April 3, 2025 • 48 mins
In this episode of Sidecar Sync, Amith and Mallory dive deep into two major AI model releases—Google’s Gemini 2.5 Pro and DeepSeek’s V3—and explore how they’re reshaping the landscape of artificial intelligence. They discuss the technical breakthroughs, from million-token context windows to open-source innovations, and what it all means for association leaders. This is a jam-packed episode with strategic insights for the forward-thinking association exec.
🔎 Check out Sidecar's AI Learning Hub and get your Association AI Professional (AAiP) certification:
https://learn.sidecar.ai
📕 Download ‘Ascend 2nd Edition: Unlocking the Power of AI for Associations’ for FREE
https://sidecar.ai/ai
📅 Find out more digitalNow 2025 and register now:
https://digitalnow.sidecar.ai/
🚀 Find out more about the upcoming Blue Cypress Innovation Hub! https://bluecypress.io/innovation-hub-chicago
💡 Join Sidcar’s AI Mastermind Group:
https://sidecar.ai/association-ai-mastermind
🛠 AI Tools and Resources Mentioned in This Episode:
GPT-4o ➡ https://openai.com
Gemini 2.5 Pro ➡ https://deepmind.google/technologies/gemini
DeepSeek-V3 ➡ https://huggingface.co/deepseek-ai
Midjourney ➡ https://www.midjourney.com
HeyGen ➡ https://www.heygen.com
Amith’s Comic Strip ➡ https://shorturl.at/2Khn5
Chapters:
00:00 - Introduction
02:02 - Catching Up: Innovation Hub in Washington DC
07:58 - Gemini 2.5 Pro
15:58 - Google’s Position and Counter-Positioning Strategy
20:57 - Visual Breakthroughs in GPT-4o’s Image Generator
24:00 - Small vs. Large Models
28:36 - DeepSeek-V3 Release: What’s New and Different
33:29 - Why Open Source Models Matter for Associations
38:04 - The Impact of Lower AI Costs on Strategy
41:23 - Dream Bigger: Breaking Constraints and Thinking Abundantly
44:48 - Final Thoughts
🚀 Sidecar on LinkedIn
https://www.linkedin.com/company/sidecar-global/
👍 Like & Subscribe!
https://x.com/sidecarglobal
https://www.youtube.com/@SidecarSync
https://sidecar.ai/
Amith Nagarajan is the Chairman of Blue Cypress https://BlueCypress.io, a family of purpose-driven companies and proud practitioners of Conscious Capitalism. The Blue Cypress companies focus on helping associations, non-profits, and other purpose-driven organizations achieve long-term success. Amith is also an active early-stage investor in B2B SaaS companies. He’s had the good fortune of nearly three decades of success as an entrepreneur and enjoys helping others in their journey.
📣 Follow Amith:
https://linkedin.com/amithnagarajan
Mallory Mejias is the Manager at Sidecar, and she's passionate about creating opportunities for association professionals to learn, grow, and better serve their members using artificial intelligence. She enjoys blending creativity and innovation to produce fresh, meaningful content for the association space.
📣 Follow Mallory:
https://linkedin.com/mallorymejias
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
You shouldn't think of AI video generation or AI
whatever as a scarce resource,that you can only use it in a
tiny number of areas.
You should be thinking that bythe end of this year, you'll
probably be able to do it acrosseverything.
Welcome to Sidecar Sync, yourweekly dose of innovation.
If you're looking for thelatest news, insights and
(00:20):
developments in the associationworld, especially those driven
by artificial intelligence,you're in the right place.
We cut through the noise tobring you the most relevant
updates, with a keen focus onhow AI and other emerging
technologies are shaping thefuture.
No fluff, just facts andinformed discussions.
I'm Amit Nagarajan, chairman ofBlue Cypress, and I'm your host
(00:42):
.
Greetings and welcome to theSidecar Sync, your home for
content at the intersection ofassociations and AI.
My name is Amit Nagarajan.
Speaker 2 (00:53):
And my name is Mallory Mejiaz.
Speaker 1 (00:54):
And we are your hosts .
Today, we are going to coversome crazy and awesome and
exciting things that arehappening at the forefront of
artificial intelligence, andwe're going to tell you how they
might apply to your world as anassociation leader.
Before we do that, though,let's take a moment to hear a
quick word from our sponsor.
Speaker 2 (01:14):
If you're listening to this podcast right now,
you're already thinkingdifferently about AI than many
of your peers, don't you wishthere was a way to showcase your
commitment to innovation andlearning?
The Association AI Professional, or AAIP, certification is
exactly that.
The AAIP certification isawarded to those who have
achieved outstanding theoreticaland practical AI knowledge.
(01:36):
As it pertains to associations,earning your AAIP certification
proves that you're at theforefront of AI in your
organization and in the greaterassociation space, giving you a
competitive edge in anincreasingly AI-driven job
market.
Join the growing group ofprofessionals who've earned
their AAIP certification andsecure your professional future
(01:58):
by heading to learnsidecarai.
Amit, how are you doing today?
Speaker 1 (02:04):
I'm doing great.
I just got back from DC.
I think I may have picked up alittle bit of a cold or
something on the way back, buthere I am.
I'm back in New Orleans anddoing well.
How about yourself?
Speaker 2 (02:14):
I'm doing pretty well myself.
Nothing like a little dose ofsidecar sink to make you feel
better, right?
Speaker 1 (02:20):
That's right.
It's like an adrenaline shot.
It's awesome.
Speaker 2 (02:22):
Exactly.
I have been itching to find outhow the Innovation Hub went.
I was not able to attend and Ihaven't really talked to you yet
, amit, about it, so this is thefirst time I'm hearing.
But how was it?
How did it go?
How was your session?
Speaker 1 (02:36):
It was amazing.
So this is the third annual DCInnovation Hub.
We have the Chicago one comingup in two weeks in Chicago April
8th, and that's also going tobe an amazing event.
It's at the American College ofSurgeons office downtown and
the DC Innovation Hub was at theAmerican Geophysical Union just
north of DuPont Circle in DC.
(02:57):
Beautiful location, amazingconference, wonderful hosts.
We thank our friends at AGU forthat and we had great turnout.
We had about 70 people show up.
This is our small communityevent we do in DC and Chicago
each year.
Our main flagship event, asmany of you know who are
listening, is Digital Now, whichwill be in Chicago this fall.
(03:19):
But back to the Innovation Hubin DC.
It was really a cool moment intime, I think, in a way Mallory,
in that we really are seeingpeople build.
They're past the contemplationstage.
Many of the people that were inthe room were there to listen
and learn, but they were alsothere to share.
That's the idea behind theinnovation hubs.
We started those as informalcommunity gatherings to
(03:41):
certainly share some content andthings that various folks
across our family of companiesare doing, but more than
anything to kind of take a feelof the community and say, hey,
like what's going on, and sopeople are talking about
deploying AI in a lot ofinteresting ways.
So it was super fun.
I learned a ton, met some greatpeople I hadn't had a chance to
meet in person before.
All around, Really well worthit, and I hope those of you that
(04:03):
are listening that are evensomewhat close to Chicago
consider joining us on April 8th.
We still have a little bit ofroom left.
Speaker 2 (04:10):
Were there any key takeaways or any like challenges
, any patterns that you noticedemerge of all the associations
there said, oh, we're reallystruggling with X, anything like
that.
Speaker 1 (04:21):
You know, I think, more than anything, people seem
ready.
So even last year at this time,the way I felt was that people
were still nervous, they werecontemplating, they were
learning, but they were stillkind of like, hmm, I don't know
if we should do this.
I'm not sure if AI is ready tosupport the critical work of our
association at scale.
(04:42):
A lot of people a year ago werealready, you know, doing
personal experiments andactually, in some cases,
significant work with tools likeClaude and ChatGPT.
But to deploy it as anorganization, right to go out
into the world and say this isour association's AI for
knowledge or this is ourassociation's tool for search or
personalization, few were doingthat.
(05:02):
And this time around, what Ifelt was that people were saying
, yeah, we are doing this.
It wasn't a question of if, andI think that's exciting because
you know AI is not going towait for any of us, whether
we're, you know, do-gooders inthe not-for-profit sector or
somebody else.
Like AI is not hanging out andjust saying, hey, we'll wait for
you however long you want.
So I was excited to see peopletaking action, because that's
(05:24):
really the learning loop, right.
We talk at Sidecar all the timeabout the AI learning journey.
It's not an event.
It's a continuous and foreverprocess, and that's true in any
domain.
It's definitely true in AI, andwhen you go and do the thing
yourself, that's when you reallylearn and you build
organizational reps.
That leads to organizationalstrength, obviously leads to
culture change.
It's cool, so I was reallypumped up about that.
(05:50):
We heard about people talkingabout a lot of different kinds
of AI as well, so some peopledoing interesting analytics
stuff, some people are doingpersonalization at scale, some
people are doing knowledgeassistant work.
So it was a lot of fun.
Speaker 2 (05:59):
That's awesome.
I feel like you and I, andthrough the podcast, through
Sidecar, getting to witness thisjourney from its inception all
the way to now, all the way intothe future.
I'm thinking of the AIMastermind group that Sidecar
runs for Association C-suiteleaders and thinking how, in its
early stages, the AI Mastermindgroup we had to lead most of
the sessions because there justweren't any use cases out there.
(06:21):
It was more abstract.
We were talking about concepts,getting your data ready, ready
strategy, and now in our AImastermind iteration, we're
having more and moreparticipants present because
they're actually like rollingout these projects themselves.
So it's been really neat to seein 75 episodes how far we've
come.
Speaker 1 (06:40):
Yeah, I think we had to with the mastermind.
We had to get the party started.
But now it's raising.
So it's pretty cool.
And that mastermind group isawesome.
So a good friend of ours, maryByers she and I are the leads of
the mastermind.
We host this virtual monthly 90minute session with a small
group of engaged leaders from avariety of different
(07:03):
associations, and if you'reinterested in going deep on AI
once a month, consider joiningthat.
We have information on ourwebsite there as well.
But, yeah, you're right, I meanthis journey that we're on.
We are witnesses to it, as youpointed out, Mallory, and we
hope to be, you know, in variousways obviously, educators and
sources of inspiration, butreally a conduit through which
(07:25):
the community can share theirexperiences with AI.
And so, to build on the point Imade with respect to Innovation
Hub, what we'd love to do ishear more from our listeners,
from the folks who watch us onYouTube.
Give us feedback on thingsyou're doing.
In some cases, we may be ableto bring you on the pod as a
guest or feature something thatyou're doing in an article on
(07:45):
the Sidecar blog, in ournewsletter, etc.
We'd love to hear from you.
That's the most powerful thingyou know.
We can talk about ourperspective as the leaders of
Sidecar and Blue Cypress.
At the end of the day, whatmatters is what you guys are
doing.
Speaker 2 (07:59):
In today's episode, we are talking about AI models.
Is that a surprise?
We're talking about the releaseof Gemini 2.5 Pro, and then
we'll be talking aboutDeepSeek's latest upgraded model
.
So, starting off with Gemini2.5 Pro, it's Google's latest
and most advanced AI modelintroduced this week to the
(08:19):
public.
