microsoft Cool.
Speaker 2 (52:57):
Beings well look, again AND i just want to emphasize
this is all accessible to. Everybody you can start playing
with it. Today make it do a raw function call
and see the output so you can see that What
i'm saying is. True don't have it do the function,
call just call it and tell it to do. Return
do a tool call and you'll see it literally Returns.
(53:18):
Jason that, says run this function with these, parameters.
Speaker 1 (53:23):
And be as careful with your prompts with your agents
as you are with your ten year old when they
tell them to go to the store and get used xy.
Speaker 2 (53:30):
Eggs or your programmer.
Speaker 1 (53:31):
Husband, yes that's, right that's you, know the one about the.
Programmer your husband and the wife, said go to the
store and get a gallon of. Milk if they have,
eggs get a. Dozen he came back with a dozen.
Gas they had.
Speaker 2 (53:46):
Eggs they had. Eggs but that's just how we have
the whole world turned into paper. Clips but if you
need to be, precise but if you think about, it
that's exactly how you should think about these. Things that's.
Right you have to be very because you want to.
Listen you want to maximize the probability of it returning
the right tokens based upon the ones you gave.
Speaker 1 (54:05):
It so we're all in for a new world of
discovery and refinement of search and refinement of. Prompts, yeah
and the whole new.
Speaker 2 (54:15):
World and if you go about retrieval augmented, generation it's
literally the process of you get a, question you programmatically
search for an, answer you put it into the, prompt
and then you get the. Answer that's. It, yeah that's
all it. Is that's How, yeah we just gave it
a fancy. Name, Yeah WELL i don't like the TERM.
Rag it sounds it sounds kind Of but, YEAH i
(54:35):
do a demo WHERE i do augmented, generation WHERE i
literally go and copy From, wikipedia BUT i think the
right answer is put it in AND i show it
AND i was, like how does it know the right? Answer
and it was, like, well because you put it. In,
okay now you understand what augmented generation. Is the only
process of retrieval was artisanal in my case BECAUSE i
went and did. It it was. Artisanal, yeah it's like
(54:59):
a fine, cheese, Right but we's That and then if
you think about we're too. Generation by putting the right
context in and then coupled with tool, calling now you
have a powerful. Combo and that's WHERE i, think if
you think about it like, this you're really going to
change the way things. Happen sure, WELL i also love
how much you dilute the pretense of all this. Technology,
(55:24):
OH i know it's the. PRETENSE i said that in
a very pretentious. WAY i. Know it's kind of.
Speaker 1 (55:30):
Amazing, actually, yeah.
Speaker 2 (55:33):
It was like a two fur three For, yeah there you.
Go but like That, latin but it's a plural of. Slezolo.
Right but if you think about, it this has been
What i've been doing for the last twenty. Years, Sure,
yes that's. True i've been trying to make this stuff
(55:55):
lose all of its pretentiousness and when you, do it
empowers you to do really. Things absolutely. Sure does the
tools only get? Better that's, Right Brave New.
Speaker 1 (56:04):
World and that's the last word from. BUILD i, Suppose
WE'RE i, Love we're. Done it's TIME uh to hit the.
Road thank you for. Listening thank, You.
Speaker 2 (56:12):
Seth it's always a, pleasure always a.
Speaker 1 (56:13):
Pleasure and we'll talk to you next time on dot net.
Rocks dot Net rocks is brought to you By Franklin's
(56:41):
net and produced By Pop, studios a full service, audio
video and post production facility located physically In New, London,
connecticut and of course in the cloud online at pwop dot.
Com visit our website at D O T N E
T R o c K s dot com FOR R
S s, feeds, downloads mobile, apps, comments and access to
(57:04):
the full archives going back to show number, one recorded
In september two thousand and. Two and make sure you
check out our. Sponsors they keep us in. Business now
go write some, code see you next. Time you Got
Jent Middle vans
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Law & Order: Criminal Justice System - Season 1 & Season 2
Season Two Out Now! Law & Order: Criminal Justice System tells the real stories behind the landmark cases that have shaped how the most dangerous and influential criminals in America are prosecuted. In its second season, the series tackles the threat of terrorism in the United States. From the rise of extremist political groups in the 60s to domestic lone wolves in the modern day, we explore how organizations like the FBI and Joint Terrorism Take Force have evolved to fight back against a multitude of terrorist threats.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com