July 7, 2025 • 116 mins
Our 215th episode with a summary and discussion of last week's big AI news! Recorded on 07/04/2025
Hosted by Andrey Kurenkov and Jeremie Harris. Feel free to email us your questions and feedback at contact@lastweekinai.com and/or hello@gladstone.ai
Read out our text newsletter and comment on the podcast at https://lastweekin.ai/.
In this episode:
- Cloudflare's new AI data scraper blocking feature, its potential implications, and technical challenges
- Meta's aggressive recruitment for its Super Intelligence Labs division is covered, highlighting key hires from OpenAI and other leaders in the field
- Anthropic loses significant talent to Cursor, with details on their new economic futures program focusing on AI's impact on the labor market
- Notable open-source AI model releases from Baidu and Tencent are also discussed, including their performance metrics and potential applications.
Timestamps + Links:
- (00:00:11) Intro / Banter
- (00:01:43) News Preview
Tools & Apps
- (00:02:55) Cloudflare Introduces Default Blocking of A.I. Data Scrapers
- (00:05:44) Runway is going to let people generate video games with AI
- (00:11:24) Google embraces AI in the classroom with new Gemini tools for educators, chatbots for students, and more
- (00:16:23) No one likes meetings. They’re sending their AI note takers instead.
- (00:18:08) Google launches Doppl, a new app that lets you visualize how an outfit might look on you
- (00:19:14) Google's Imagen 4 text-to-image model promises 'significantly improved' boring images
Applications & Business
- (00:22:18) Mark Zuckerberg announces his AI ‘superintelligence’ super-group
- (00:29:35) Anthropic Revenue Hits $4 Billion Annual Pace as Competition With Cursor Intensifies
- (00:35:10) As job losses loom, Anthropic launches program to track AI’s economic fallout
- (00:38:04) OpenAI says it has no plan to use Google's in-house chip
- (00:41:08) Nvidia stakes new startup that flips script on data center power
- (00:44:11) TSMC Arizona Chips Are Reportedly Being Flown Back to Taiwan For Packaging; U.S. Semiconductor Supply Chain Still Remains Dependent on Taiwan
Projects & Open Source
- (00:46:57) Baidu releases open source model family ERNIE 4.5
- (00:51:55) Tencent Open Sources Hunyuan-A13B: A 13B Active Parameter MoE Model with Dual-Mode Reasoning and 256K Context
- (00:57:09)
As usual, in this episode, we willsummarize and discuss some of last
week's most interesting AI news.
You can check out the episodedescription for the timestamps
and links to all the stories.
I am one of your hosts, Andre Ov.
(00:31):
I studied AI in grad school and Inow work at Regenerative AI startup.
And what's up everybody?
My name's Jeremy.
The other co-host of the podcast.
Yeah, co-founder of Gladstone,ai, ai, national security stuff,
and all that jazz and more.
There's other jazz too.
What a week we, I, I gotta sayI did, so I did my prep one day
earlier than usual this week.
(00:52):
Mm-hmm.
And there's some weeks where that'sfine and nothing happens on the, you
know, Thursday before the show, andthen there's other weeks where it's
just like a giant double middle finger.
Fuck you.
All the news happens.
And, and maybe you missed acouple big stories earlier.
So this is what happened.
I had to frantically catch up this week onthe Baidu releases, the Tencent releases,
(01:14):
all the, like a bunch of big stuff.
And so, yeah, it just seems likea, it was gonna be not that big
of a week until suddenly it was.
And, and here we are soexcited for this one for sure.
Yeah, this is, uh, we have quite afew stories on the block and it's a
week where there was no gigantic news,nothing that was like the talk of a town.
but there was a lot of stuff thathappened that is worth talking about.
(01:38):
So this will be a prettydecently sized episode.
I expect,
to give people a preview ofwhat we'll be talking about.
Tools and apps.
There was actually no releaseunlike, I dunno, the last few months.
You know, we had Gemini, CLI,variations of cloud code, et cetera.
Nothing really big like that this week.
But some interesting kind of smallertools, applications in business.
(02:01):
We'll chat a bit more about Meta's,super intelligence push, which
was one of the big kind of fun,slightly drama things of a week.
and some, news about philanthropicand, uh, as usual hardware stuff.
Project open source, as youmentioned, some pretty big stories
there with, uh, Ernie 4.5 and IonA 13 B, and quite a few others.
(02:26):
Like we have five stories inthere, research and investments,
some more scaling research andsome more research on sort of the
place where we are at with LLMs.
Finally, policy andsafety, quite a packed.
Section there as well with somediscussions of Safety Institutes
of China, security risks,the state of AI for security.
(02:51):
Lots of various stuff.
So yeah, it'll be apretty, pretty good week.
And let's go ahead and get into it.
So in tools and apps, first storyis actually not a tool most people
use, but if you're building atool, you might be using it.
So I thought we should cover it.
CloudFlare has introduced defaultblocking of AI data scrapers.
(03:12):
So this is a setting that allowswebsites to automatically block AI
companies from scraping their data.
And that would require websiteowners to explicitly grant access
to bots for data collection.
This is, uh, kind of a big deal.
CloudFlare is a very big company.
(03:32):
Lots of websites go through.
And so the defacto, I guess standard forwebsites up to now has still been to a
large extent that, you know, unless youbuilt it in yourself, AI companies will be
able to eat up your data and they probablyhave, I guess that is about to change.
(03:54):
Yeah, I think the, the ultimatequestion is always gonna be to what
extent can you meaningfully determinewhat is AI traffic versus what is not.
That's already challenging.
And then it's just gonna get harderas, as we get into, computer use.
Right?
You can really have, ultimately theinputs come from the very same channels as
human computer usage in the limit, right?
(04:16):
And, and even mimic you know,human delays on clicking on
things, human-like movement.
This is sort of all part of the,uh, the end of the capcha era,
if you will, in every, everypossible sense of the term, right?
So I think there's a, a question asto how long this will meaningfully
be a constraint, especially given themassive economic incentives to scrape.
(04:37):
But still, I like it'sa, this is is a big move.
It's precedent setting, and it'salso, as you say, cloud CloudFlare
is already like a massive player.
So it is the, the default nowfor the internet in some sense.
Yeah, so phase transition a little bithere, but I, I think it will be temporary
again, don't, I don't expect in the longrun that we won't be able to have bots
and in fact, you know, open AI or whoeverelse, if they're dedicated enough, we'll
(04:59):
be able to get by this pretty easily.
Right.
And I think at this point, they'vealready scraped most of the internet,
so we can, I think previously theywould've had to have pretty cheap
bots or, or simple bots, right?
To just go through everything fast.
Now they can probably get aroundit, but, uh, if you're being
sort of a. Good, and lawful.
(05:21):
Then you do have to kind ofmake some, uh, I think specific
request types of your bot.
So technically speaking, if youare going by the rules, this
should be able to block you.
and yeah, as you said,cloud player, really big.
Apparently Cloudflare's Networkhandles approximately 20%
of global internet traffic.
(05:42):
So yeah, substantial
And moving on Next, we haveactually a fun tool that, uh, isn't
out yet, but is coming runway.
The company that's focused onmaking it easier for people to edit
videos of AI and generate videoslately is going to get into gaming.
(06:02):
So they have announced the idea of lettingpeople generate video games with ai.
Apparently this is just a, aplan to release a new interactive
gaming experience next week.
So we'll see when it comes out.
seems a lot like, uh, if you rememberAI Dungeon basically sort of interactive
(06:22):
story that generate text and imagesand you go on a little kind of d
and d esque adventure where you.
Try to go through some scenario.
It's like video gamein a very loose sense.
Uh, and so the preview of what we'veseen will be, screenshots, uh, indicators
(06:42):
like that, but more polished, whichpersonally I think is, is pretty cool.
That's, uh, a fun use of AI that hasmaybe been, not entirely explored
to the extent that it can be.
Yeah.
the context here too is so they,they've been known for moving into
Hollywood and helping big studiosput together movies cheaper.
(07:03):
As, you know, as they say, like,like 40% faster, something like that.
They're contrasting though the speedat which gaming companies are moving to
adopt this with the traditional kind ofHollywood movie industry as being a lot
faster, which does make sense, right?
I mean, you think about, the baggagethat's inherited by the Hollywood
kind of movie complex, right?
There's a lot of stuff.
Even just looking at the laborunion side, you know, screen
(07:25):
actors, guild, all that stuff justsort of slows your implementation
ability in a lot of, a lot of ways.
Whereas in the gaming industry sort,that's less of an issue, right?
You have a lot of indie gamers,for example, are gonna be very,
very quick to pick this stuff up,and even big kind of AAA studios.
So really interesting fromjust a, a pacing standpoint.
just another note here, they have beenin talks, uh, runway has apparently with
(07:48):
meta about a possible acquisition there.
It seems like, and this alluded to inthis article, kind of as a non-sequitur,
but it's still, worth flagging.
So they were talkingabout it and Valenzuela.
So one of the co-founders of Runwaysays, I think we have more interesting
intellectual challenges being independentand remaining independent for now.
So seems like the talk of the metaacquisition now sort of falling
(08:11):
through based on that whichitself is interesting, right?
Because for meta to start.
Playing with runway.
You could see why they wanna do that,obviously for content generation, uh,
especially given meta's recent challengesin AI and, and having, uh, actually sort
of high-end generative capabilities.
It could be an interesting acquisitiontarget, but especially on the gaming
side, starts to look a little bit in away like Netflix's move into, into gaming.
(08:34):
You know, that's kind of one angle youcan start to imagine meta playing with,
but that looks like it's not gonna happen.
So, uh, so interesting that fell through.
At the same time, meta obviouslydeveloping its internal capacity for
AI on the super intelligence side ofthings with all those acquisitions.
So, uh, kind an interesting storyabout many different things.
But runway definitely is aninteresting company to watch.
Yeah, I'm, I'm a fan.
(08:55):
They've been around for quite a whilein the landscape of AI startups, and
have focused very much on sort of aprofessional side of tooling rather
than the cool side of generative models.
for the most extent, they do havetheir own generative models, which
are not, frontier that aren't quiteas good, but they focus a lot on
kind of deep integration to makethem usable as part of a more sort
(09:18):
of editing tool suite kind of setup.