Gemini 2.5 Pro is part of agreater trend of thinking models
or reasoning models, which areadvanced AI systems designed to
mimic certain aspects of humanthought processes, particularly
in problem solving and logicalreasoning.
These models use complexalgorithms and techniques to
analyze information, drawconclusions and make decisions
(08:40):
based on that analysis decisionsbased on that analysis.
We've chatted at length aboutClaude 3.7, OpenAI's O1, and now
Gemini 2.5 Pro is joining theranks.
The model uses techniques likereinforcement learning and chain
of thought prompting tosimulate reasoning.
This process involves the modelthinking through its responses
by verifying facts and logicallydeducing answers before
(09:02):
providing them With a largecontext window of 1 million
tokens.
Gemini 2.5 Pro can processextensive amounts of data, and
not just text.
It can also process audioimages, videos and large
datasets like entire coderepositories.
The model's agentic codingabilities allow it to create
complex applications like fullyfunctional video games from a
(09:25):
single prompt.
I did a little brief test with2.5 Pro in the Google AI studio
before we recorded this pod, andso far I like it.
It breaks down exactly whatit's thinking, which a lot of
the models that I mentioned doas well.
You all know that we produceblogs from our Sidecar Sync
transcripts.
If you've listened to thepodcast before, we've talked
(09:46):
about this several times.
Typically, my go-to model forthat would be Claude, but I
decided to run a littleexperiment with 2.5 Pro and I
was quite impressed, mostlybecause I like to start that
process by asking identify threetopics from this podcast that
we could write a blog about,instead of just asking it to
generate an entire blog at once.
Identify three topics from thispodcast that we could write a
blog about, instead of justasking it to generate, you know,
an entire blog at once, and itgave me some really compelling
(10:09):
topic ideas, and then it alsoprovided support from the
podcast.
So at minute 10, Amit mentionedthis.
This is why I included it inthis topic, so I thought that
was interesting.
That's something that Claudedoes not typically do when I use
it.
Amit, what are your initialthoughts with 2.5 Pro?
Amit Bhandari.
Speaker 1 (10:27):
I'm pretty impressed with it.
I haven't personally sent it asingle prompt, so I'll disclose
that I intend to over the nextcouple of days.
But I have seen several videosof demos of Gemini 2.5 Pro you,
an AI.
The context window reallyrefers to the amount of
(10:53):
short-term memory it has.
So when you send a prompt inand kind of the history of your
conversation with that model notall conversations, by the way,
but that specific conversationyou're on in chat, gpt, in cloud
or in the Google Studio theaggregation of all of the back
and forth you have with themodel that all has to fit into
(11:14):
what's called a context windowand there's a variety of
techniques for dealing withreally long conversations.
But the idea is that the biggerthe context window in theory,
the more powerful the modelcould be because it has more of
this short-term memory.
So when we talk about a milliontokens a token being
approximately equal to a word,for our purposes that means
(11:36):
that's a lot of words.
That's something on the orderof magnitude of 15 to 20
business books.
It's a lot of content, whereasthe other models that are out
there that are similar inintelligence, like Cloud 3.7 and
GPT-4.0, those tend to belimited to 128,000 tokens, so
about an eighth of the totalcapability of these models,
(11:57):
sorry, of this particular model,gemini 2.5 Pro.
Now, google has been a leaderin long context models for some
time, since the first Geminirelease.
Actually, they had very largecontext windows.
Some of the ways to think aboutthis for associations is that
if we have really complex taskswe want to take on, where we
want to feed in many differentpieces of content let's say,
from our journals or transcriptsfrom conferences, and we want
(12:21):
to be able to look across a lotof content at the same time,
gemini is a tool that stands onits own at the moment because
these other tools are limitedessentially to this fairly small
context window on the one hand.
But remember original ChatGPT,if you recall, had 4K of context
, so 4,000 tokens.
It was very, very limited.
So in any event, the point Iwould make is that by itself is
(12:43):
cool and really more thananything.
That's not new, that's just afeature of Gemini that seems to
be a key differentiator for big,complex pieces of content, but
really the intelligence of themodel is pretty amazing.
So one of the people I follow onYouTube, this guy named Matt
Berman.
If you like slightly moretechnical content.
(13:04):
He's a great YouTuber to follow.
I watch quite a few of hisvideos and he breaks down fairly
complex topics in a really niceway, in my opinion.
Anyway.
So he had this video showcasingusing Gemini 2.5 Pro for coding
, and I tend to look at thoseexamples fairly quickly when a
new model comes out, becausecoding is both it has to
(13:24):
represent, both the ability todo like fairly complex reasoning
but also to understand prettycomplex prompts, especially with
these days.
You know, people are putting inrequests to coding tools like
this to do very complex thingsLike the two examples in his
videos.
One was a Rubik's Cubesimulator, which was essentially
a three-dimensional Rubik'sCube simulator of any number of
(13:47):
dimensions, so it could be threeby three, six by six, a hundred
by a hundred, and the AI wasasked to build the codes that
you could visually representthis.
You could spin it around, youcould see it from any angle, you
could zoom in and out, youcould pan tilt, et cetera, and
it did that.
And then, on top of that, thecode it was requested of the AI
(14:08):
to make it so that the AI itselfcould solve the Rubik's Cube,
so it could randomize the cubestate and then it could solve it
and you can watch it visuallysolving the Rubik's Cube, which
is pretty cool.
What's impressive about this isthat in this video it was a
single shot right, so it wasjust a prompt, and then
immediately had a working pieceof code in a browser in a single
(14:29):
HTML page that did this.
That's a non-trivial bit ofsoftware development to write
that right.
Even a really good developerwould take quite a bit of time
to build something of that orderof magnitude.
The other example was alsosimilarly three dimensional
visual kind of thing in thebrowser to build a Lego
simulator, to be able to snapLego bricks together of any
(14:50):
sizes and shapes and colors, andwhat he demonstrated there was
pretty cool.
So I found that particularexample really compelling, both
because it was clearly showingthe ability to do fairly complex
reasoning with that level ofcoding.
This is not a trivial codingexercise.
Like you know, some simplergames that people have asked
like build a snake game inPython, which is I wouldn't say
(15:16):
that's trivial, but it's fairlysimple.
Comparatively speaking, this isan order of magnitude more
complex.
So that was impressive.
I think your example was greatand when we, when we're talking
about this, just continuousevolution of these models.
The thing we always have topoint out is there's so many
options, right.
So this is now Google gettinginto the game in a way that I
think really puts them more onthe radar for a lot of people.
(15:36):
We talk about OpenAI, we talkabout Claude, we talk about the
open source models, but you know, google's just been kind of
behind, at least in the publicperception and in terms of usage
.
Certainly they've been adistant, maybe not even third
place is what I was going to saymaybe fourth or fifth.
So I think this is going to putthem back on the map for a lot
of people to really consider andthink deeply about.
Speaker 2 (15:59):
And that was one of my follow-up questions too is,
at least from my perspective, itsounds like from yours too.
Google seems to be late to theparty oftentimes when it comes
to AI, but when they arrive tothe party, right, they're
well-dressed, people want tohang out with them, like this
model is really impressive, butit seems late.
So is that kind of Google's MOyou would say, like do they take
(16:20):
more time to bake thingsbecause they want to put out
something really quality?
Speaker 1 (16:25):
I definitely think that is an element of it, that I
think, in fairness to them, Ithink that's a part of what they
need to do because of theirscale and because of their brand
that they want to put thingsout that are fairly well thought
out.
The flip side of it is, I thinkthey were just behind and you
know they were not behind interms of fundamental AI research
.
They've been leaders in that inmany ways for years and years,
(16:48):
and, for those that aren'tfamiliar, google actually
invented the transformerarchitecture, which is the type
of neural network that haspowered all of the language
models you've been hearing about, including the original chat
GPT, and it still powers thevast majority of language models
.
Back in 2017, they inventedthat architecture, and so they
know a thing or two about AI.
(17:08):
These are some really smartfolks with a lot of resources,
but, you know, one of the thingsto maybe consider also is
organizationally.
These guys are the incumbentsin search, and so when they saw
LLM start to scale quickly,they're thinking about their own
business model, and so it allof a sudden made it, you know, a
strategically critical thing tobe in this game.
(17:33):
So I don't know if they'rethinking about it more from the
perspective of how do we use itwithin Google search and other
Google products, and that'stheir number one priority versus
producing models for the restof the world to use, as compared
to OpenAI and Anthropic, whoseonly purpose is to produce
models other people use.
So I don't know if it might bethat perhaps you know it's also
in any organization, no matterhow big and how well resourced
(17:55):
and how smart the people are,you have to, you have to make
priority choices.
I'm entirely speculating aboutthis, but I sense that there's
elements of that going on.
Speaker 2 (18:04):
It seems like Google might need to displace itself
which we've talked about beforeon the pod, this idea of counter
positioning to displace its ownsearch function or at least
greatly change it.
And it's kind of already donethat with AI overviews, which
have gotten better in mypersonal opinion.
But that'll be interesting towatch it play out.
Speaker 1 (18:21):
Definitely.
Yeah, I think that you knowcounter-positioning oneself.
It's better to be Netflix thanBlockbuster, than Blockbuster.
But it's a better idea topotentially say, hey, how can we
be?
You know, how can weessentially be the company that
creates the new business modelthat has superior customer
experience ourselves.
(18:41):
So I definitely see Googleheading in that direction.
You know, my bottom line is thatthere's so much happening in
the area of models that you know, even those of us who spend all
of our time thinking about thisand talking about it and
playing with these models andbuilding software on top of
these models, we can't keep up.
And so just this week, you know, aside from the two models that
we're talking about, with thenew DeepSeek V3 and also Gemini
(19:04):
2.5 Pro, we also had a releasefrom OpenAI with the GPT-4.0 new
form of image generation, andthat is it's a really stunning
capability.
So it's a capability.
If you haven't experimentedwith it since it came out, I
really recommend that you get onChatGPT and ask it to create an
image, or take an existingimage, drop it in and ask it to
(19:26):
modify it.
It's pretty stunning and that'sgotten quite a bit of attention
, but it's hard to keep up withall this stuff.
The advice that I always givepeople about that issue of how
to keep up with this is thatlook for the trend line right,
look for the pattern and lookfor how you would like to use
these models, not just thespecific model.
(19:47):
What are the problems you'retrying to solve today?
If it's a use case, you alreadyhave working totally fine for
the last 12 months with GPT-4,maybe you're not so excited by
like this new model.
That's an even better blog,when the existing blog was
pretty darn good.
Of course, you know those ofyou listening to this podcast
probably are always looking totake the next step and improve
(20:08):
and evolve what you're doing.
But what I also think isinteresting and important is
think about the things youcouldn't do with AI.
So my example where I did havehands on experience in the last
24 hours, mallory is I used thenew Chachi PT 4.0 image
generator to create a comicstrip and I threw it on LinkedIn
(20:28):
.
It's not perfect, it's notgreat, but it just talks about
this conundrum that associationsare in with respect to AMS
replacement.
Right, our favorite topic ofthe pod for a lot of people
other than AI, of course, is AMSreplacement, and you know we're
not trying to, you know, pickon it, but it's a tough, tough
thing to do and it's really longand really expensive and the
(20:49):
value creation oftentimes ismarginal at best.