So, curious to see what this hassadly don't have access to this yet.
Uh, I'm seeing you don't haveaccess to game worlds yet.
want to try it out when I can.
Yeah.
And one, one last comment too on, um.
on the Frontier model side withcompanies like Runway, and you're
talking about the, the tooling as beingthe thing they focused on historically.
(09:40):
Less so.
The models, you know, there's this, uh,famous story of Microsoft in the early
days as they're deciding which way they'regonna go strategically as a company,
they end up going in the direction of.
Obviously, as we know, making softwarethat's really expensive in high
margin on the basis that there'sa bunch of companies that can make
(10:01):
laptops like, or, or computers, right?
So the, the famous phrase hereis commoditize your complement.
So software is complimentary to hardware,and if there's a bunch of different
companies making hardware, then theprice of that hardware gets driven down.
Not to zero, but it gets driven down.
The margins get driven down to zero.
So you've got a bunch of peoplecompeting really hard to make really
(10:21):
cheap hardware, but if you own theintegration point of that stack, which
is the software, the operating systemlayer, the application layer, Microsoft
does, your margins can crush it, right?
So this in a sense, is what alot of companies are trying to
figure out in the AI landscape.
You've got all this competition atthe model layer, a million companies
making AI models, especially computervision, you know, especially video
(10:43):
generation, that sort of thing.
But where is the value bottleneck, right?
Where is that aggregation point?
What is the compliment?
To the commoditized models, andmaybe it's the tool chains, maybe
it's some kind of home base for useraccounts or something like that.
It's really as yet kind of unclear innascent, obviously hardware is one,
one choke point in, in the value chain.
(11:04):
But, this is all sort of part of,you can think of runway strategy
there as being, okay, let's notcompete with everybody else on the
thing that's already commoditized.
Let's try to focus on an area that'sless touched and that may become
a choke hold in the value chain.
that's the only way you can reallycompete if you're not open ai, if you're
not anthropic, right, that you need thateither scale or favorable kind of value,
aggregation points in the, in the chain.
(11:24):
Moving on a couple quicker stories.
First up, we have Google embraces AIin a classroom with new Gemini tools
for educators, chat bots and students.
So they've introduced 30 newAI tools for educators, at,
apparently an tech conference.
and that includes a version of theGemini app tailored for education.
(11:48):
Gemini AI Suite is now available forfree for all Google workspace for
education accounts, and that has variousfeatures like lesson plan generation,
personalized content creation.
teachers can create custom AIexperts called gems, which you can
also do in in Gemini and so on.
yeah, it's, it's a pretty.
(12:08):
Sort of customized, wrapped,version of their tooling.
If you go and, and see, they haveactually a little kinda special UI for
classrooms in this suite of errors.
And, with Gemini there is a ton ofsort of prebuilt things like, uh,
generating a lesson plan, a quiz,brainstorming project ideas, lots of
(12:30):
sort of suggestions for use cases.
And this is following up, I believe,uh, Ivo Open ai Andro also has
a EDU version of their service.
Kind of notable in a sense because,that's one of the big use cases.
Students are using this stuff likecrazy teachers also using it quite
(12:51):
a bit, from what I know in termsof grading and and preparation and
so on, they certainly need the helpin terms of managing the workload.
So having more and more sort of, nativefor education versions of these things
is significant because I do think, youknow, the educational sensor, uh, sector
(13:12):
needs to figure out what to do with alums.
How do you change education now that.
You know, homework is, let's saymuch easier to do than it used to be.
Yes, that's right.
Yeah.
It's, it's the, the ageold question, right?
Are LLMs calculators or, or are they Yeah.
Something that actually atrophies,brain function and that sort of thing.
But yeah, one of the, I think one ofthe interesting questions is when do we
(13:35):
really move from, you know, the framehere is Google AI tools for educators.
When do we move from that frameto Google AI tools as educators?
Because we are on that continuum.
Uh, you know, speaking from painfulexperience, doing my, my physics degrees
like, at least in physics, and, andAndre, you can tell me how this isn't
Cs, but like, profs are, are liketerrible teachers on average, right?
(13:57):
You'll have like, you know, two orthree profs who knock your socks off and
then the rest of them are just garbage.
Like, you know, they wanna do research.
They're not there to, to teach.
Really, it's not their passion.
It's not what they're great at.
And so they're kind of competingagainst a suite of products that
has been increasingly optimizedto do a better job than they can.
And at a certain point, I wonderif, you know, teachers, professors,
(14:19):
all the stuff, there's a transientwhere they basically are just Sherpas
guiding you towards the best kindof generative AI tools that they're
aware of to get you started over time.
Even that function obviouslygets automated away.
but I would expect it'll come witha lot of resistance, especially
from, you know, on, on the teacherend because, well, I mean, you've
got unions, you've got entrenchedinterests, nobody wants to be replaced.
(14:39):
But that transition point is gonnabe really interesting because the
people who are by and large doing themeasuring of the effectiveness of these
tools, right, who are writing thoseeducational fury papers or whatever,
are people in the system who have anincentive to potentially, eventually
pretend that these tools can'tautomate quite as much as they can.
So I actually think this is gonna be aninteresting point of friction as we look
(15:00):
at how can regulation, how can entrenchedinterests, slow adoption where really
you would want more adoption faster.
Anyway.
My 2 cents, but there it is.
Yeah.
I, I wouldn't say most professors areterrible, but, uh, it's, I'll tell you,
in physics, in physics, it feels that way.
Yeah.
I think in N Cs, because their classesare, are super popular, typically, maybe
(15:22):
there's, there's more and there's, youknow, many TG students who are very
motivated to do a lot of work, right?
But professors are typically overloadedand have very limited bandwidth.
And that's, I think true for alot of teachers across Absolutely.
Education in general.
so these tools that arespecifically aimed ED educators
could hopefully help with that.
(15:43):
And one of the thing noted in thearticle is that this is coming alongside
updates to manage Chromebooks more onthe side of students, including, uh,
a new teaching mode where teachers cancorrectly, uh, direct with students.
And, I think that reminds me that,uh, Google had, has made some inroads
(16:06):
on the device side of Chromebooks.
They are at least, to some extent,I don't know the comparison of,
uh, Chromebook versus, tablet, butpretty significant inroads in terms
of what, devices students get to use.
So together this, thiscould be significant.
And next meeting.
Not a new tool, but I thinka, a fun discussion of where
(16:30):
certain tools are being used.
The title of the articleis No one likes meetings.
They're sending theirAI note takers instead.
And this, starts with a little, funanecdote of a person named Clifton
Sellers going to a Zoom meetingwhere there are more AI note taking
(16:51):
bots than human participants.
So this is, if you're like,at a Zoom meeting and a teams
meeting, we will meet, et cetera.
You can presumably often be seeing this.
Now in, in a lot of major companies,there's many providers of these things,
including Google Meet and Teams andZoom themselves, plus some others.
(17:14):
So yeah, I'm using to me as, as someonewho works at a tiny startup where
excessive meetings aren't really a thing.
Like if you go to a meeting, you'regonna be doing some talking and,
and some, uh, actual useful work.
But, uh, having been or having knownpeople who work at big companies and
(17:36):
the general reputation they have kindof not too surprising to see the trend.
I've always sort of struggled with,obviously, you know, meetings are, are,
uh, are the enemy, by default in startups.
Like people just naturallyhave an aversion to not having
meetings unless they're necessary.
So I, I definitely, this is like a moreforeign thing, but I, yeah, I mean, you
(17:56):
know, it seems like, something that, thatcould easily happen and I wonder what the
failure modes end up looking like wheneverybody, or most everybody in these
meetings is just an AI agent, but yeah.
Mm-hmm.
couple more stories from Google,just there wasn't any major stuff,
so it wound up being a little heavy.
Uh, on the Google side, we havea new app from 'EM called doppel,
(18:21):
which uses AI to visualize howdifferent outfits might look on you.
This is an app that is being releasedboth an iOS and Android in the US
and yeah, let's you try out outfits.
You upload a full body photo of yourselfand you can use images or screenshots of
outfits to see how they, uh, would look.
(18:43):
This is generating actually bothstatic images and AI generated
videos, which is, uh, kind of neat.
Yeah, something we've seenalready in the AI space.
I'm sure this has already been built bycompanies and so on for quite a while,
but this is, obviously better than whatyou might get with the use of VO and, and
(19:06):
tools like, uh, Google Gemini, editing.
Nowadays, you'll have a very,very good preview, with ai.
And the last story issomehow Google's image four.
So this kind of sneaked under the radar.
Google has introduced their latest textimage model, Ima Imagine four, which
(19:27):
is, you know, the follow up to theirmain line of text to image generators.
Uh, now they had Imaginethree for a while.
Now we have Imagine Four,also Imagine Four Ultra.
And they say that this is better at,handling, you know, very specific prompts.
Basically what we've seen with AI imagegeneration, now, the focus is on prompt
(19:51):
adherence and being very good at thingslike spatial layout, preserving text
like all of these, subtler details thanjust making realistic looking stuff.
And yeah, it's, it seems like it shouldbe a big deal, but nobody was really
excited about it from what I could see.
(20:11):
And it does sort of not introduceanything fundamentally new from Imagine
Free or anything else we've seen.
It's just the latest and.
Greatest.
Yeah.
Again, with these image model, and,and this is strictly a function of
the fact that, you know, I don'tspend my time focusing on image
generation in my work, right?
This is like something, I'm, I'mfollowing this more or less as a passive
(20:34):
observer, but incremental advantageof these models over each other is
something that feels pretty opaque to me.
Like, how is imagine three betterthan imagine or imagine four better
than imagine three on things like,like writing, you know, what kinds of
things could imagine three write thatImagine four can't is a bit fuzzy to me.
I, I, I, I get the impressionwe're saturating this stuff.
(20:55):
Surely we must, but who knows?
Maybe at a certain point you wannabe able to, you know, put paragraphs
and paragraphs on an image andhave it be faithfully represented.
I imagine that's gonna be the case.
You know, if you look at like, movie,content, like, you know, assets that
you wanna put in a movie or something.
don't know, uh, super high faithfulness,high resolution use cases, but
uh mm-hmm The images that theyshow are beautiful, no question.
(21:17):
They always are.
So I'm kind of like, looks good.
Yeah.
We do have some impressiveexamples, uh, in the release.
So with Imagine four Ultra, we havea prompt, like a free panel cosmic
epic comic panel, one tiny Stardustin Nebula radar shows anomaly text
(21:38):
anomaly detected hall, text Stardustpilot whispers panel two violinist
and Leviathan emerges console.
Red text warning panel free.
Leviathan Chase is shipped throughAsteroids console text shield, critical
screen text debate, pile of screams, SFX.
Anyway, and it's, you know, it doesall that very they faithfully, there's
(22:02):
some quirks in the rendering thatyou might say are not quite right,
but way, way, way beyond what you'vewould've been able to do previously
with, and, things where you involvecomposition and and placement and so on.
And moving on toapplications and business.
First story is about Meta and MarkZuckerberg's Drive to set up the Meta
(22:28):
Super Intelligence Labs division.
So we've known about this some time.
We've been covering some of the stories.
I believe we talked last week aboutsome of the hires that have been
announced in terms of people from OpenAI.
But this week, it kind ofgot formally kicked off.
Uh, Zuckerberg announced it internallyto Meta, and a bunch more people,
(22:51):
have been, announced to have joined.
And in fact, I believethis is a new update too.
In addition to Alexander Wang, former CEOof scale ai, Nat Friedman, former CEO of
GitHub will also be joining to lead thisdivision, which now has, 11 AI focused
(23:12):
employees from, philanthropic GoogleDeepMind, and in particular open ai.
We have like eight people from openai from across the company, people
who've worked on very significant,things like or three mini and, G
four oh and and so on and so on.
there's been some leaks as tothe pay packages, not quite.
(23:36):
As absurd as, uh, I believe whatSam Altman was saying in terms
of offering people a hundredmil upfront, but still higher.
I, I, I've, I've heard, Dylan Patel.
Mm-hmm.
Right.
Was on a podcast recently talking aboutwhat he had heard was a $1 billion
package that was pitched at somebody.
I don't think they took it.
But it was pitched at somebody.
(23:56):
And then I've seen other headlines about300 million, like, so it seems like
there's a range, and there are somepeople who have been offered more and
some a lot less, which, you know, it'swhat you'd expect, but it, it's, uh,
it seems very ambiguous at this point.
It's, it's significant to the extentthat, within OpenAI, Sam Ottman sent a
memo, to the staff on Saturday and, andbasically addressed this saying that
(24:20):
yeah, meta has been pretty aggressivelyrecruiting, senior researchers, and it
was cast as apparently a quote from hismemo is, someone has broken into our home.
And, uh, later, I think thatwas Mark Chen, if I recall.
Like Mark Chen said, it feels likesomeone's broken into our home.
(24:40):
Yeah, yeah, it is.
From messages on Slack.
So yeah, it's clearly a big, effortand a big investment on Meta Spartan.
The stock hit an all time high onthese announcements, so I guess it's.
Paying off from a stock perspective.
I mean, you know, I think I, I alludedto this last episode, but I mean, what a
(25:03):
repudiation of y Koon's philosophy, whichcould not have been less open ai, right?
And then all of a sudden,basically Zuck says, you know what?
We're doing super intelligence.
We're calling it that.
And also we're gonna do it with likeeight out of 11 or something of our
hires are gonna be open AI guys.
And then Alex Wang, like one of thefew non-op AI guys is like literally
(25:26):
from a company called Scale ai.
But who also has the, uh, inclinationtowards the super intelligence
perspective that that deviatesquite significantly from, Y Koon.
So it's quite an interesting situation.
I'm very curious where Y Koon ends up.
He's been interestingly silent this wholetime that's kind of noteworthy for a guy
who's usually bombastic on social media.
(25:47):
Yeah, like, you know, obviously my, mybiases are well known on this podcast,
but, it does feel like it's, it's metasort of coming around to that view
after spending so long in the woods.
But, but this is a reallyinteresting series of, of
acquisitions of personnel, right?
So, what you have to do ifyou're meta is you have to.
Shake up the, the game boardin a serious way, right?
People were just not interested injoining a fast follower open source
(26:12):
lab or even a slow follower opensource lab as it started to look like.
And so, you know, you gottahire the best, get people jazzed
about working at Meta again.
And that means you needto refound the company.
They were refound the AI partof the company and make it
clear that you have top cover.
The pitch now is gonna be verycompelling actually, right?
You look at the caliber of the peoplewho've joined the fact that Zuck
owns the majority voting shares,for meta so he can make unilateral
(26:36):
decisions that other people just can't.
That's a really interesting pitch, Andso, you know, you've got this massive
compute fleet, you've got a lot of data.
'cause you're meta, you've gotzucks backing and now this, this
really interesting team behind you.
The big open question remains, whatis Meta's position on alignment on
technical safety, on security, right?
I mean, again, Koons is well known, fairlydismissive of it with some asterisks here
(26:59):
and there and, you know, there's nuance.
But Alex Wang, you know, certainlysafe super intelligence, Daniel
Gross coming over from there.
These are, and you know, one anthropichire, I believe, uh, these are, are
places, including DeepMind that havehistorically oriented towards more
of an alignment friendly perspective.
I'm personally really curious whatwe'll hear in the coming weeks and
months about their position on this.
(27:20):
But they are a live player.
I would, you know, call 'em a tierthree player for now, along with.
You know, SSI and, to some degree XIwith, you know, maybe Anthropic, I
would put it number one right now.
And, and opening AI numbertwo, something like that.
You could debate those allday, but it seems like they've,
they've made themselves a, alive proposition with this play.
That's at least howit, how it reads to me.
Mm-hmm.
(27:40):
Yeah.
No and Lacoon is interesting.
Now, as far as I know, meta willstill have fair, their internal
AI research division thatpublishes quite a lot of research.
And, to be fair, they've published alot of very significant research on
all s They're not sort of refusinglms, but Yna Kuna has pretty famously
(28:03):
been arguing that LMS will not leadto, a GI or a SI and some other
techniques beyond that are needed.
So this will live alongside the existingresearch group and presumably be
less focused on publication and, andtraditional academia and more focused
on going head to head with open ai,philanthropic in terms of building
(28:25):
out the tech and probably not doing asmuch research on the alignment side as
philanthropic for instance has been doing.
and on the compensation side, yeah,reports are saying not so much upfront
offers, but you see offers upwards of ahundred million on the first year, 300
(28:45):
million over the span of four years.
Yeah.
And there's of course nuances thereof like stock versus cash and so on.
But, some crazy big numbers and Ithink numbers that they probably
need to make for people to leaveopen ai, philanthropic and deep mind.
Because at the end of the day, fromwhat I know of people, who go to
(29:06):
these social companies, typicallythey're a bit more startup oriented.
They're more interestedin smaller companies.
They're sometimes they have somebackground at Google, for instance.
But even DeepMind is a bitdifferent from Google itself, right?
As being more of a researchdivision up until recently.
So, in some sense, not too surprising,they had to go all out to convince
(29:30):
people to flip from deep mindphilanthropic and open AI to go to meta
And next up actually, let'stalk about philanthropic.
We've got some news that they've losta major, pool of talent or a couple
major leadership, positions for talent.
Uh, and those went toHemisphere, the makers of Cursor.
(29:52):
So Cursor has hired Boris Cherney, who leddevelop development of Claude Code, and
he'll be joining as Chief architect andhead of engineering starting this week.
And Kat w the product manager forCloud code philanthropic is joining
as the head of product at Sphe.
(30:14):
for reference, if you dunno, cursor isthe leading, or maybe not the leading.
I'm not sure how the market stacksagainst VS Code, but at least among
people who use AI heavily, cursor isseen as the tool for ai for coding.
the development environmentthat is leading the pack in
terms of quality and have grownexplosively by valuation is, crazy.
(30:38):
Cloud code did disruptthat to some extent.
I think certainly as someone who hasused Cursor, once I've adopted cloud
code, I find myself using it a lotless and even moving back to VS code.
So kind of makes a lot of sense for'em to make this aggressive move.
The article also notes that this ishappening as Philanthropics revenue is
(31:01):
hitting 4 billion, annual, which is.
Quite a lot.
Yes.
You heard, you heard it here firstguys last 4 billion is quite a lot.
Sometimes I gotta put in thatcommentary, you know, that's
really deep insightful stuff.
You know, I feel like we've gottenso desensitized to hearing like
(31:24):
numbers, like billions and gigawattsand all this stuff like it.
It's, uh, it happenedpretty fast, didn't it?
Over the last few years.
Yeah.
And that's, that's revenue too, right?
We're not talking about valuation.
Of course.
The, the big question is, hasbeen historically, can these
companies live up to the hype?
Can they actually generate revenue?
Andros done reallyimpressively on that front.
I think their last fundraise hadthem at a $60 billion valuation.
(31:46):
I believe they'recurrently fundraising now.
So we'll see what, what, uh, youknow, that valuation looks like.
But they're, they're also forcontext, I vaguely remember they're
like burning net, uh, somethinglike three to 5 billion or so.
So, you know, there's still a lot of,a lot of burn going on for all of CapEx
and, and, and all the talent as well.
But on the talent side, yeah, thisis interesting 'cause Anthropic has
(32:07):
been by far and away the big winner.
And we, we covered this, Ithink a few episodes back on the
recruitment or talent wars, right?
Like something like a, an eightto one ratio of people leaving
open AI to anthropic versus peopleleaving anthropic for open ai.
So far, far more people movingtowards anthropic than away from it.
Similar, in fact, even more extremenumbers, relative to DeepMind.
(32:28):
And so it's rare that you seethese big poachings from Anthropic.
And, and part of that, presumablytoo is that philanthropic does have
a kind of it's clear what they standfor in terms of the, the safety side,
the security side, and all that.