And so you know why do thatwhen you could do a lot with
those resources experimentingwith AI.
So that's what I just basicallysaid.
What I just said to open AI'smodel yesterday and I got a four
panel comic strip.
Then I said, hey, give me fourmore panels that kind of
conclude, like what happens twoyears later if all you do is
(21:10):
focus in your AMS and don't doAI, so go check out my LinkedIn
if you want to see the comicstrip.
But in the past I've had thisidea to create comic strips or
infographics or whatever, andnone of the image generators
could do it.
So had I not experimented withthis, last night I would still
be thinking in that mindset thatwouldn't it be cool if I had
this creative outlet to createcomic strips or infographics
(21:32):
with just an AI model?
But in my brain I would say, oh, but they do a horrible job
with text.
Oh, but they can't do differentstyles, like a comic strip style
is very different than the kindof you know.
We all know these AI imageshave started to look like for
the last couple of years.
They're very, very similar.
But now my brain is all firedup about, like, all these new
capabilities we have, right, soexperimentation is good, but the
(21:53):
trend line is it's not justwe're excited about OpenAI's
image generator.
All the other ones are going tobe like that within half a
minute, right, you're going tosee the same thing for mid
journey.
Probably within days You'regoing to see the same thing from
all the other multimodal models.
We know Anthropic is no slouchwith their cloud product, so
(22:19):
it's exciting.
So the trend line is imagegeneration really can be used
for things that even, like a dayago or two days ago couldn't be
used.
And that's what I keep lookingfor is what are the next use
cases?
That are the next unlocks?
Speaker 2 (22:25):
Yeah, that's a great point.
Your comic strip was good.
I highly encourage you all togo to Amis LinkedIn.
We can include that in the shownotes as well for you to check
out.
I have never asked Midjourney,which is my preferred image
generator, to do a comic strip,but I'm pretty sure it wouldn't
do a good job with it, based onhow frequently I use it.
So I'm going to use GPT-4'simage generator the next time I
(22:47):
publish a blog, which might betoday, and see what I can come
up with.
Speaker 1 (22:55):
The publish a blog, which might be today, and see
what I can come up with.
The text side of it is reallycompelling because the image
generation, even in the past,when we've had really stunning
images generated by AI, whenthey have attempted to
incorporate text, whether it's asign in the background.
It's always been, you know, kindof garbled up, and even when
you prompted the AI to say donot include text, it would still
oftentimes include text.
So this really does represent apretty significant leap in
image generation and that'suseful in so many respects.
(23:17):
You know, you think of it aswell again, when we're going
from something that was oncescarce to something that is
abundant.
You know, a comic strip wouldtake a lot of work to put
together.
Right, you need talentedillustrators, you need the idea
for the comic strip, you needall this stuff, and I certainly
wouldn't be attempting to dothat.
I really can't do much morethan a stick figure to save my
(23:38):
life, you know.
But I have ideas and so I'd loveto be able to express those
ideas in different ways that areboth interesting and, hopefully
, effective in communicating.
So I find this really, reallyexciting.
So then you take the comicstrip and you say, hey, take an
image to video generator and nowanimate that comic strip and
add audio to it, and you know,so, on and on and on.
(23:58):
This goes right.
Speaker 2 (24:00):
Wow, talking about trend line, I want to zoom out a
little bit.
It seems like we've beentalking a lot lately about these
thinking models that Imentioned, or reasoning models,
which seem to be, in general,pretty large models in terms of
parameter size, but in the priorfew months we spent a lot of
time talking about small models.
So I'm kind of wondering howyou see small models fitting in
(24:21):
to these greater thinking andreasoning model conversations
and do you see like that chainof thought, prompting and
reasoning as something that willbe needed in small models, or
should we just leave that to thebig ones?
Speaker 1 (24:36):
Well, I think that the simulation of the chain of
thought, reasoning, is somethingthat some of the small models
are starting to do in some ways,where they're not really doing
reasoning in the same sense asthe bigger models.
Now, gemini, 2.5, pro, all usethis internal thinking process
(25:02):
where, rather than trying togive you an answer as quickly as
possible, they decide that theproblem is complex enough, that
they're going to break down theproblem into steps and they're
going to solve each of the stepsone at a time, typically
sequentially, although notalways, and then bring back the
results from each of the steps,compile a result and then
evaluate the result and thenpossibly iterate again and again
until the model determines it'sgotten to a good result.
(25:23):
And this process is thisreasoning process, as it's
called is compute, intensive, ittakes more time, it takes more
compute and it producesremarkably improved results
compared to just quickestpossible answers, which is you
know, we've compared it in thepast to the idea of system one
and system two, thinking of theintuitive, reactive thinking
(25:45):
versus the reasoning stylethinking that our biological
brains are doing so well, Ithink.
Coming back to your questionabout small models, will they
incorporate actual reasoningprocesses?
I'd be surprised if they didn'tat some level.
But at the same time, part ofwhat we are seeing is people
distilling into smaller modelsthings that reflect the
(26:06):
intelligence of larger models,and we've seen that from big
LLMs to small LLMs for some timenow, for a couple of years,
where the small models keepgetting smarter and the model
architectures are getting better.
In some ways they're gettingsmarter, but it's also about how
the data set that the smallmodel is being trained on
reflects back on what the bigmodel knows.
(26:29):
So, for example, lama 3.3, the$70 billion parameter kind of
mid-sized model that wasreleased in December is smarter
than the LAMA $405 billionparameter model that was
released last April.
And the way they did that isthere were some advancements in
the tech, but what Meta did isthey used LAMA 405b to generate
a bunch of sample data that theythen trained the new model on.
(26:51):
So this distillation process isa way of really taking the
essence of the intelligence ofthe larger model and packing it
into a much smaller frame, and Ithink you're going to see that
with reasoning.
Specifically, you know, toforeshadow our next topic with
respect to DeepSeek v3, that'sactually a lot of what they did
taking R1's reasoning anddropping some of that
(27:11):
intelligence into anon-reasoning model in v3.
But you're going to see moreand more of this.
What I keep coming back to andthis is particularly important
for our friends in the world ofassociations and nonprofits is
that you know this is a somewhattechnical topic it's important
to understand as a businessleader, because we're at this
forefront of this emergingtechnology.
We want to know what's possible,how we can use it to better
(27:34):
serve our constituents, and onand on.
But eventually and I think thateventually might be in the next
12 to 24 months you're notgoing to be talking too much
about is the model a reasoningmodel or is it a straight
inference model?
I think most of these modelswill have a reasoning ability
and how much of that reasoningability they use or don't use
will be dependent on what youask it to do, and that's
(27:55):
essentially what Cloud 3.7 does,right?
So Cloud 3.7 knows that itneeds to go into thinking mode,
just like you or I would if I'mgiven a really complicated math
problem or something else thattakes, you know, time to reason
through.
I'm going to do that.
I'm not just going to guess atthe answer, which is essentially
what the, you know, the fastinference mode is what language
models have been doing up untilthese reasoning models.
Speaker 2 (28:16):
Yeah, I think you're right.
Looking back and again, this AIjourney that we've been on
since the beginning, I rememberyou and I would talk about
multimodal models and modelsthat can understand text and
audio and images, and I feellike that's just a commonplace
thing now.
So maybe you're right, 12 to 24months we won't even be having
this conversation, but that's agood segue for topic two of
(28:38):
today, which is DeepSeek's V3that Amit mentioned.
We saw this model upgrade thisweek as well from a Chinese AI
startup called DeepSeek, whichwe recently covered in a prior
episode, for its R1 modelreleased.
To give you a quick recap ofthat, its R1 model just cost $6
million to train over 89 timescheaper than OpenAI's rumored
(29:03):
$500 million budget for its O1model and the release of R1 led
to a $1 trillion drop in the USstock market.
So lots of waves were made withDeepSeek and the plot continues
to thicken.
So DeepSeek has now recentlyupgraded its v3 large language
model to DeepSeek v3-0324.
(29:25):
Again, we love these AI modelnames.
This new version is seen asanother competitive attack on
major AI players like OpenAI andAnthropic.
Deepseek's v3 offers enhancedreasoning and coding
capabilities compared to itspredecessor.
The model has 685 billionparameters, up from 671 billion
(29:46):
in the previous version, andit's important to note that
these models are open source, sothe model is available under
the MIT license on platformslike Hugging Face.
Deepseq V3 challenges thedominance of US AI companies by
offering competitive performanceat lower costs, and the model's
release underscores the growingpresence of Chinese AI startups
(30:08):
in the global AI scene, perhapsshifting the balance of power
in tech development.
I feel like Amit with DeepSeekthe things to note are lower
cost, competitive capabilitiesand the fact that it's open
source.
Do you agree with that?
Speaker 1 (30:24):
I think.
So, you know, I think credit isdefinitely due to these folks.
They are brilliant.
The papers they put out theredetail, you know, with a lot of
granularity how they built thesemodels.
So they've open sourced notonly the code and the weights of
the models so anyone can run itanywhere, but they also have
published a ton of researchexplaining how they've advanced
(30:45):
their technology.
So this is not a copycat or aclone.
They have fundamentallyadvanced the science of AI and
they're contributing back to theglobal community.
So I view this as a verypositive thing and I'm hoping to
see this from a lot of otherparts of the world jumping in
because, especially as you see,the resource constraints seem to
decrease over time as peopleget really creative when they
(31:07):
have limited resources, and thatleads to fewer resources being
needed or being perceived asbeing needed Right, and that, in
turn, leads to more innovation,which leads to more and more
growth, and that's that'sexciting.
So to me, I think that that'sone really important thing to
recognize.
One of the things that theseguys have done a really good job
(31:27):
of is they've advanced the waymixture of expert models, or MOE
models, work.
It is a 600 plus billionparameter model, but you can
still run it on fairly.
You know fairly reasonablehardware pretty quickly because
it only has about I think it'slike 32 billion parameters are
active on a per token basis.
So what that means essentiallyis it's really behaves like a 32
(31:50):
billion parameter model andthat's and that's not a
particularly huge model.
It's half the size of Lama3.370 that I just mentioned a
few moments ago.
So that's important becausethey've been able to really
optimize with this mixture ofexperts model and do some other
things around efficiency.
They have a highly performantand really intelligent model.
(32:18):
All the other players that areout there are really paying
close attention to everythingDeepSeek is doing, incorporating
their ideas.
So probably today already we'veseen OpenAI and Anthropic and
Google incorporate DeepSeek'sideas, and on and on right.
And so the proprietary vendorsdon't really share what they're
doing, but I guarantee they'retaking advantage of open source
and the open source communitykeeps on compounding on itself.
So I would point out, the opensource bit probably is the most
(32:42):
important part of it, becausethese things are moving so
quickly.
Speaker 2 (32:47):
And when you say mixture of experts, architecture
, that just means parts of themodel, the only parts that
activate within the model, orthe parts that are needed for
that prompt.
Speaker 1 (32:55):
Exactly so.
You know, if I say, for example, I would like DeepSeek to write
code, maybe there's a portionof that model 600 plus billion
parameters that are focused oncoding, and only those.