And so people are not necessarilyjust working there for money.
At least that's that's what I'veseen quite clearly talking to
folks at all these different labs.
But this is, this is interesting, right?
So you've got cursor movingin on the, on this talent.
(32:50):
Obviously all of Cursor's chips hereare on the code generation stuff,
so they can afford to kind of plowmore money into these poachings.
And when we, when we talk about, youknow, cursor hitting $500 million in a
RR compared to Philanthropics 4 billion,that may sound smaller, but you know,
it's over 10% of what Anthropics making.
So these are like, you know,they're, they're in the kind of,
(33:11):
maybe not quite the same orbit, butapproaching that, that same orbit.
So, in any case, we'll seewhat ends up happening here.
But, the cursor anthropic thing, youcan sense the nerves, the anxiety about,
you know, maybe upsetting anthropicthat's coming from any Sphere, which
is the, the company that owns Cursor.
So one of their co-founders, Sule, Asif,I'm probably butchering that, did refer
(33:32):
to Anthropic as quote one of our closestpartners, in the context of this article.
So you can kind of see it's like hownobody wants to piss off Jensen right now,
or Nvidia you know, sort of similar thing.
There's a lot of dependency here.
And these are, these are dicey moves.
Yeah, exactly.
if you use Cursor, the VIN tool, youknow, it allows you to use any of
LMS really, but from a Vinet, uh,you can choose which LM you want to
(33:55):
use Gemini or, or cloud and so on.
And, that I think is a big reason why youwouldn't wanna upset this relationship.
Uh, it is pretty important for.
Any sphere and cursor to remain friendlywith the providers of Frontier LMS at the
same time it does feel like cloud codeand, and generally the movement towards
(34:18):
agentic, tooling that even goes beyondthe traditional developing environment
with Cloud Code being kinda a standalonething outside of where you look and
write code, which is what Cursor is.
and has really kind of upped theamount to which, agents in agent
AI can do a lot of coding withoutyou manually supervising it.
(34:42):
So yeah, to me it's the surprisingbit is not just that they hired these
people, but they, gave them theseroles of chief architect and head
of engineering and head of product.
I mean, that is big, right?
Well, it again, you know, you gotpeople who work for stuff that's not
necessarily just money and if you'regonna convince them to move over Yeah.
(35:03):
They're, they're gonna need asense that they're able to shape
the direction of the company.
Mm-hmm.
So this is part of the, you know,part of the pay package in a sense.
And next story also about philanthropic.
Actually kind of not directly relatedto any business updates, but the
related to economy, I guess in general,philanthropic slash the economic features
(35:24):
program to research AI's impact onthe labor market and global economy.
They want to provide, yeah,basically evidence, based insights
into the economic effects of ai.
And, this is coming after in recentmonths, philanthropic, CEO, Dario
Ade has been talking about prettydire consequences of ai saying
(35:48):
that AI could eliminate half of allentry level white collar jobs in
the next, uh, one to five years.
And, and that wouldunemployment up to, 20%.
So not as surprising.
They are moving in thisdirection, I suppose.
And, you know, we just talked about cloudcode and, and cursor definitely gonna
(36:12):
be a hit on the job market for that.
I think it's already here.
In fact, if you're an entrylevel engineer, if you're in
the programming, uh, businessand more junior, it's probably
getting a lot harder to get a job.
I had a conversation with somebodypretty senior up at, at one of the
Frontier Labs who just straight upsaid, our expectation is that we will
(36:35):
never hire another developer with lessthan 10 years of experience again.
that's pretty amazing, right?
Like, and, and obviouslythat's the Frontier Lab.
They have access to the internal modelsthat they build all the best stuff.
But if that's not a canary in a coalmine, like, I don't know what is right?
Like the, even if you froze AIcapabilities today I think it's quite
credible that, yeah, you would seethat kind of effect at a minimum
(36:57):
propagate, uh, throughout the economy.
And that's, that's a big issue, right?
Those entry level roles are.
Or how people jump to obviouslymid-levels of, of seniority.
But, so this is a, a really a reallyinteresting play from Anthropic.
It's, it's a kind ofbig package of things.
So they're looking at, uh, puttingtogether this economic futures program
that includes economic futures researchawards, which are like five, uh, sorry,
(37:18):
$50,000 grants for empirical research oneconomic impacts from ai, which again,
I think is this really important space.
Like the empirical side.
There's a lot of theoryhappening right now.
You know, you got Epic AI coming outwith their plots, and then you've got,
uh, AI 2027 coming out with their plotsand people are comparing plots and, you
know, its semi, it's semi evidence-based.
(37:38):
But then also, you know, we need justmore empirical research to ground these
predictions and analysis in they'realso setting up symposia so they can
bring people together to talk aboutstuff and setting up, uh, strategic
partnerships with research institutions.
So, finding all these ways to kindof, drum up more interest and, uh,
resources pointed in this direction,which again, seems like, uh, something
that's probably helpful if you believeanything like the 2027 timelines that,
(38:01):
uh, a lot of people in the space do.
And moving to some hardware stuff.
First we have the story aboutOpenAI is saying it has no plan
to use Google's in-house chip.
So apparently it has been recent reportssaying that OpenAI had been considering
using tpu, had been testing tpu, analternative to GPUs with, uh, open
(38:27):
air, having invested heavily in NvidiaGPUs, as have most other companies.
open AI is also developing itsown ship, to compete with VTPU.
I guess this came out in the contextof Open, having signed up for Google
Cloud Services, in continued trendof splitting slightly from Microsoft.
(38:48):
We have been on Azure from earlyon as a primary source of compute.
So yeah, significant, perhaps primarilybecause of, what we know about them
trying to create an alternative to TPUs.
And it's also, it's alsointeresting from a, a almost a
historical standpoint, right?
(39:09):
You've got two trendsthat are now intersecting.
On one hand, we've got this trend of,OpenAI trying to shake itself loose
from Microsoft more and more, right?
We've had a lot of stories that havecovered this in different ways, but
fundamentally, you know, Microsoftwas supposed to be their cloud partner
of choice now getting a, a bit ofcold feet in terms of providing all
the infrastructure support that isneeded for things like Stargate.
(39:31):
And so they've allowed OpenAI towork with Oracle, and allowed is
the right word here, by the way.
Microsoft does have right of firstrefusal on these, these, uh, uh, big
kind of infrastructure buildup contracts.
And, and that's where, you know,Oracle and Cruso and all these guys
are coming together to do Stargateand OpenAI, and is then looking for
other partners outside Microsoft.
At the same time, Google is now juststarting to push out in the direction
(39:54):
of third party partnerships, trying tofind ways to make their TPUs available
to companies that are not Google.
That historically was not a thing, right?
Google was a lot more protectiveabout access to their TPUs.
And so, these two, two trends are kindof intersecting each other and, and
causing this, what otherwise would'vebeen an almost unthinkable partnership.
Because Google so famously is thehome of Google DeepMind, which for so
(40:17):
long was the one big rival to open ai.
And so I guess time heals allwounds, and at least they've
been testing out these TPUs.
It does seem that they're not gonna.
Actually go for them ultimately,which itself is interesting.
You know, this is something that, theinformation, I won't say got it wrong,
but certainly their initial headlineearlier this week made it seem like OpenAI
was actually gonna go forward with this.
(40:37):
And there was some corrections,uh, issued in that article.
So, yeah.
Last quick note, I guess isthat, you mentioned OpenAI
is developing their own chip.
They're ready to hit thattape out milestone this year.
So tape out is when you ship thefinalized design for manufacturing,
presumably in this case, the TSMC.
They are partnering with Broadcom onthis, but they basically have their chip
(40:58):
design that will be finished this year.
That's a big, big deal.
And then we'll have to see how long ittakes to ramp up production and all that.
But that's a, a big thing andsomething you can, you can believe
we'll be keeping an eye on too.
And next up, Nvidia is one of theinvestors in a new startup that
has just emerged out of stealth.
The startup is Emerald ai and theirfocus is on kinda a deeper connection
(41:25):
of data centers into the energy grid.
So they provide software that allowsyou to change, AI workloads at and
between facilities and basicallyconnected to the local power usage to.
Not put as much strain on the grid,which is increasingly the case at,
(41:48):
I suppose, us and, some specificregions where the hyperscaler
are building their data centers.
Yeah.
This is kind of aninteresting play, right?
When you have so much power thatis being soaked up by data centers
and that power is being used invery inconsistent ways, right?
Like during a training run.
I mean, locally you have these giant boombust cycles of power consumption with, you
(42:11):
know, back propagation and all this shit.
but over, over longer stretches oftime, you also just have like, you know,
sometimes a training run is, is goingsometimes, you know, there's fine tuning.
GPU utilization isn't alwaysas high as it could be.
And so you've got high amounts ofvariation and what you might think
of as a 50 megawatt or eventuallya one gigawatt data center is
(42:31):
not exactly always consuming 50megawatts or one gigawatt of energy.
And so the question then becomes,okay, well if data center number one
is sort of in a slower period anddata center number two is ramping
up, can we arbitrage over that?
Can we have some kind oforchestration function on the grid
that allows us to, to sort of loadbalance in a more dynamic way?
(42:54):
And that's really whatEmerald AI is doing.
They think their modeling suggeststhat they could unlock up to a hundred
gigawatts, of US data center energysupply, which would take loads off
the, the grid and allow you essentiallyto build, to free up more gigawatts
for fresh data center builds.
That's super interesting.
A structural challenge, vulnerabilitythey may have is if we just develop more
(43:18):
efficient, high utilization algorithmsand hardware stacks for these models
such that you're, there's just likeless value in arbitraging between them.
But this is certainly areally interesting play.
Like, haven't seen anything like this.
It almost, it almost turns loadbalancing into a fresh source of
significant power on the grid.
It's the equivalent of building likehuge amount of, of power, right?
(43:39):
A hundred gigawatts.
I, if this works as, as promised.
Interestingly, the founder and CEOis Varun Ra, who has a background in
physics and has worked as a senioraide to Excl Envoy, John Carey.
So yeah, some notable names amongst thebackers of the startup, beyond Nvidia.