You know 30 billion parametersget activated.
If I ask it to, you know, writea poem, it might be a different
section of that.
You know that synthetic brainessentially that gets activated.
(33:17):
So it's a very large brain, soto speak, but it only uses
portions of it at a time andit's not running the whole thing
.
And you know, most of these AIarchitectures are MOE or mixture
of experts models.
What these guys have been doinga good job with is making them
more efficient, making them moreperformance.
Part of what's happening is islike who makes the decision
about which section of the modelshould activate for a given
(33:40):
token, and that's something that, of course, is really important
, because if you don't routethat to the right part of the
net, you don't actually getgreat results.
So I would give these guys alot of props, as I've been doing
, in terms of their scientificadvancements.
You know my point of view interms of open AI, and probably
Anthropic as well, but more soopen AI is that.
You know my point of view interms of open AI, and probably
anthropic as well.
But more so open AI is that youknow you have these companies
(34:02):
that have had I wouldn't sayunlimited, but they've had
substantial resources and theirperception has been that they
need those substantial resourcesto produce these.
You know world-class gains andhere you have a challenger with
far fewer resources, not justmoney, but also they used
equipment that's considerablyless powerful.
Due to export restrictions,they don't have access to the
(34:25):
latest chips, at least that'swhat's been reported.
So you know, that tells you alot.
Again, constraints can beincredibly powerful.
If you give people a verynarrow time frame to do a job, a
lot of times you get a betterresult than if you say you know
how much time you need or if yougive them longer timeframes.
I'm a big fan of setting narrowdeadlines for small chunks of
(34:45):
work.
I don't like saying, hey,what's your, what's your
priority for the next 12 months.
I'd rather know what you'regoing to do in the next week.
Not that I don't think aboutthe next 12 months, but like
it's more about how do you, howdo you put a constraint?
That's near term, and thesmaller the constraints usually
the more creative people get.
Speaker 2 (35:01):
Necessity breeds invention, which we've said on
the pod before.
Amit, I didn't realize this andmaybe you had heard it, but
there was a bill introduced inthe House last month potentially
banning federal employees fromusing a Chinese AI platform,
DeepSeek, on their governmentissued devices.
So I imagine there might besome organizations that would be
dissuaded from using DeepSeekon their government-issued
devices.
So I imagine there might besome organizations that would be
(35:22):
dissuaded from using DeepSeekmodels under the threat of a
potential ban.
But on the flip side, themodels are open source.
So can you explain whether thatis a valid concern or not, or
what to keep in mind there?
Speaker 1 (35:32):
It's totally a valid concern if you're using their
website.
So if you're going todeepseekai, I think is the
website, if you go to thatwebsite, that model that you're
going to deepseekai, I think isthe website that, if you go to
that website, that model thatyou're interacting with is
hosted in China, which is, youknow.
It's not an inherently badthing to have a model hosted in
another country, but the pointwould be that if you're a
(35:52):
government employee and you areasking a model, something
related to you know, whatever itis you're working on, maybe
that isn't the best thing tosend overseas, right?
Maybe that's something weshould be running within the
United States and preferably in,you know, a government cloud of
some sort that's secure.
So I think that's one piece ofit, but that should be separated
from the idea of the modelitself.
(36:14):
It is an open source, openweights model.
It can be reproduced and youknow you can run it anywhere.
You can run it in your own datacenter, you can run it in a
public cloud infrastructure likeAzure, aws, gcp, oracle, et
cetera.
You can run it.
You mentioned Hugging Faceearlier.
They provide a variety ofservices as well.
There's a lot of places you canrun these models and so to ban
(36:37):
the model, I think, is reallymisinformed, because the model
itself is just a piece ofsoftware.
That's totally something youcan inspect.
One of the really nice thingsabout AI models is that if you
say it's open source, by the way, versus open weights, they
sound like similar things.
The model itself.
The actual number of lines ofcode behind these models is
(36:58):
remarkably small.
It's in the single digitthousands typically, or even
smaller, and so the modelsthemselves don't have a lot of
code.
It's all about the weights.
So the weights are, of course,like harder to understand, right
?
They're just a bunch of numbers.
But you if, like someone wasworried about like a backdoor
existing or like phone home youknow phones home and sends your
(37:18):
data back.
That's not a thing Like you canverify that the software is not
doing that.
You can also contain thesemodels in ways when you
inference them in your ownhardware so that there's no
possibility of them doing that.
So I think that the mindsetshould be that we care deeply
about where we inference thesemodels.
That should be done, thoughtabout from a security
perspective, that should bethought from a vendor trust and
(37:40):
vendor reliability perspective,and we should also deeply care
about which models we use.
But you know, whereas I probablywould not sign up to do any
workloads for any of ourproducts in China right now, for
a variety of reasons, one ofwhich is just whether or not
that will continue to beavailable, but also it's a
question of sensitivity,obviously, of the data.
But I'd be totally happy to runthese models as long as they're
(38:04):
inferenced in places where Ibelieve there's, you know, a
better degree of transparency,visibility, control, et cetera.
So you know, we've talked aboutGrok, we've talked about AWS
and Azure Foundry.
You can run most of thesemodels in most of those places.
Speaker 2 (38:18):
Okay, so it wouldn't make sense for the government to
ban source code or like theactual weights.
Speaker 1 (38:24):
I don't believe that's what's being discussed.
I think it's access to thewebsite.
So that's why I think it'simportant to separate using the
model on your own equipment oron equipment that's run by
someone you trust versusconnecting to DeepSeek's website
and using it as a consumer.
I believe that is what the billis intending to ban.
Speaker 2 (38:41):
Mm I want to zoom out again to the trend line because
, as you mentioned with thefirst topic, that's what's
important to keep in mind.
This competition as a wholeseems to be driving AI costs
down, down, down, really quicklyno-transcript.
Speaker 1 (39:28):
Having the knowledge that the costs are going down at
such a rapid rate should openyour mind to the possibilities
of applications that youcouldn't afford previously.
So if you were to say, hey, wewant to go through every piece
of content your organization hasever published and we want to
auto-generate a taxonomy from it, or we want to do other things
with it, right In the past maybethat would have been
(39:50):
prohibitive from a costperspective.
Maybe the quality wasn't greatenough either.
But let's just say that youthink the quality is awesome now
, but the cost might've been afactor where you're like oh,
that would cost us $3 million togo through all that content and
now it doesn't.
Right Now maybe it costs you$3,000 or something like that.
So these shifts in cost shouldopen your mind up to at-scale
(40:13):
applications, doing things atscale.
Or what about all theunstructured data that you have
in your email or elsewhere?
What can you do with that kindof stuff?
So the way I look at it is theapplications.
When cost is going down andperformance and speed are going
up.
It opens up new possibilities.
So that's importantconceptually, because a lot of
what we're talking about in thispod and in the rest of our work
(40:34):
at Sidecar is plotting yourcourse, really your strategic
direction, understanding thetools, understanding the
technology, understanding howthey all fit together, but
thinking about where you shouldgo with this.
So clearly, we want to optimizecurrent business processes.
That's the obvious part, right?
We want to make what we dohappen faster, better, cheaper,
right?
Traditionally, you'd pick maybeone or two of those, but you
(40:57):
don't get all three.
Now you can get all three.
But the bigger question is whatshould you be doing?
As opposed to how do you dowhat you currently do better?
And the what should you bedoing is informed by constraints
, and the constraints are thecapabilities of these AIs, but
it's also the cost.
So knowing that the cost keepsrelentlessly going down should
open your mind to thepossibilities that six months,
(41:18):
12 months, 24 months from now,video generation will be close
to free.
For example, right now, if youwant to do a lot of hours of
video generation with HeyGen,you have to sign up for plans
that will cost you in the fiveor even six figures.
That's a constraint.
We're, by the way, extensivelyusing HeyGen in a variety of our
content production, for our AIlearning hub and for other
(41:39):
things, and we're making theinvestment because we think it's
worth it, but we also know thatit's not really a long-term
recurring investment becausethere is competition in every
one of these dimensions.
So that's part.
One is that I think you canthink bigger and think a little
bit longer term, something thatyou know you wouldn't say oh
well, I was going to invest amillion dollars in a new AMS and
next year it's going to be$100,000 and the year after it's
(42:01):
$10,000.
That's not a thing, right?
That's not the way people thinkand that's not the way these
systems tend to work.
But with AI, that's literallywhat's happening with the
fundamental models, because ofcompetition, because of the
amount of capital being thrownat it.
So that's true for what I'dcall kind of the raw materials
of AI.
It's the fundamental buildingblocks, really the models at
that layer.
(42:21):
It's also happening withinference competition, because
so many people are going afterbeing your cloud provider for AI
, whether it's the hyperscalermajor players like AWS and Azure
, or if it's people who have anew approach to hardware, like
our friends at the Grok which isGROQ, to always repeat that for
clarity or a number of otherfolks that are out there doing
(42:43):
cool things with inference.
That's a hyper competitivemarket.
You're going to keep seeingcosts come down there.
But to come back to the secondpart of your question about like
well, how come we aren't seeingcosts for the application layer
which consumes these?
Ais is for one, it's marketeconomics.
There's fewer competitors thanthere are at the model level.
(43:05):
There's also more complexity interms of integration with
systems.
There's kind of the nuance ofwhat a lot of times is referred
to as the last mile of thesolution actually solving the
problem.
So you can take the model thathas Gemini 2.5 Pro caliber
thinking, but then to wire it upinto your ecosystem and make it
work just right for you.
Still today there arespecialized pieces of software
(43:27):
that have to be integrated.
There's a lot of labor that'srequired.
There's domain knowledge interms of what the best processes
are.
So the actual end solutions, Ithink you are going to see them
come down in cost overall, butit's slower.
It's kind of like if we said,hey, let's just imagine that the
price of steel was close tozero all of a sudden, would that
(43:48):
mean that the price of carsimmediately goes close to zero
as well?
You know, when you're furtherdown in the supply chain, you
know that can happen.
Right.
Cost of materials can go up anddown.
Sometimes it has a rippleeffect and sometimes it takes a
lot longer to work its waythrough the system, but
ultimately this is a really goodthing.
The component costs of theultimate solution obviously are
(44:10):
a major factor in that solution,but more than anything, it's
this abundance mindset that I'mtrying to really hammer into
people's minds that youshouldn't think of AI video
generation or AI whatever as ascarce resource, that you can
only use it in a tiny number ofareas.
You should be thinking that bythe end of this year you'll
probably be able to do it acrosseverything.
(44:30):
So every single blog post thatyou have, let's have a great
video on that post, every singletime, completely AI generated.
Today that might cost you tensof thousands of dollars.
By the end of the year it'llprobably be considerably less.
Speaker 2 (44:45):
I love that.
I call it my version of that.
So you said dream bigger.
I call it often breaking yourbrain, because it's constantly,
every day, all day, having tosay, like what else could I do
with this model, breaking thesebarriers of what you thought was
possible, especially when wetalked about member services.
Right, having a member serviceagent that doesn't just answer
basic questions, that actuallyhas access to your whole
(45:06):
knowledge repository and cananswer domain specific questions
, that's something that you haveto really open your mind to.
So I agree, I think that'sessential and actually, now that
I'm thinking about it, that itcould be an interesting take,
maybe for for Digital Now 2025.
I don't know if we have a themehammered out, but kind of the
idea of dream bigger, breakingyour brain.
Speaker 1 (45:27):
I like that.
I like breaking your brain.
I like that.
I like breaking your brain.
That sounds fun.
Speaker 2 (45:30):
Uh-huh, I don't know.
So maybe you all are the firstto hear our theme for this year.
But with that, everybody, thankyou for tuning in to today's
episode.
I know Amit mentioned Grok witha Q.
We do have a special interviewcoming up.
Not sure when we're going topost it, but in the next few
weeks or so.
So be on the lookout for thatand we will see you all next
week.
Speaker 1 (45:52):
Thanks for tuning into Sidecar Sync this week.
Looking to dive deeper?
Download your free copy of ournew book Ascend Unlocking the
Power of AI for Associations atascendbookorg.
It's packed with insights topower your association's journey
with AI.
And remember, sidecar is herewith more resources, from
webinars to boot camps, to helpyou stay ahead in the
(46:14):
association world.
We'll catch you in the nextepisode.
Until then, keep learning, keepgrowing and keep disrupting.
You shouldn't think of AI video generation or AI
whatever as a scarce resource,that you can only use it in a
tiny number of areas.
You should be thinking that bythe end of this year, you'll
probably be able to do it acrosseverything.
Welcome to Sidecar Sync, yourweekly dose of innovation.
If you're looking for thelatest news, insights and
(00:20):
developments in the associationworld, especially those driven
by artificial intelligence,you're in the right place.
We cut through the noise tobring you the most relevant
updates, with a keen focus onhow AI and other emerging
technologies are shaping thefuture.
No fluff, just facts andinformed discussions.
I'm Amit Nagarajan, chairman ofBlue Cypress, and I'm your host
(00:42):
.
Greetings and welcome to theSidecar Sync, your home for
content at the intersection ofassociations and AI.
My name is Amit Nagarajan.
Speaker 2 (00:53):
And my name is Mallory Mejiaz.
Speaker 1 (00:54):
And we are your hosts .
Today, we are going to coversome crazy and awesome and
exciting things that arehappening at the forefront of
artificial intelligence, andwe're going to tell you how they
might apply to your world as anassociation leader.
Before we do that, though,let's take a moment to hear a
quick word from our sponsor.
Speaker 2 (01:14):
If you're listening to this podcast right now,
you're already thinkingdifferently about AI than many
of your peers, don't you wishthere was a way to showcase your
commitment to innovation andlearning?
The Association AI Professional, or AAIP, certification is
exactly that.
The AAIP certification isawarded to those who have
achieved outstanding theoreticaland practical AI knowledge.
(01:36):
As it pertains to associations,earning your AAIP certification
proves that you're at theforefront of AI in your
organization and in the greaterassociation space, giving you a
competitive edge in anincreasingly AI-driven job
market.
Join the growing group ofprofessionals who've earned
their AAIP certification andsecure your professional future
(01:58):
by heading to learnsidecarai.
Amit, how are you doing today?
Speaker 1 (02:04):
I'm doing great.
I just got back from DC.
I think I may have picked up alittle bit of a cold or
something on the way back, buthere I am.
I'm back in New Orleans anddoing well.
How about yourself?
Speaker 2 (02:14):
I'm doing pretty well myself.
Nothing like a little dose ofsidecar sink to make you feel
better, right?
Speaker 1 (02:20):
That's right.
It's like an adrenaline shot.
It's awesome.
Speaker 2 (02:22):
Exactly.
I have been itching to find outhow the Innovation Hub went.
I was not able to attend and Ihaven't really talked to you yet
, amit, about it, so this is thefirst time I'm hearing.
But how was it?
How did it go?
How was your session?
Speaker 1 (02:36):
It was amazing.
So this is the third annual DCInnovation Hub.
We have the Chicago one comingup in two weeks in Chicago April
8th, and that's also going tobe an amazing event.
It's at the American College ofSurgeons office downtown and
the DC Innovation Hub was at theAmerican Geophysical Union just
north of DuPont Circle in DC.
(02:57):
Beautiful location, amazingconference, wonderful hosts.
We thank our friends at AGU forthat and we had great turnout.
We had about 70 people show up.
This is our small communityevent we do in DC and Chicago
each year.
Our main flagship event, asmany of you know who are
listening, is Digital Now, whichwill be in Chicago this fall.
(03:19):
But back to the Innovation Hubin DC.
It was really a cool moment intime, I think, in a way Mallory,
in that we really are seeingpeople build.
They're past the contemplationstage.
Many of the people that were inthe room were there to listen
and learn, but they were alsothere to share.
That's the idea behind theinnovation hubs.
We started those as informalcommunity gatherings to
(03:41):
certainly share some content andthings that various folks
across our family of companiesare doing, but more than
anything to kind of take a feelof the community and say, hey,
like what's going on, and sopeople are talking about
deploying AI in a lot ofinteresting ways.
So it was super fun.
I learned a ton, met some greatpeople I hadn't had a chance to
meet in person before.
All around, Really well worthit, and I hope those of you that
(04:03):
are listening that are evensomewhat close to Chicago
consider joining us on April 8th.
We still have a little bit ofroom left.
Speaker 2 (04:10):
Were there any key takeaways or any like challenges
, any patterns that you noticedemerge of all the associations
there said, oh, we're reallystruggling with X, anything like
that.
Speaker 1 (04:21):
You know, I think, more than anything, people seem
ready.
So even last year at this time,the way I felt was that people
were still nervous, they werecontemplating, they were
learning, but they were stillkind of like, hmm, I don't know
if we should do this.
I'm not sure if AI is ready tosupport the critical work of our
association at scale.
(04:42):
A lot of people a year ago werealready, you know, doing
personal experiments andactually, in some cases,
significant work with tools likeClaude and ChatGPT.
But to deploy it as anorganization, right to go out
into the world and say this isour association's AI for
knowledge or this is ourassociation's tool for search or
personalization, few were doingthat.
(05:02):
And this time around, what Ifelt was that people were saying
, yeah, we are doing this.
It wasn't a question of if, andI think that's exciting because
you know AI is not going towait for any of us, whether
we're, you know, do-gooders inthe not-for-profit sector or
somebody else.
Like AI is not hanging out andjust saying, hey, we'll wait for
you however long you want.
So I was excited to see peopletaking action, because that's
(05:24):
really the learning loop, right.
We talk at Sidecar all the timeabout the AI learning journey.
It's not an event.
It's a continuous and foreverprocess, and that's true in any
domain.
It's definitely true in AI, andwhen you go and do the thing
yourself, that's when you reallylearn and you build
organizational reps.
That leads to organizationalstrength, obviously leads to
culture change.
It's cool, so I was reallypumped up about that.
(05:50):
We heard about people talkingabout a lot of different kinds
of AI as well, so some peopledoing interesting analytics
stuff, some people are doingpersonalization at scale, some
people are doing knowledgeassistant work.
So it was a lot of fun.
Speaker 2 (05:59):
That's awesome.
I feel like you and I, andthrough the podcast, through
Sidecar, getting to witness thisjourney from its inception all
the way to now, all the way intothe future.
I'm thinking of the AIMastermind group that Sidecar
runs for Association C-suiteleaders and thinking how, in its
early stages, the AI Mastermindgroup we had to lead most of
the sessions because there justweren't any use cases out there.
(06:21):
It was more abstract.
We were talking about concepts,getting your data ready, ready
strategy, and now in our AImastermind iteration, we're
having more and moreparticipants present because
they're actually like rollingout these projects themselves.
So it's been really neat to seein 75 episodes how far we've
come.
Speaker 1 (06:40):
Yeah, I think we had to with the mastermind.
We had to get the party started.
But now it's raising.
So it's pretty cool.
And that mastermind group isawesome.
So a good friend of ours, maryByers she and I are the leads of
the mastermind.
We host this virtual monthly 90minute session with a small
group of engaged leaders from avariety of different
(07:03):
associations, and if you'reinterested in going deep on AI
once a month, consider joiningthat.
We have information on ourwebsite there as well.
But, yeah, you're right, I meanthis journey that we're on.
We are witnesses to it, as youpointed out, Mallory, and we
hope to be, you know, in variousways obviously, educators and
sources of inspiration, butreally a conduit through which
(07:25):
the community can share theirexperiences with AI.
And so, to build on the point Imade with respect to Innovation
Hub, what we'd love to do ishear more from our listeners,
from the folks who watch us onYouTube.
Give us feedback on thingsyou're doing.
In some cases, we may be ableto bring you on the pod as a
guest or feature something thatyou're doing in an article on
(07:45):
the Sidecar blog, in ournewsletter, etc.
We'd love to hear from you.
That's the most powerful thingyou know.
We can talk about ourperspective as the leaders of
Sidecar and Blue Cypress.
At the end of the day, whatmatters is what you guys are
doing.
Speaker 2 (07:59):
In today's episode, we are talking about AI models.
Is that a surprise?
We're talking about the releaseof Gemini 2.5 Pro, and then
we'll be talking aboutDeepSeek's latest upgraded model
.
So, starting off with Gemini2.5 Pro, it's Google's latest
and most advanced AI modelintroduced this week to the
(08:19):
public.
Gemini 2.5 Pro is part of agreater trend of thinking models
or reasoning models, which areadvanced AI systems designed to
mimic certain aspects of humanthought processes, particularly
in problem solving and logicalreasoning.
These models use complexalgorithms and techniques to
analyze information, drawconclusions and make decisions
(08:40):
based on that analysis decisionsbased on that analysis.
We've chatted at length aboutClaude 3.7, OpenAI's O1, and now
Gemini 2.5 Pro is joining theranks.
The model uses techniques likereinforcement learning and chain
of thought prompting tosimulate reasoning.
This process involves the modelthinking through its responses
by verifying facts and logicallydeducing answers before
(09:02):
providing them With a largecontext window of 1 million
tokens.
Gemini 2.5 Pro can processextensive amounts of data, and
not just text.
It can also process audioimages, videos and large
datasets like entire coderepositories.
The model's agentic codingabilities allow it to create
complex applications like fullyfunctional video games from a
(09:25):
single prompt.
I did a little brief test with2.5 Pro in the Google AI studio
before we recorded this pod, andso far I like it.
It breaks down exactly whatit's thinking, which a lot of
the models that I mentioned doas well.
You all know that we produceblogs from our Sidecar Sync
transcripts.
If you've listened to thepodcast before, we've talked
(09:46):
about this several times.
Typically, my go-to model forthat would be Claude, but I
decided to run a littleexperiment with 2.5 Pro and I
was quite impressed, mostlybecause I like to start that
process by asking identify threetopics from this podcast that
we could write a blog about,instead of just asking it to
generate an entire blog at once.
Identify three topics from thispodcast that we could write a
blog about, instead of justasking it to generate, you know,
an entire blog at once, and itgave me some really compelling
(10:09):
topic ideas, and then it alsoprovided support from the
podcast.
So at minute 10, Amit mentionedthis.
This is why I included it inthis topic, so I thought that
was interesting.
That's something that Claudedoes not typically do when I use
it.