(44:04):
They include some individuals likeGoogle Chief Scientists, Jeff Dean,
Fefe Ali, and, and some other ones.
And one last story alsodealing with hardware.
The story is that TSMC Arizonaships are reportedly being flown
back to Taiwan for packaging.
So packaging, as we've covered in our longago hardware episode, which we refer to
(44:30):
like every other, or every single episode,uh, packaging is an important step,
in the production pipeline for chips.
And the fact that the SMC is flying chipsmade in Arizona back to the one basically
means that, Taiwan is still, let's sayyou could say a bottleneck where you
(44:50):
could say, you know, is crucial to the,chip, supply chain for AI in particular.
Yeah, and this is actually something thatwe called out in our latest investigation,
you know, the America's Super Intelligenceproject thing, where we looked into
supply chain risks among other things.
You know, thinking about how,how China could undermine
(45:11):
American Super Intelligenceresearch, uh, or add risk to it.
You know, we pointed out like.
Hey, everybody's talking aboutpackaging as if it's the solved problem.
But if you actually look under the hood,there are structural reasons why you
might think it could take a little bitlonger than expected, to have that online
and that creates issues just like this.
So essentially go back to our hardwareepisode to look at the details,
but you've got two fundamental corekinds of chips on a Nvidia GPU.
(45:37):
You've got the memory, the HBM, andthen you've got the logic dye that
actually does the computing, right?
So the memory stores the data, thelogic die does the computing, and
there's this dance between them.
They've gotta talk toeach other crazy fast.
'cause the logic die has tofetch data from memory to do
math on it and then send it back.
So to get them to talk to eachother really fast, you have to
set them both on substrate aspart of this packaging process.
(46:01):
The packaging process requirestoday this process called COOs.
And, and you know, COOs L is kind ofthe latest, well, I guess COOs R is.
But anyway, there's a bunch ofCOOs processes that can only be
done really in Taiwan at scale.
And so we've onshore a lot ofthe fab for the logic dies.
What we haven't done though,is onshore the packaging.
And so we still have toship them back to Taiwan.
(46:23):
By the way, big winnerfrom all this is Taiwan's.
Ava Air, which is the demand forair logistic services, has massively
risen recently, and this is why.
So anyway, kind of, kind of interesting,there are, by the way, plans to
increase and ramp up packaging quickly.
including by the way, a $165 billioninvestment that TSMC announced for the us.
(46:43):
there were plans to create advancedpackaging fabs that could do COOs here,
but there's kind of been no progress yet.
So until that gets solved, youknow, it's all well and good to have
great logic dyes being FD onshore.
But you can't package 'em,you're not making chips.
And moving on to projects and opensource, as we mentioned at the beginning.
First up, we have the, uh, announcementof Ernie 4.5, the release from Baidu
(47:10):
of a whole bunch of models underan Apache 2.0 license, Apache 2.0,
meaning you can use it for whatever,including commercial, uh, directions.
And, uh, the largest of these, there 10models here in total, including just text
to text, VMs, image plus text to text,MOE models, dense models, it's, and so on.
(47:34):
But the big one is.
A model with a total of 424billion total parameters with up
to 47 billion active parameters.
And they release a bit of evaluationcomparing primarily to other open
source, LLMs without a reasoning focus.
(47:54):
So they compare to deepsseek V three que free.
and for those, models in the particular,they have better performance on
most of the typical benchmarks.
And, yeah, this is somewhat significantbecause, you know, we have llama, we
have deep seek V three, on the bigmodel side of, of like actually quite
(48:19):
big models with 400 plus parameters.
Now there's another option for peoplewho want to develop seemingly a
quite good option to build on top of.
Yeah, and this is actually, uh, kind ofan interesting structural choice, right?
So they have the full version.
This is like Ernie 4.5, 300billion parameter version has
(48:40):
47 billion active parameters.
And so it's half the size asdeep seek V three, which is the
base model for R one, of course.
But uh, it's got more activeparameters and so it's maybe
got three times on a, you know.
Total parameter basis, the number ofactive parameters as deeps seek V three.
And in a smaller scale.
(49:00):
So sort of interesting trade off therein terms of, you know, like, let's
say, uh, flops versus versus memory.
Anyway, they, they do make a pointof saying it outperforms deep
seek V three, the full versionon the vast majority benchmarks.
22 outta 28.
I will say this is notsuper shocking, right?
Deep.
I mean, V three cameout in late, late 2024.
(49:22):
Yeah, that's right.
So that what, six months ago?
More than six months ago.
And so when you have a, a modelthat comes out six months later,
it's like, oh, okay, can BB three?
Sure.
I mean sure, fine, but, uh, it'snot exactly a ringing endorsement.
It doesn't mean that much at this point.
And this article also comments,I don't know, this is a
terribly meaningful comment.
There are no published comparisons withindustry leaders like Deeps seeks R
(49:44):
one open Eyes oh three or cloud four.
Well, yeah, because these areage agentic models, right?
So that's not apples to apples.
It is fair to compare the basemodel earning 4.5 directly
to deeps, seek V three.
What's gonna be interesting is whenthey come out with X one, so this is
going to be the age agentic model thatwill be directly comparable to R one.
It's not part of the current release yet.
But we'll see, you know, we'll seewhat the performance is like there.
(50:06):
Yeah, ultimately, think we gotta waitand see what the actual usage is.
Obviously usage of, of deep seek R one isdown significant, I think like something
30% or, or something, quarter overquarter, in the last couple of months.
So.
You know, there is a bit of an effecthere of the export controls, the,
the lack of an R two release that'scompetitive with some of the, the other
(50:27):
open source models that have come out.
So we'll see, this may be part ofthe export control story as well.
Right.
And, it's also a bit of a shift for Baidu.
to my knowledge, we haven'topen sourced Ernie, for context.
Baidu of course, is sort ofcomparable to Google within China.
Gigantic super, majorcompany in the cloud space.
Uh, and they do search and they'vebeen pretty early to the LLM
(50:51):
game, like this is Ernie 4.5.
You know, they released Ernie quite awhile ago, not too much after OpenAI
Hatch, LGBT, this is their first at leastmajor entry in the open source space.
In addition to the models, theyreleased a technical report that was
like 40 free pages, not counting theappendix, going into a lot of detail
(51:16):
on the training, similar to what yousaw of deep seek that went into all the
nitty gritty of the infra and so on.
and they also release tooling, so theyrelease, uh, the training framework.
they have earning kit, theyhave some other stuff as well.
Fast deploy deployment toolkit.
So a bit of a shift interms of the dynamics and.
(51:38):
Uh, yeah.
Over this year, I guess since Vthree and R one, we've seen more
and more open source releases ofbig models coming out of China.
They're kind of taking the lead onwhere you can squeeze out performance,
as opposed to, let's say, laa.
And speaking of that nextstory is about another Chinese
(51:59):
giant open sourcing something.
This time it's Tencent.
another, you know, one of other giganticcompanies that is really something that
a lot of, maybe a majority of people use.
And they have introduced IONA 13 B. this is NMOE model.
(52:19):
So this has 80 billion total parameters.
Only 13 billion are active at a time.
apparently it's been trained quite a lot.
20, 20 trillion token pre-training phase.
And it, uh, notably has a lotof training for, uh, thinking.
So they did post-training withreinforcement learning for tasks,
(52:44):
specific rewards, and they supportfast and slow thinking out of a box.
They say that they have a state of artagentic performance on benchmarks like
math, cmaf and GPQA suppressing evenbigger models, which as you said, I think.
Would be very, believable where, you know,given that it's been quite a while since
(53:08):
R one, there've been a lot of insightsas to post training for reasoning.
no reason really why we shouldn'tbe able to squeeze out comparable
performance with smaller models.
And that's seems to be the case here.
Also released with a permissive license.
Yeah, it's, it's an interesting release.
I mean, the, the benchmarks,as you say, do look good.
(53:29):
All the standard caveats that youmentioned, the most recent model that
they compare it to here is Quinn three.
The, the full kind of, uh, 22 billionactive parameter version and there
compares favorably but not, it's notlike it blows it outta the water.
It's better in someareas, worse than others.
It's more sort of the incremental movementthat you would expect from the space.
I don't think that there'sa, a huge, huge story here.
(53:50):
Some of the standout features, well,the, one of the standout features is
that the thing that's not a standoutfeature, which is that fundamentally
this reflects yet another model build inthe Chinese ecosystem that effectively
mirrors the deep seek, curriculum or,you know, training approach, right?
So you got like, yes,a pre-training phase.
You have long context adaptation andthen, like fine tuning and then there's
(54:14):
gonna be an RL stage that is as yetunreleased with fast and kneeling as well.
So, fast and kneeling is thisidea where you rapidly, towards
the end of the training process,you decrease the learning rate,
so the model gets updated less.
Let's quickly, let's say,as it gets trained on tokens
towards the end of the, the run.
And, and the idea here is just like, asthe model gains, more experience learns
(54:35):
more and more, like over time you shouldexpect it to be making smaller and
smaller adjustments, sort of honing inon, on what it needs to look like rather
than continuing to kind of flop around.
So they, they do that.
the second standout feature is thefact that they use this thing called
dual mode, chain of thought, right?
So you got two different modes.
One, and it's the same model, right?
So it's not like we're shipping.
(54:56):
A query to one model to do fast thinkingand another model to do slow thinking.
it's like one model that can route toeither sort of sub circuits within itself.
So one is this like low latency, fastthinking mode for more routine queries.
And then you've got this a slowthinking mode for multi-step reasoning.
So you can control this with a tag system.
(55:16):
There's like this no think tagfor, you know, faster inference,
and then for more reflectivereasoning, you pass it a think tag.
So it's pretty straightforward to use.
They also use reinforced learning,uh, reinforcement learning for task
specific reward models, or sorry,with task specific reward models.
So basically you have like reward modelsthat are being used for the RL loop.
(55:37):
And those are designed kind ofbess in a bespoke way for a bunch
of different tasks, uh, thatyou're training the model towards.
So nothing too shocking, eitherfrom the standpoint of the, the
architecture or the performance orthe optimization routine or the data.
But overall it's anotherimpressive player.
You know, Tencent showing in, theyactually can contribute, they can
(55:59):
play in the, at least the open sourcegame, uh, with other big companies.
Yeah, they did release a paperalongside this, that has at decent
detail, particularly on the trainingwith regards to this, Specialty, uh,
I forget the term, but with regardsto the training in particular,
they, uh, seem to generalize a bit.
So they focus on a whole range oftasks, not just math and coding, but
(56:23):
also creative writing, knowledge based,qa, multi-term dialogues, and so on.
And, and for each of those,they do supervised training
and reinforcement learning.
Uh, so they might be trying togeneralize to domains outside
the typical reasoning, things.
And the focus is, uh, moreon the, efficiency side.
(56:44):
So the abstract is kickingoff at the very beginning.
they presented as a model that optimizesthe trade-off between computational
efficiency and model performance.
So the gist here is, four 13 billionactivated parameters during inference,
a huge amount of parameters.
(57:05):
This seems to give youquite good outcomes.
Next we've got another,open source release.
This time it's more of an RLtrained coding agents coming from
together, ai, they call it deepSWE, deep Software, engineer.
And this is based on, on top ofa Quinn 3 32 B language model.
(57:31):
So the focus is on particularlythe training of a model two b,
a software engineering agent.
And, uh, among, uh, open weightmodels, this is a leading one on
benchmarks like SB bench verified.
I will say, I think we don't yet havebenchmarks that really capture the types
(57:54):
of things you can see cloud code doingin terms of very impressive, uh, tool
calling, uh, code-based exploration.
A lot of these benchmarksfocus more on solving GitHub
issues and things like that.
Not quite the same, but uh, regardless.
Yeah, clearly, making a push ina direction of having a cloud
code, capable, uh, open model.
(58:18):
Yeah, and this is, uh, noteworthythat it comes from together.
Ai, we've covered a lot oftheir stuff in the past, right?
They have this, philosophy of tryingto do like, kind of fully distributed
open source AI training, aggregating,compute from around the world.
There are a couple of different, differentcompanies that are pushing in this
direction as well, uh, philosophically.
But pretty impressive results.
And trained on 64 h 100 GPUs over 60 days.
(58:39):
So, you know, not a huge,huge workload here, uh, but
certainly non-trivial either.
It's sort of like meso.
anyway, also a partnership betweentogether a AI and agenta, which
is this other company that'sfocuses more on frameworks for
post-training, language model agents.
They have this framework called RLLM,uh, that, uh, that was apparently used
(59:02):
here, presumably somewhat experimentally.
So, yeah.
Pretty cool.
continue to see more ofthese open source RL agents.
N nothing like, nothing jaw droppingin terms of performance here.
It, it looks like it's solid.
It, it's again, a solidincremental, boost to performance.
And especially on suite bench, right?
So this is, a benchmark that doesmatter a lot because it is, or I
(59:25):
should say suite bench verified, right?
This OpenAI kind of.
Scrubbed version of theoriginal SW bench benchmark.
it mirrors pretty closely some standardsoftware engineering practices.
So to the extent that you're seeingthis model score close to 60% now,
uh, as an open source model that,you know, that's a pretty interesting
and pretty, pretty big deal.
So, uh, yeah, I mean a pretty solid moveby together ai they continue to impress.
(59:47):
Yeah, and they also released apretty nice report on the model,
with a bunch of offers, and notquite as deep as these previous ones
in terms of the details disclosed.
But, some pretty nice details inparticular about the training recipe
and the empirical results of training.
So pretty useful for insightsand, and research in the area.
(01:00:12):
Next we've got, some papers talkinginto some slightly different directions.
First, we have GLM 4.1 V thinkingtowards versatile multimodal
reasoning with scalable rl.
So this is getting into thevision language space we've seen.
(01:00:33):
More research, uh, moving beyondjust reasoning and text and, uh,
towards reasoning with images.
And that is, the frontier of research.
and so this is, uh, they have a nineB thinking model that, apparently is
outperforming, uh, a bunch of alphamodels, even surpassing really big
(01:00:55):
ones like Quin 2.5, VL 72 B. So, prettykind of, uh, meaningful results in the
reasoning with visual, uh, input space.
And they open source both forreasoning model and the base model.
Yeah, there's, I think a lot that'sinteresting about this, this paper and
(01:01:18):
this sort of philosophy, this approach.
So, it is still modular, sothey have a, a vision encoder.
The goal here is to combine orhave a model that can look at
images and texts at the same timeand, and do reasoning on them.
So yet they have a visionencoder that basically looks at
it could take videos actually.
So if you're familiar with convolutions,convolutional neural networks, typically
(01:01:40):
those are two dimensional convolutions.
So you kind of look at a, a patchof an image and you sort of,
anyway, you do some math on it.
You apply like a filter to that patchand it gives you a smaller patch, and
then you can apply another filter tothat smaller patch and on and on and on.
And that's kind of howconvolutional nuts are made.
Well, this is that, but instead of atwo dimensional patch of image, it's
(01:02:00):
a three dimensional patch of video.
So in the same way you're gonna take wherethe third dimension is time, basically.
you can do s strided, convolutions if.
You know, strided Convolutionsdoesn't make sense to you.
It's fine, don't worry about, butfundamentally, you're looking at three
dimensional patch across time now of, uh,of your image or therefore of your video.
And from there they're able to encode youknow, the image of the video in the model.
(01:02:25):
And then they take that encoding andthen they pass it through basically
just a simple feedforward neuralnetwork, like a, an MLP to map it in,
to get the dimensionality, to match thedimensionality of their language model.
So basically they're just liketaking now the, the vision encoded
yeah, latent representation.
And they're mapping it to something thatmatches the dimensionality of the latent
(01:02:47):
representation of the language model.
This, uh, GLM based language modelthat they use, which is from, JPU ai.
And then they're just gonnaconcatenate those two together.
So you now have a unified imageand language representation that
then you can basically just do,do LLM work on in the usual way.
The language model itselfthough is really interesting.
(01:03:08):
They use bi-directional attention,which, because we're doing,
a speed run on this story.
It's in, it's in the, uh, lightning round.
I'm not gonna get into why that is.
If you know bidirectional attention, then,Anyway, it, it just makes sense to do
this when you're doing reasoning on imagesmore than just going full auto aggressive.
And, and just like a decoderonly model, Anyway, the, the, the
(01:03:28):
reasoning behind that is, is sort ofintuitive, but, I guess only if you
have the intuition, it's intuitive.
If, if it's intuitive, you know?
Yeah.
Like, if this is a main storywe, we we're going Yeah.
Yeah.
Yeah.
Lots, lots going on at opensource this week, clearly.
Uh, and as you said, this is from gpu, AIand, Chico University, they release Yeah.
(01:03:50):
Uh, a pretty detailedreport on this one as well.
Uh, 18 pages and they show how itcan be used for things like long
document understanding we agents video,understanding coding, uh, stuff like that.
so seems to be a prettystrong entry in that space.
And last up, just one more new model.
(01:04:15):
This one coming out of the US for once,uh, actually from Apple in collaboration
with the University of Hong Kong.
Not really the US even, you know, mostly,mostly the paper title is Diffuse Coder
Understanding and Approving MaskedDiffusion Models for Code Generation.
(01:04:37):
So they trained a 7 billionparameter diffusion, LLM.
quick recap.
Most l LMS do other aggressive, inference.
Basically, you train the model topredict, one token at a time or maybe a
few tokens at a time, given your input.
Diffusion is totally different.
You're sort of like predictingeverything all at once.
(01:04:58):
Typically what you do for images,but very rarely what you do for text.
But, you can do it for text and there'sincreasingly research on that direction.
This is, some of that research.
They show that, you know, someinsights into training effectively
and, get pretty good resultseven compared to non diffusion
(01:05:18):
models like Quinn 2.5, uh, coder.
Yeah.
And we've, we've skipped over, Ithink, so this is the third paper this
week that has its own take on GRPO.
So this like group relative policyoptimization, the reinforcement learning
optimization routine that deep seekfamously used and kind of made popular.
(01:05:43):
Anyway, this is its own variant.
I feel like that's maybe nextepisode we should carve out just
like five minutes because the, theGRPO stuff I is really important.
It's, it's something that you'll, youknow, you'll wanna understand to, to
get what the differences between thesedifferent papers and their approaches.
But yeah, yeah, one these days, uh, somany topics we could do deep dive on.
(01:06:07):
You know, we should do areasoning, Episode clearly.
Probably.
Yeah.
What should be in that?
Yeah.
Yeah.
Might we'll try.
You know, we are always wantto do more, but life is busy.
Yeah.
And moving on, research and advancementsand speaking of doing more so we've
got this paper wider or deeper scaling.
LLM inference, time compute withadaptive branching tree search.
(01:06:32):
Okay, so Monte Carlo tree search is thiswell trodden path in reasoning where you
have, you know, imagine that you startoff with like, some attempted solution.
you now have the choice as to whether youwant to like make, let's say you make a
modification of the solution, you try torefine it, and now you have the choices
(01:06:52):
to whether you continue to try to modifythe thing you just modified or whether
you go back to the original branch andtry to spawn another variant of it, right?
So do you explore more from thatstarting from that original node?
Or do you go deeper and exploitmore, go deeper down the path, right?
That you had sort of started down.
(01:07:13):
And there are two extremes, right?
Like there's some models thatjust continuously refine the
same pro or not prompt the sameoutput more and more and more.
And in that sense, they riskgetting stuck in a rut, right?
They don't do much exploration.
another extreme variant to thisis to say, okay, let's just like.
Sample a crap ton, like generatea bunch of different possible, uh,
solutions starting from the sameprompt, but never go deeper in,
(01:07:36):
you know, and refine any of them.
And that's the other kind ofmore exploration heavy thing.
Now, in practice, what you wannado is balance those two, right?
This is the classic reinforcementlearning And so what this paper's going
to do is ask the question, what if wecould vary in a principled way, the
number of different, kind of branchesthat we try versus the depth that
(01:07:56):
we push at in each branch, right?