Amit, what are your initialthoughts with 2.5 Pro?
Amit Bhandari.
Speaker 1 (10:27):
I'm pretty impressed with it.
I haven't personally sent it asingle prompt, so I'll disclose
that I intend to over the nextcouple of days.
But I have seen several videosof demos of Gemini 2.5 Pro you,
an AI.
The context window reallyrefers to the amount of
(10:53):
short-term memory it has.
So when you send a prompt inand kind of the history of your
conversation with that model notall conversations, by the way,
but that specific conversationyou're on in chat, gpt, in cloud
or in the Google Studio theaggregation of all of the back
and forth you have with themodel that all has to fit into
(11:14):
what's called a context windowand there's a variety of
techniques for dealing withreally long conversations.
But the idea is that the biggerthe context window in theory,
the more powerful the modelcould be because it has more of
this short-term memory.
So when we talk about a milliontokens a token being
approximately equal to a word,for our purposes that means
(11:36):
that's a lot of words.
That's something on the orderof magnitude of 15 to 20
business books.
It's a lot of content, whereasthe other models that are out
there that are similar inintelligence, like Cloud 3.7 and
GPT-4.0, those tend to belimited to 128,000 tokens, so
about an eighth of the totalcapability of these models,
(11:57):
sorry, of this particular model,gemini 2.5 Pro.
Now, google has been a leaderin long context models for some
time, since the first Geminirelease.
Actually, they had very largecontext windows.
Some of the ways to think aboutthis for associations is that
if we have really complex taskswe want to take on, where we
want to feed in many differentpieces of content let's say,
from our journals or transcriptsfrom conferences, and we want
(12:21):
to be able to look across a lotof content at the same time,
gemini is a tool that stands onits own at the moment because
these other tools are limitedessentially to this fairly small
context window on the one hand.
But remember original ChatGPT,if you recall, had 4K of context
, so 4,000 tokens.
It was very, very limited.
So in any event, the point Iwould make is that by itself is
(12:43):
cool and really more thananything.
That's not new, that's just afeature of Gemini that seems to
be a key differentiator for big,complex pieces of content, but
really the intelligence of themodel is pretty amazing.
So one of the people I follow onYouTube, this guy named Matt
Berman.
If you like slightly moretechnical content.
(13:04):
He's a great YouTuber to follow.
I watch quite a few of hisvideos and he breaks down fairly
complex topics in a really niceway, in my opinion.
Anyway.
So he had this video showcasingusing Gemini 2.5 Pro for coding
, and I tend to look at thoseexamples fairly quickly when a
new model comes out, becausecoding is both it has to
(13:24):
represent, both the ability todo like fairly complex reasoning
but also to understand prettycomplex prompts, especially with
these days.
You know, people are putting inrequests to coding tools like
this to do very complex thingsLike the two examples in his
videos.
One was a Rubik's Cubesimulator, which was essentially
a three-dimensional Rubik'sCube simulator of any number of
(13:47):
dimensions, so it could be threeby three, six by six, a hundred
by a hundred, and the AI wasasked to build the codes that
you could visually representthis.
You could spin it around, youcould see it from any angle, you
could zoom in and out, youcould pan tilt, et cetera, and
it did that.
And then, on top of that, thecode it was requested of the AI
(14:08):
to make it so that the AI itselfcould solve the Rubik's Cube,
so it could randomize the cubestate and then it could solve it
and you can watch it visuallysolving the Rubik's Cube, which
is pretty cool.
What's impressive about this isthat in this video it was a
single shot right, so it wasjust a prompt, and then
immediately had a working pieceof code in a browser in a single
(14:29):
HTML page that did this.
That's a non-trivial bit ofsoftware development to write
that right.
Even a really good developerwould take quite a bit of time
to build something of that orderof magnitude.
The other example was alsosimilarly three dimensional
visual kind of thing in thebrowser to build a Lego
simulator, to be able to snapLego bricks together of any
(14:50):
sizes and shapes and colors, andwhat he demonstrated there was
pretty cool.
So I found that particularexample really compelling, both
because it was clearly showingthe ability to do fairly complex
reasoning with that level ofcoding.
This is not a trivial codingexercise.
Like you know, some simplergames that people have asked
like build a snake game inPython, which is I wouldn't say
(15:16):
that's trivial, but it's fairlysimple.
Comparatively speaking, this isan order of magnitude more
complex.
So that was impressive.
I think your example was greatand when we, when we're talking
about this, just continuousevolution of these models.
The thing we always have topoint out is there's so many
options, right.
So this is now Google gettinginto the game in a way that I
think really puts them more onthe radar for a lot of people.
(15:36):
We talk about OpenAI, we talkabout Claude, we talk about the
open source models, but you know, google's just been kind of
behind, at least in the publicperception and in terms of usage
.
Certainly they've been adistant, maybe not even third
place is what I was going to saymaybe fourth or fifth.
So I think this is going to putthem back on the map for a lot
of people to really consider andthink deeply about.
Speaker 2 (15:59):
And that was one of my follow-up questions too is,
at least from my perspective, itsounds like from yours too.
Google seems to be late to theparty oftentimes when it comes
to AI, but when they arrive tothe party, right, they're
well-dressed, people want tohang out with them, like this
model is really impressive, butit seems late.
So is that kind of Google's MOyou would say, like do they take
(16:20):
more time to bake thingsbecause they want to put out
something really quality?
Speaker 1 (16:25):
I definitely think that is an element of it, that I
think, in fairness to them, Ithink that's a part of what they
need to do because of theirscale and because of their brand
that they want to put thingsout that are fairly well thought
out.
The flip side of it is, I thinkthey were just behind and you
know they were not behind interms of fundamental AI research
.
They've been leaders in that inmany ways for years and years,
(16:48):
and, for those that aren'tfamiliar, google actually
invented the transformerarchitecture, which is the type
of neural network that haspowered all of the language
models you've been hearing about, including the original chat
GPT, and it still powers thevast majority of language models
.
Back in 2017, they inventedthat architecture, and so they
know a thing or two about AI.
(17:08):
These are some really smartfolks with a lot of resources,
but, you know, one of the thingsto maybe consider also is
organizationally.
These guys are the incumbentsin search, and so when they saw
LLM start to scale quickly,they're thinking about their own
business model, and so it allof a sudden made it, you know, a
strategically critical thing tobe in this game.
(17:33):
So I don't know if they'rethinking about it more from the
perspective of how do we use itwithin Google search and other
Google products, and that'stheir number one priority versus
producing models for the restof the world to use, as compared
to OpenAI and Anthropic, whoseonly purpose is to produce
models other people use.
So I don't know if it might bethat perhaps you know it's also
in any organization, no matterhow big and how well resourced
(17:55):
and how smart the people are,you have to, you have to make
priority choices.
I'm entirely speculating aboutthis, but I sense that there's
elements of that going on.
Speaker 2 (18:04):
It seems like Google might need to displace itself
which we've talked about beforeon the pod, this idea of counter
positioning to displace its ownsearch function or at least
greatly change it.
And it's kind of already donethat with AI overviews, which
have gotten better in mypersonal opinion.
But that'll be interesting towatch it play out.
Speaker 1 (18:21):
Definitely.
Yeah, I think that you knowcounter-positioning oneself.
It's better to be Netflix thanBlockbuster, than Blockbuster.
But it's a better idea topotentially say, hey, how can we
be?
You know, how can weessentially be the company that
creates the new business modelthat has superior customer
experience ourselves.
(18:41):
So I definitely see Googleheading in that direction.
You know, my bottom line is thatthere's so much happening in
the area of models that you know, even those of us who spend all
of our time thinking about thisand talking about it and
playing with these models andbuilding software on top of
these models, we can't keep up.
And so just this week, you know, aside from the two models that
we're talking about, with thenew DeepSeek V3 and also Gemini
(19:04):
2.5 Pro, we also had a releasefrom OpenAI with the GPT-4.0 new
form of image generation, andthat is it's a really stunning
capability.
So it's a capability.
If you haven't experimentedwith it since it came out, I
really recommend that you get onChatGPT and ask it to create an
image, or take an existingimage, drop it in and ask it to
(19:26):
modify it.
It's pretty stunning and that'sgotten quite a bit of attention
, but it's hard to keep up withall this stuff.
The advice that I always givepeople about that issue of how
to keep up with this is thatlook for the trend line right,
look for the pattern and lookfor how you would like to use
these models, not just thespecific model.
(19:47):
What are the problems you'retrying to solve today?
If it's a use case, you alreadyhave working totally fine for
the last 12 months with GPT-4,maybe you're not so excited by
like this new model.
That's an even better blog,when the existing blog was
pretty darn good.
Of course, you know those ofyou listening to this podcast
probably are always looking totake the next step and improve
(20:08):
and evolve what you're doing.
But what I also think isinteresting and important is
think about the things youcouldn't do with AI.
So my example where I did havehands on experience in the last
24 hours, mallory is I used thenew Chachi PT 4.0 image
generator to create a comicstrip and I threw it on LinkedIn
(20:28):
.
It's not perfect, it's notgreat, but it just talks about
this conundrum that associationsare in with respect to AMS
replacement.
Right, our favorite topic ofthe pod for a lot of people
other than AI, of course, is AMSreplacement, and you know we're
not trying to, you know, pickon it, but it's a tough, tough
thing to do and it's really longand really expensive and the
(20:49):
value creation oftentimes ismarginal at best.
And so you know why do thatwhen you could do a lot with
those resources experimentingwith AI.
So that's what I just basicallysaid.
What I just said to open AI'smodel yesterday and I got a four
panel comic strip.
Then I said, hey, give me fourmore panels that kind of
conclude, like what happens twoyears later if all you do is
(21:10):
focus in your AMS and don't doAI, so go check out my LinkedIn
if you want to see the comicstrip.
But in the past I've had thisidea to create comic strips or
infographics or whatever, andnone of the image generators
could do it.
So had I not experimented withthis, last night I would still
be thinking in that mindset thatwouldn't it be cool if I had
this creative outlet to createcomic strips or infographics
(21:32):
with just an AI model?
But in my brain I would say, oh, but they do a horrible job
with text.
Oh, but they can't do differentstyles, like a comic strip style
is very different than the kindof you know.
We all know these AI imageshave started to look like for
the last couple of years.
They're very, very similar.
But now my brain is all firedup about, like, all these new
capabilities we have, right, soexperimentation is good, but the
(21:53):
trend line is it's not justwe're excited about OpenAI's
image generator.
All the other ones are going tobe like that within half a
minute, right, you're going tosee the same thing for mid
journey.
Probably within days You'regoing to see the same thing from
all the other multimodal models.
We know Anthropic is no slouchwith their cloud product, so
(22:19):
it's exciting.
So the trend line is imagegeneration really can be used
for things that even, like a dayago or two days ago couldn't be
used.
And that's what I keep lookingfor is what are the next use
cases?
That are the next unlocks?
Speaker 2 (22:25):
Yeah, that's a great point.
Your comic strip was good.
I highly encourage you all togo to Amis LinkedIn.
We can include that in the shownotes as well for you to check
out.
I have never asked Midjourney,which is my preferred image
generator, to do a comic strip,but I'm pretty sure it wouldn't
do a good job with it, based onhow frequently I use it.