What if we could trade off in a principledway, exploration versus exploitation
in a tree search setting, right?
So that's what they're gonna do.
Normally what you do in multicolor treesearch is you'll like fix the number of
child nodes with a fixed hyper parameter.
So you'll, you'll do that, thenmaybe you go to the next level,
or you pick one branch and expand.
(01:08:16):
But what they wanna do hereis kind of dynamically do it.
And so.
It gets a bit involved, but thefundamentals are, they actually have an
internal model that is being trained alongwith the main training loop to predict the
value of creating a new node of creating,like doing more exploration in other
words, versus doing more exploitation.
(01:08:38):
And they have different modelsessentially that predict the value
of each, for each part of the tree.
So they kind of decompose the tree andgo like, okay, in this chunk of the
tree, you know, what is the value ofadding a new node versus the value of
pushing forward in the existing nodes.
And essentially this is themechanism that they're gonna use.
really interesting actually, thedetails, if you wanna look into how they
(01:08:59):
implement it, it's quite interesting.
It's called, they, they use this techniquecalled Thompson's, sampling, which
is, sort of Bayesian friendly way ofmaking this kind of principle trade off.
And so, the results are really impressive.
they end up basically like.
outperforming on average allmodels in terms of, uh, the
benchmarks that they look at.
The best average ranksreally across all benchmarks.
(01:09:21):
They, yeah.
Anyway, they, they compare it toall the variants we talked about,
just repeated sampling from the samenode or just like doing sequential
refinement the whole way through.
Or a standard Monte Carlo tree searchwhere you fix the number of, uh, branches
at each layer, all that sort of thing.
of the really amazing things is the arc aGI sort of experimentation that they do.
(01:09:42):
one last detail I gotta give you.
So as they're doing this right, you,you imagine like you start off at
a given node, you can ask yourself,should I should I try refining the
solution or should I try going up alevel and spawning a new alternative?
Another question you could askyourself if you're spawning a new
alternative is, which model shouldI use to spawn this new alternative?
(01:10:04):
And they're actually gonna include that.
They're gonna allow the model.
To build another model topredict which model it should
use to spawn those branches.
And in that way, essentiallyintegrate together.
They're, they're creating like acomplex of systems here or a complex
of models integrating togethermany submodels where at each node
you have the choice of explore.
(01:10:26):
and then within explore, whichmodel do I use to explore or
exploit and refine, right?
So that, that's kind of a veryunusual sort of hybrid model approach.
On, uh, on Arc a GI two, they end updoing like really well, like scoring
around 30% on, Pass at two 50.
So 250 at bats here.
With a, basically a number ofLLM calls budget of 250 calls.
(01:10:49):
They end up hitting like around30%, which is pretty wild.
That performance, by the way, is highestwhen you allow the model to use the most.
Or the framework, I shouldsay, to use the most models.
So when they combine O four miniGemini 2.5 Pro and deep seek R
one, the May 28th version they endup getting the best performance.
As you start dropping modelsfrom that list, you start to see
(01:11:10):
progressively worse performance.
And so this really is a frameworkfor getting many models to play nice
together and to choose which model touse to expand nodes and also whether
to expand nodes or iteratively refined.
So it, it's a really interesting paperin a sense, a meta paper or a metamodel.
I just, just thought thisis a, a fascinating read.
(01:11:31):
Yeah.
This is, uh, somewhat significant,practically speaking because
this is one of the standard waysto structure reasoning, right?
And, on the usage side, what this lookslike is on their benchmarks, for instance,
on coding and on RKGI is sort of quasipuzzles that measure intelligence.
(01:11:53):
The meaning of tree search would be,you know, a given node is a solution and
you're able to sample multiple solutions.
That's the breadth, the width,or you can iterate on a solution
that's with depth and you, youshould be getting some feedback such
as test outcomes or, or scoring.
And one of the limitationsof searches you do need.
(01:12:17):
A scoring function of some sortto base the, tree expansion on.
So as you said, uh, anotherinteresting bit here is kind of
the adaptation of multicolor treesearch to the context of LLMs.
It's not super kind of intuitive becauseyou can, you know, within a node also
(01:12:37):
generate more tokens potentially.
and there's various nuances, quite abit of detail in the implementation.
But the gist is it allows youto, in a very principled, sort
of classic way, Monte collegeresearch is, with default, right?
It's, it's one of the big algorithms as isThompson sampling very highly researched.
(01:12:58):
So they, very much like adapted ina very clean way for reasoning and,
and seemed to get good results.
Next we've got the automatedLLM speed running benchmark,
reproducing nano GPT improvements.
So they want to evaluate an AIagent's ability to reproduce
(01:13:22):
scientific results, and theyfocus on not OGBT speed run tasks.
so they want to test the reproducibilityof scientific results by ai.
That's a motivation, right, to kind ofget the results of a given, research paper
(01:13:43):
with your own implementation using ai.
And this, uh, non GPT Speed Runtask is designed to assess the
efficiency and accuracy of AI modelsin repeating scientific improvements.
So, now we have abenchmark for doing that.
And as with any benchmark, it setsa goal to target and, at least
(01:14:05):
for now, has room for improvement.
Yeah, I like to think of this as a sortof companion to or compliment to the,
uh, meter AI evals when they look atyou know, what is the time horizon that
AI agents can successfully automate on?
So, you know, tasks that take humans anhours tasks that take humans two hours
it, they tend to look at AI research tasksfor that because they're interested in
(01:14:28):
the sort of recursive self-improvementloop where we get to the point where AI
is just automating all of AI research.
This is in that same spirit.
And so nano GPT is essentially a, likea version or an an instantiation of,
GPT two, an implementation I shouldsay of GT two in PyTorch that Andre
Carpathy put together a while ago.
And, so what's happened sinceGT two came out is we've had a
(01:14:51):
series of breakthroughs, right?
A series of innovations wherewe've iteratively improved the
training time for GT two from 45minutes to under three minutes.
And there are 21 successiveoptimization records that have happened.
So you think of it as like 21, stepsdown the ladder that people have taken.
(01:15:12):
And each of the tasks associatedwith this benchmark tasks, whatever
agent you're testing with reproducingone of those 21 steps, right?
So, starting with the originalg PT two or at least its nano
g PT version, modify that.
In order to achieve the first speedboost that we got historically.
(01:15:32):
And then you see, okay, well what fractionof that speed boost did the model,
did the agent successfully recover?
And so that enters your, sortof evaluation framework and
then you, you try the next one.
Okay, so, so now that we're here, let's,let's reset, let's give the model the
correct sort of version of GPT twothat reflects that next step, that next
breakthrough, and start from there.
Now replicate the second breakthrough.
(01:15:54):
See how, what fraction of thatacceleration the model is able to recover.
And then repeat all the way down theline 21 times and then add up together.
Or one eval would be to add uptogether all the time savings that
your agent was able to achieve.
Compare that to the, time savings ofabout 42 minutes that was actually
achieved by, human researchers over time.
And then use that essentiallyto compare different models.
(01:16:16):
And so, yeah, this isreally quite interesting.
O three mini is the best performer.
They, they looked at here.
It recovered about 40% ofthe speed up with hints.
So this is a variant that they have.
They have versions with likethree different layers of hints
that you can give to the model.
and then also.
Versions with no hints at all.
And, you know, threedifferent layers of hints.
It's like one is a descriptionof the pseudo code that would
(01:16:38):
be required to reproduce it.
the second layer is naturallanguage explanations, and the third
layer is a full mini paper formatwith all the technical details.
So gradually kind of givingthe model more and more.
But there is a version where theyjust, zero shot it and you're
not given any, any more context.
And so, CLO 3.7 sonnet alsocomparable to O three mini
apparently in some conditions.
(01:16:59):
but deep cq R one basically just likedoes really badly, sometimes does even
worse with hints than without, whichitself, it's kind of interesting.
And then Gemini 2.5 Pro, basicallyjust bombed it, got basically
zero in their aggregate measures.
And so all kinds of interestingobservations about what works, what
doesn't worth looking at the paper.
This is kind of more, I think ofit as more encyclopedic knowledge
(01:17:21):
that will go bad pretty quickly.
but certainly the benchmark itselfseems like a really important
and interesting contributionthat I would keep an eye out on.
you know, again, think ofit as a compliment to those,
very famous meter evals.
I think it's a greatway of looking at that.
Right.
And also just a fun, I guess.
Way to do this.
I did not know that there was aspeed running challenge to train GPD
(01:17:42):
two in the shortest amount of time.
That's where Nano G PT comes from.
'cause by these days, GPD two isconsidered nano, uh, I believe
GP two was what, like one ishbillion, 2 billion parameters.
So the, the version they do here isthe, 124 million parameter version.
Right.
But yeah, tiny by today's standards.
(01:18:03):
Right.
And speaking of meta and theirevaluation suite, that is our next story.
We have just an update.
So they posted saying that, just toquickly recap, they released a paper
several months ago now measuring AIability to compete to complete long tasks.
(01:18:24):
And they basically have a task suitewhere they roughly know how long a
given task takes, could be five minutes,could be 10 minutes, could be an hour.
And they measured the ability ofvarious models to reliably complete
those tasks with, for instance, likethere's a 50% chance you are going
to get this done in an hour or less.
(01:18:46):
So, since that release, there'sbeen a couple months have passed
and they published an update where.
Claude four Opus now reaches,50% time horizon of 80 minutes.
So 50% chance it completes an 80minute task, uh, in a time span.
(01:19:06):
Sonnet reaches, uh, the sonnetfour reaches the 65 minute, point.
So they're now exceeding an hour,you know, slotting into the,
trend, the kind of prediction,fit that came out with a paper.
Yeah, this is really interesting becausethere's so few data points, right?
(01:19:28):
Like every frontier modelbasically is a data point.
That's, that's where the bar is.
And so figuring out exactly what thetrend says is really hard and even small.
If you look at the plots, eventhese like small adjustments, right?
In the slope of that log plot arethe difference between hitting, you
know, uh, a SI or, or hitting, let'ssay, AI agents that can do month long
(01:19:51):
tasks coherently in say, two yearsversus three or four years, right?