So I'm going to use GPT-4'simage generator the next time I
(22:47):
publish a blog, which might betoday, and see what I can come
up with.
Speaker 1 (22:55):
The publish a blog, which might be today, and see
what I can come up with.
The text side of it is reallycompelling because the image
generation, even in the past,when we've had really stunning
images generated by AI, whenthey have attempted to
incorporate text, whether it's asign in the background.
It's always been, you know, kindof garbled up, and even when
you prompted the AI to say donot include text, it would still
oftentimes include text.
So this really does represent apretty significant leap in
image generation and that'suseful in so many respects.
(23:17):
You know, you think of it aswell again, when we're going
from something that was oncescarce to something that is
abundant.
You know, a comic strip wouldtake a lot of work to put
together.
Right, you need talentedillustrators, you need the idea
for the comic strip, you needall this stuff, and I certainly
wouldn't be attempting to dothat.
I really can't do much morethan a stick figure to save my
(23:38):
life, you know.
But I have ideas and so I'd loveto be able to express those
ideas in different ways that areboth interesting and, hopefully
, effective in communicating.
So I find this really, reallyexciting.
So then you take the comicstrip and you say, hey, take an
image to video generator and nowanimate that comic strip and
add audio to it, and you know,so, on and on and on.
(23:58):
This goes right.
Speaker 2 (24:00):
Wow, talking about trend line, I want to zoom out a
little bit.
It seems like we've beentalking a lot lately about these
thinking models that Imentioned, or reasoning models,
which seem to be, in general,pretty large models in terms of
parameter size, but in the priorfew months we spent a lot of
time talking about small models.
So I'm kind of wondering howyou see small models fitting in
(24:21):
to these greater thinking andreasoning model conversations
and do you see like that chainof thought, prompting and
reasoning as something that willbe needed in small models, or
should we just leave that to thebig ones?
Speaker 1 (24:36):
Well, I think that the simulation of the chain of
thought, reasoning, is somethingthat some of the small models
are starting to do in some ways,where they're not really doing
reasoning in the same sense asthe bigger models.
Now, gemini, 2.5, pro, all usethis internal thinking process
(25:02):
where, rather than trying togive you an answer as quickly as
possible, they decide that theproblem is complex enough, that
they're going to break down theproblem into steps and they're
going to solve each of the stepsone at a time, typically
sequentially, although notalways, and then bring back the
results from each of the steps,compile a result and then
evaluate the result and thenpossibly iterate again and again
until the model determines it'sgotten to a good result.
(25:23):
And this process is thisreasoning process, as it's
called is compute, intensive, ittakes more time, it takes more
compute and it producesremarkably improved results
compared to just quickestpossible answers, which is you
know, we've compared it in thepast to the idea of system one
and system two, thinking of theintuitive, reactive thinking
(25:45):
versus the reasoning stylethinking that our biological
brains are doing so well, Ithink.
Coming back to your questionabout small models, will they
incorporate actual reasoningprocesses?
I'd be surprised if they didn'tat some level.
But at the same time, part ofwhat we are seeing is people
distilling into smaller modelsthings that reflect the
(26:06):
intelligence of larger models,and we've seen that from big
LLMs to small LLMs for some timenow, for a couple of years,
where the small models keepgetting smarter and the model
architectures are getting better.
In some ways they're gettingsmarter, but it's also about how
the data set that the smallmodel is being trained on
reflects back on what the bigmodel knows.
(26:29):
So, for example, lama 3.3, the$70 billion parameter kind of
mid-sized model that wasreleased in December is smarter
than the LAMA $405 billionparameter model that was
released last April.
And the way they did that isthere were some advancements in
the tech, but what Meta did isthey used LAMA 405b to generate
a bunch of sample data that theythen trained the new model on.
(26:51):
So this distillation process isa way of really taking the
essence of the intelligence ofthe larger model and packing it
into a much smaller frame, and Ithink you're going to see that
with reasoning.
Specifically, you know, toforeshadow our next topic with
respect to DeepSeek v3, that'sactually a lot of what they did
taking R1's reasoning anddropping some of that
(27:11):
intelligence into anon-reasoning model in v3.
But you're going to see moreand more of this.
What I keep coming back to andthis is particularly important
for our friends in the world ofassociations and nonprofits is
that you know this is a somewhattechnical topic it's important
to understand as a businessleader, because we're at this
forefront of this emergingtechnology.
We want to know what's possible,how we can use it to better
(27:34):
serve our constituents, and onand on.
But eventually and I think thateventually might be in the next
12 to 24 months you're notgoing to be talking too much
about is the model a reasoningmodel or is it a straight
inference model?
I think most of these modelswill have a reasoning ability
and how much of that reasoningability they use or don't use
will be dependent on what youask it to do, and that's
(27:55):
essentially what Cloud 3.7 does,right?
So Cloud 3.7 knows that itneeds to go into thinking mode,
just like you or I would if I'mgiven a really complicated math
problem or something else thattakes, you know, time to reason
through.
I'm going to do that.
I'm not just going to guess atthe answer, which is essentially
what the, you know, the fastinference mode is what language
models have been doing up untilthese reasoning models.
Speaker 2 (28:16):
Yeah, I think you're right.
Looking back and again, this AIjourney that we've been on
since the beginning, I rememberyou and I would talk about
multimodal models and modelsthat can understand text and
audio and images, and I feellike that's just a commonplace
thing now.
So maybe you're right, 12 to 24months we won't even be having
this conversation, but that's agood segue for topic two of
(28:38):
today, which is DeepSeek's V3that Amit mentioned.
We saw this model upgrade thisweek as well from a Chinese AI
startup called DeepSeek, whichwe recently covered in a prior
episode, for its R1 modelreleased.
To give you a quick recap ofthat, its R1 model just cost $6
million to train over 89 timescheaper than OpenAI's rumored
(29:03):
$500 million budget for its O1model and the release of R1 led
to a $1 trillion drop in the USstock market.
So lots of waves were made withDeepSeek and the plot continues
to thicken.
So DeepSeek has now recentlyupgraded its v3 large language
model to DeepSeek v3-0324.
(29:25):
Again, we love these AI modelnames.
This new version is seen asanother competitive attack on
major AI players like OpenAI andAnthropic.
Deepseek's v3 offers enhancedreasoning and coding
capabilities compared to itspredecessor.
The model has 685 billionparameters, up from 671 billion
(29:46):
in the previous version, andit's important to note that
these models are open source, sothe model is available under
the MIT license on platformslike Hugging Face.
Deepseq V3 challenges thedominance of US AI companies by
offering competitive performanceat lower costs, and the model's
release underscores the growingpresence of Chinese AI startups
(30:08):
in the global AI scene, perhapsshifting the balance of power
in tech development.
I feel like Amit with DeepSeekthe things to note are lower
cost, competitive capabilitiesand the fact that it's open
source.
Do you agree with that?
Speaker 1 (30:24):
I think.
So, you know, I think credit isdefinitely due to these folks.
They are brilliant.
The papers they put out theredetail, you know, with a lot of
granularity how they built thesemodels.
So they've open sourced notonly the code and the weights of
the models so anyone can run itanywhere, but they also have
published a ton of researchexplaining how they've advanced
(30:45):
their technology.
So this is not a copycat or aclone.
They have fundamentallyadvanced the science of AI and
they're contributing back to theglobal community.
So I view this as a verypositive thing and I'm hoping to
see this from a lot of otherparts of the world jumping in
because, especially as you see,the resource constraints seem to
decrease over time as peopleget really creative when they
(31:07):
have limited resources, and thatleads to fewer resources being
needed or being perceived asbeing needed Right, and that, in
turn, leads to more innovation,which leads to more and more
growth, and that's that'sexciting.
So to me, I think that that'sone really important thing to
recognize.
One of the things that theseguys have done a really good job
(31:27):
of is they've advanced the waymixture of expert models, or MOE
models, work.
It is a 600 plus billionparameter model, but you can
still run it on fairly.
You know fairly reasonablehardware pretty quickly because
it only has about I think it'slike 32 billion parameters are
active on a per token basis.
So what that means essentiallyis it's really behaves like a 32
(31:50):
billion parameter model andthat's and that's not a
particularly huge model.
It's half the size of Lama3.370 that I just mentioned a
few moments ago.
So that's important becausethey've been able to really
optimize with this mixture ofexperts model and do some other
things around efficiency.
They have a highly performantand really intelligent model.
(32:18):
All the other players that areout there are really paying
close attention to everythingDeepSeek is doing, incorporating
their ideas.
So probably today already we'veseen OpenAI and Anthropic and
Google incorporate DeepSeek'sideas, and on and on right.
And so the proprietary vendorsdon't really share what they're
doing, but I guarantee they'retaking advantage of open source
and the open source communitykeeps on compounding on itself.
So I would point out, the opensource bit probably is the most
(32:42):
important part of it, becausethese things are moving so
quickly.
Speaker 2 (32:47):
And when you say mixture of experts, architecture
, that just means parts of themodel, the only parts that
activate within the model, orthe parts that are needed for
that prompt.
Speaker 1 (32:55):
Exactly so.
You know, if I say, for example, I would like DeepSeek to write
code, maybe there's a portionof that model 600 plus billion
parameters that are focused oncoding, and only those.
You know 30 billion parametersget activated.
If I ask it to, you know, writea poem, it might be a different
section of that.
You know that synthetic brainessentially that gets activated.
(33:17):
So it's a very large brain, soto speak, but it only uses
portions of it at a time andit's not running the whole thing
.
And you know, most of these AIarchitectures are MOE or mixture
of experts models.
What these guys have been doinga good job with is making them
more efficient, making them moreperformance.
Part of what's happening is islike who makes the decision
about which section of the modelshould activate for a given
(33:40):
token, and that's something that, of course, is really important
, because if you don't routethat to the right part of the
net, you don't actually getgreat results.
So I would give these guys alot of props, as I've been doing
, in terms of their scientificadvancements.
You know my point of view interms of open AI, and probably
Anthropic as well, but more soopen AI is that.
You know my point of view interms of open AI, and probably
anthropic as well.
But more so open AI is that youknow you have these companies
(34:02):
that have had I wouldn't sayunlimited, but they've had
substantial resources and theirperception has been that they
need those substantial resourcesto produce these.
You know world-class gains andhere you have a challenger with
far fewer resources, not justmoney, but also they used
equipment that's considerablyless powerful.
Due to export restrictions,they don't have access to the
(34:25):
latest chips, at least that'swhat's been reported.
So you know, that tells you alot.
Again, constraints can beincredibly powerful.
If you give people a verynarrow time frame to do a job, a
lot of times you get a betterresult than if you say you know
how much time you need or if yougive them longer timeframes.
I'm a big fan of setting narrowdeadlines for small chunks of
(34:45):
work.
I don't like saying, hey,what's your, what's your
priority for the next 12 months.
I'd rather know what you'regoing to do in the next week.
Not that I don't think aboutthe next 12 months, but like
it's more about how do you, howdo you put a constraint?
That's near term, and thesmaller the constraints usually
the more creative people get.
Speaker 2 (35:01):
Necessity breeds invention, which we've said on
the pod before.
Amit, I didn't realize this andmaybe you had heard it, but
there was a bill introduced inthe House last month potentially
banning federal employees fromusing a Chinese AI platform,
DeepSeek, on their governmentissued devices.
So I imagine there might besome organizations that would be
dissuaded from using DeepSeekon their government-issued
devices.
So I imagine there might besome organizations that would be
(35:22):
dissuaded from using DeepSeekmodels under the threat of a
potential ban.
But on the flip side, themodels are open source.
So can you explain whether thatis a valid concern or not, or
what to keep in mind there?
Speaker 1 (35:32):
It's totally a valid concern if you're using their
website.
So if you're going todeepseekai, I think is the
website, if you go to thatwebsite, that model that you're
going to deepseekai, I think isthe website that, if you go to
that website, that model thatyou're interacting with is
hosted in China, which is, youknow.
It's not an inherently badthing to have a model hosted in
another country, but the pointwould be that if you're a
(35:52):
government employee and you areasking a model, something
related to you know, whatever itis you're working on, maybe
that isn't the best thing tosend overseas, right?
Maybe that's something weshould be running within the
United States and preferably in,you know, a government cloud of
some sort that's secure.
So I think that's one piece ofit, but that should be separated
from the idea of the modelitself.
(36:14):
It is an open source, openweights model.
It can be reproduced and youknow you can run it anywhere.
You can run it in your own datacenter, you can run it in a
public cloud infrastructure likeAzure, aws, gcp, oracle, et
cetera.
You can run it.
You mentioned Hugging Faceearlier.
They provide a variety ofservices as well.
There's a lot of places you canrun these models and so to ban
(36:37):
the model, I think, is reallymisinformed, because the model
itself is just a piece ofsoftware.
That's totally something youcan inspect.
One of the really nice thingsabout AI models is that if you
say it's open source, by the way, versus open weights, they
sound like similar things.
The model itself.
The actual number of lines ofcode behind these models is
(36:58):
remarkably small.
It's in the single digitthousands typically, or even
smaller, and so the modelsthemselves don't have a lot of
code.
It's all about the weights.
So the weights are, of course,like harder to understand, right
?
They're just a bunch of numbers.
But you if, like someone wasworried about like a backdoor
existing or like phone home youknow phones home and sends your
(37:18):
data back.
That's not a thing Like you canverify that the software is not
doing that.
You can also contain thesemodels in ways when you
inference them in your ownhardware so that there's no
possibility of them doing that.
So I think that the mindsetshould be that we care deeply
about where we inference thesemodels.
That should be done, thoughtabout from a security
perspective, that should bethought from a vendor trust and
(37:40):
vendor reliability perspective,and we should also deeply care
about which models we use.
But you know, whereas I probablywould not sign up to do any
workloads for any of ourproducts in China right now, for
a variety of reasons, one ofwhich is just whether or not
that will continue to beavailable, but also it's a
question of sensitivity,obviously, of the data.
But I'd be totally happy to runthese models as long as they're
(38:04):
inferenced in places where Ibelieve there's, you know, a
better degree of transparency,visibility, control, et cetera.
So you know, we've talked aboutGrok, we've talked about AWS
and Azure Foundry.
You can run most of thesemodels in most of those places.
Speaker 2 (38:18):
Okay, so it wouldn't make sense for the government to
ban source code or like theactual weights.
Speaker 1 (38:24):
I don't believe that's what's being discussed.
I think it's access to thewebsite.
So that's why I think it'simportant to separate using the
model on your own equipment oron equipment that's run by
someone you trust versusconnecting to DeepSeek's website
and using it as a consumer.
I believe that is what the billis intending to ban.
Speaker 2 (38:41):
Mm I want to zoom out again to the trend line because
, as you mentioned with thefirst topic, that's what's
important to keep in mind.
This competition as a wholeseems to be driving AI costs
down, down, down, really quicklyno-transcript.
Speaker 1 (39:28):
Having the knowledge that the costs are going down at
such a rapid rate should openyour mind to the possibilities
of applications that youcouldn't afford previously.
So if you were to say, hey, wewant to go through every piece
of content your organization hasever published and we want to
auto-generate a taxonomy from it, or we want to do other things
with it, right In the past maybethat would have been
(39:50):
prohibitive from a costperspective.
Maybe the quality wasn't greatenough either.
But let's just say that youthink the quality is awesome now
, but the cost might've been afactor where you're like oh,
that would cost us $3 million togo through all that content and
now it doesn't.
Right Now maybe it costs you$3,000 or something like that.
So these shifts in cost shouldopen your mind up to at-scale
(40:13):
applications, doing things atscale.
Or what about all theunstructured data that you have
in your email or elsewhere?
What can you do with that kindof stuff?
So the way I look at it is theapplications.
When cost is going down andperformance and speed are going
up.
It opens up new possibilities.
So that's importantconceptually, because a lot of
what we're talking about in thispod and in the rest of our work
(40:34):
at Sidecar is plotting yourcourse, really your strategic
direction, understanding thetools, understanding the
technology, understanding howthey all fit together, but
thinking about where you shouldgo with this.
So clearly, we want to optimizecurrent business processes.
That's the obvious part, right?
We want to make what we dohappen faster, better, cheaper,
right?
Traditionally, you'd pick maybeone or two of those, but you
(40:57):
don't get all three.
Now you can get all three.
But the bigger question is whatshould you be doing?
As opposed to how do you dowhat you currently do better?
And the what should you bedoing is informed by constraints
, and the constraints are thecapabilities of these AIs, but
it's also the cost.
So knowing that the cost keepsrelentlessly going down should
open your mind to thepossibilities that six months,
(41:18):
12 months, 24 months from now,video generation will be close
to free.
For example, right now, if youwant to do a lot of hours of
video generation with HeyGen,you have to sign up for plans
that will cost you in the fiveor even six figures.
That's a constraint.
We're, by the way, extensivelyusing HeyGen in a variety of our
content production, for our AIlearning hub and for other
(41:39):
things, and we're making theinvestment because we think it's
worth it, but we also know thatit's not really a long-term
recurring investment becausethere is competition in every
one of these dimensions.
So that's part.
One is that I think you canthink bigger and think a little
bit longer term, something thatyou know you wouldn't say oh
well, I was going to invest amillion dollars in a new AMS and
next year it's going to be$100,000 and the year after it's
(42:01):
$10,000.
That's not a thing, right?
That's not the way people thinkand that's not the way these
systems tend to work.
But with AI, that's literallywhat's happening with the
fundamental models, because ofcompetition, because of the
amount of capital being thrownat it.
So that's true for what I'dcall kind of the raw materials
of AI.
It's the fundamental buildingblocks, really the models at
that layer.
(42:21):
It's also happening withinference competition, because
so many people are going afterbeing your cloud provider for AI
, whether it's the hyperscalermajor players like AWS and Azure
, or if it's people who have anew approach to hardware, like
our friends at the Grok which isGROQ, to always repeat that for
clarity or a number of otherfolks that are out there doing
(42:43):
cool things with inference.
That's a hyper competitivemarket.
You're going to keep seeingcosts come down there.
But to come back to the secondpart of your question about like
well, how come we aren't seeingcosts for the application layer
which consumes these?
Ais is for one, it's marketeconomics.
There's fewer competitors thanthere are at the model level.
(43:05):
There's also more complexity interms of integration with
systems.
There's kind of the nuance ofwhat a lot of times is referred
to as the last mile of thesolution actually solving the
problem.
So you can take the model thathas Gemini 2.5 Pro caliber
thinking, but then to wire it upinto your ecosystem and make it
work just right for you.
Still today there arespecialized pieces of software
(43:27):
that have to be integrated.
There's a lot of labor that'srequired.
There's domain knowledge interms of what the best processes
are.
So the actual end solutions, Ithink you are going to see them
come down in cost overall, butit's slower.
It's kind of like if we said,hey, let's just imagine that the
price of steel was close tozero all of a sudden, would that
(43:48):
mean that the price of carsimmediately goes close to zero
as well?
You know, when you're furtherdown in the supply chain, you
know that can happen.
Right.
Cost of materials can go up anddown.
Sometimes it has a rippleeffect and sometimes it takes a
lot longer to work its waythrough the system, but
ultimately this is a really goodthing.
The component costs of theultimate solution obviously are
(44:10):
a major factor in that solution,but more than anything, it's
this abundance mindset that I'mtrying to really hammer into
people's minds that youshouldn't think of AI video
generation or AI whatever as ascarce resource, that you can
only use it in a tiny number ofareas.
You should be thinking that bythe end of this year you'll
probably be able to do it acrosseverything.
(44:30):
So every single blog post thatyou have, let's have a great
video on that post, every singletime, completely AI generated.
Today that might cost you tensof thousands of dollars.
By the end of the year it'llprobably be considerably less.
Speaker 2 (44:45):
I love that.
I call it my version of that.
So you said dream bigger.
I call it often breaking yourbrain, because it's constantly,
every day, all day, having tosay, like what else could I do
with this model, breaking thesebarriers of what you thought was
possible, especially when wetalked about member services.
Right, having a member serviceagent that doesn't just answer
basic questions, that actuallyhas access to your whole
(45:06):
knowledge repository and cananswer domain specific questions
, that's something that you haveto really open your mind to.
So I agree, I think that'sessential and actually, now that
I'm thinking about it, that itcould be an interesting take,
maybe for for Digital Now 2025.
I don't know if we have a themehammered out, but kind of the
idea of dream bigger, breakingyour brain.
Speaker 1 (45:27):
I like that.
I like breaking your brain.
I like that.
I like breaking your brain.
That sounds fun.
Speaker 2 (45:30):
Uh-huh, I don't know.
So maybe you all are the firstto hear our theme for this year.
But with that, everybody, thankyou for tuning in to today's
episode.
I know Amit mentioned Grok witha Q.
We do have a special interviewcoming up.
Not sure when we're going topost it, but in the next few
weeks or so.
So be on the lookout for thatand we will see you all next
week.
Speaker 1 (45:52):
Thanks for tuning into Sidecar Sync this week.
Looking to dive deeper?
Download your free copy of ournew book Ascend Unlocking the
Power of AI for Associations atascendbookorg.
It's packed with insights topower your association's journey
with AI.
And remember, sidecar is herewith more resources, from
webinars to boot camps, to helpyou stay ahead in the
(46:14):
association world.
We'll catch you in the nextepisode.
Until then, keep learning, keepgrowing and keep disrupting.
Popular Podcasts
Are You A Charlotte?
In 1997, actress Kristin Davis’ life was forever changed when she took on the role of Charlotte York in Sex and the City. As we watched Carrie, Samantha, Miranda and Charlotte navigate relationships in NYC, the show helped push once unacceptable conversation topics out of the shadows and altered the narrative around women and sex. We all saw ourselves in them as they searched for fulfillment in life, sex and friendships. Now, Kristin Davis wants to connect with you, the fans, and share untold stories and all the behind the scenes. Together, with Kristin and special guests, what will begin with Sex and the City will evolve into talks about themes that are still so relevant today. "Are you a Charlotte?" is much more than just rewatching this beloved show, it brings the past and the present together as we talk with heart, humor and of course some optimism.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
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