So like, things can bevery sensitive to that.
So these small little updates, everytime you get a new a new model, you
wanna fit it to the plot really quicklyand be like, oh, how does that affect
the, you know, the slope of the curve.
Certainly when oh three cameout, that was a big, big update.
If you look at the, plot, you know, ohthree seems to, in concert with other
(01:20:13):
models like Sonet 3.7 and oh one, inSonet 3.5 from back in October, really
seems to suggest there's an even steeper,trend than is otherwise, indicated.
Claude four Opus does not,by the way, beat O three.
It actually, so O three isabove an hour and a half.
Claude four Opus, uh, is, uh, fewminutes shy of that, like a, I don't
(01:20:33):
know what, an hour and 15 you said Andre?
Something like that.
Mm-hmm.
it's notable, you know, it hasbeen a little while and we're
not, we're not exceeding now.
I would say that's still withinnoise looking at the plots, but it
could make you update a little bit.
This, this is the debate that'shappening right now, right?
People are trying to figureout what this means exactly
and probably overthinking it.
We're gonna have to waituntil the next open AI or next
(01:20:54):
philanthropic ag agent model drops.
But definitely you know, this is somethingreally, really to, to keep an eye on
just because of the implications, right?
If these, if these curves reallydo curve the way they seem to,
then we could be in for a hell ofa party over the next few years.
And how much of a party, well,that's contingent on a relatively
small number of data points.
(01:21:15):
Yeah.
And, uh, you know, this is a trickything to evaluate obviously, because,
you know, you have a set of tasks thatthey evaluate on and, you know, how
do you really know how long it takes?
But my two sense is, it seems prettyplausible from using cloud code like that.
It can autonomouslyfinish a one hour task.
It's definitely gettingthere, in my opinion.
(01:21:37):
And next story, a research paper.
And this one is titled PerformancePrediction for Large Systems
Via Text to Text Regression.
The gist of it is the focus is onpredicting the outcomes of some
sort of configuration of a system.
So for instance, you have a cluster,you have some way to configure
(01:21:59):
a cluster, and you wanna be ableto predict latency of that setup.
tricky task, very usefultask to do well on.
And they are training a model that doesthat for you based on, system logs and
things like that go from previous datato a prediction of, your performance
(01:22:21):
with a new setup and they get, as youmight expect, really good, correlation
and performance with this approach.
Yeah, I guess the core of this is adebate that's been happening for a long
time, in especially language modeling,but elsewhere too, as to whether the,
decoder only architecture is the rightway to go, or whether you should use
(01:22:44):
an encoder decoder structure, right?
So an encoder decoder is model that willstart by, through many layers, specialize
in just generating a really good encodingof an input, and then have separate
layers that specialize in massaging,that encoding to turn it into a decoding
in different clumps, essentially indifferent stages with an optimization
routine that reflects that intent.
(01:23:05):
So the advantage people will claimfor the decoder only version is you
have essentially like an integratedthing, like one model that's able
to to kind of address dependenciesand interactions between the lowest
layers and, and the highest layerswithout having to go through a, a
bottleneck where you have a well-definedencoded latent representation.
(01:23:27):
whereas the encoder decoder sidewould say, well, uh, it's good
to have specialization so youcan make a really good encoding.
And then separate that out from the,the decoder step, which is kind of
this fundamentally different operation.
And I mean, the answer is towhich of these approaches is best?
Does seem to be context dependent.
Certainly this paperstrengthens that argument.
Uh, what they're gonna do here is havetwo encoder layers, again, that specialize
(01:23:50):
in just taking in this semi-structureddata about the state of a system.
As you said, they used Google'sBorg compute, architecture as
their testing ground for this.
So they use system logs as inputs.
They use all kind of like, theequivalent of like check engine lights
and things like this that they feed in.
And some of this data is in text form, bythe way, but this is not a language model.
(01:24:11):
It, it's not gonna learn tounderstand the semantics of the text.
It's only gonna learn to understand thoseindicators insofar as they are correlated
to the one metric, the number that themodel ultimately is going to predict.
And so this is not anauto regressive model.
it's taking these raw inputs and it'spredicting a number, which is like kind
of a measure of the efficiency of theoverall system, the predictive efficiency.
(01:24:34):
So you've got two encoderlayers, two decoder layers for
60 million parameters in total.
So relatively small model.
so this is a really effective way itseems of making these predictions on
how the system is gonna, is gonna work.
The Borg cluster scheduling system, thisis the kind of source of data they use.
It's like big, cluster or sort of, uh,orchestration system that Google uses.
(01:24:57):
Is, all the, the raw data that they'reusing to, to train this, they do
it using cross entropy loss overresponse tokens that they get from
the system that indicate how it'sdoing, that indicate like what the,
status is of the overall system.
And so, anyhow, it it'spretty interesting.
It's, it is more of a, a niche.
Sort of application, it's showingthat just using a raw LLM trained
(01:25:19):
from scratch on text data is notnecessarily, the best play if you
have a, a more structured problem.
Again, this is not perfectly structured.
You still have language inputs, but thoselanguage inputs, again, are you, you think
of them more as categorical variablesand that's sort of the interpretational
frame that they're applying here.
Next up just a couple more papers.
(01:25:41):
The next one is, does math reasoningimprove general LLM capabilities?
Understanding transferability of LLMreasoning, we've really gotta get going.
So just a very short gist of thispaper, they are exploring if you trained
specifically to do better on math,are you gonna be able to do better
(01:26:02):
on reasoning in general outside ofmath, like, I don't know, uh, science
problems, for instance, or coding.
They find that depending onhow you do it, uh, you might
actually get negative transfer.
So supervised trainingjust doesn't work as well.
compared to reinforcement learning.
Reinforcement learning kind of has amore subtle effect that doesn't mess up.
(01:26:26):
your initial model as much and generallyseems to result in better transfer.
And last up, correlated errors inlarge language models and other.
A kinda empirical analysis paper.
They are looking at thecorrelation, uh, of the errors
among different large languagemodels using several, uh, data sets.
(01:26:51):
And they look at 349 LLMs on12,000 multiple choice questions.
And so the question is like,I'm wrong, different lms.
How similar are they interms of what they get wrong?
They found that thecorrelation is pretty high.
Uh, models agree on incorrect answers.
About 60% of the timeon the helm leaderboard.
(01:27:15):
So much more likely thanthen, random chance.
And, suppose not entirely surprisingperhaps, but still interesting from,
empirical analysis perspective.
Yeah, absolutely.
I mean, this may, it remains true tooregardless of architectures used and
and model developers, which sort ofleaves, I mean, and, and presumably
(01:27:38):
optimization routines as well, sothat, that basically leaves like.
The data.
Right?
And, and that it kind of makesa lot of sense right there.
There's only so much internetdata and everybody will be using
a highly overlapping data setsas part of their, their training.
So in some sense, maybenot surprising in others.
I mean, one of the things you would expectis, like, I'd be interested in seeing the
(01:27:59):
overlap of this with just general, likefrequency of errors from these models.
Because as the frequency of errorsdrops, the errors themselves get
more and more rare and scarce.
And so you're getting a more and moredistilled picture of the, I mean,
I don't want to, it's definitelynot the irreducible entropy of the
training, data, but it's, it's sortof gesturing in that direction.
(01:28:19):
So anyway, kind of, kind ofinterested in, in, uh, what that
looks like if, if they're plottedtogether, but interesting paper.
Yeah.
Yeah.
And they, they, uh, focus, inparticular on this area of, uh,
job applicant screening, as kindof a outcome of this analysis.
And as you might expect, having lowercorrelation is better because if you have
(01:28:43):
lower error correlation, it means youcan look at several LMS and potentially
avoid an error because one LM getsit wrong and everyone gets it right.
And the gist of it is you know, youneed to sample quite a few LLMs to be
able to get to lower error because ofa decently high correlation among them.
(01:29:04):
And up next onto the policy andsafety section, we're starting with
forecasting biosecurity risks from LLMs.
Okay.
So there's been a lot of talkabout whether LLMs actually do make
it more likely that, bad actorsare gonna be able to design or
release more dangerous bio weapons.
And, you know, famously therewas this like study from a year
(01:29:27):
ago that said, guys, don't worry.
It's, or not, don't worry,but like, good news.
There's no meaningful uplift from, andI think at the time it was like G PT
four or something, and then opening AIcame out with something a few months or
weeks later saying, actually we triedthis, or something analogous to it.
And we have access to the fullunlocked version of gbd four.
And we do get a quite significantdetectable increase in the probability
(01:29:51):
that people with a, a little bitof training or, or a significant
amount of training are able toaccess, dangerous bio weapons.
And so, this is another take on this.
and we've seen, by the way, otherbenchmarks too, in system cards from
anthropic, from opening AI sincethat have meaningfully increased that
even further and quite significantly.
So this is another take on it.
Instead of throwing these, throwingthese models directly at tasks that
(01:30:12):
people think will be correlatedwith high bio weapon risk.
What they're doing is they're turning to abunch of experts biosecurity and biology.
So 46 domain experts in biosecurity andbiology and 22 expert forecasters, right?
The so-called super forecasters.
And what they look at is, okay, forall these folks, we want to get you
(01:30:33):
to predict, the probability of a humancaused epidemic causing over a hundred
thousand deaths, by, I think it's 2028.
Yeah.
And so they, that's the,you know, the base question.
A whole bunch of other questions,correlated with that or, or
that, that follow from that.
But that's kind of the base question.
That's the, the meat and potatoes.
(01:30:53):
And then they divide the group up or, orthe overall group up into, in different
ways to look at how how accuracy, or how,sorry, how that predicted probability
changes depending on who you ask, right?
So, for example, the overall, assessedprobability was, uh, somewhere between
it was around 1.5% with AI and 0.3%without, so they're predicting a very
(01:31:16):
significant increase in the probabilityof a, you know, 100,000 deaths from,
again, a human caused epidemic by2028, which is quite significant.
but it turns out that the people who most.
Believe who, who assign the highestprobability to this, are also the people
who are more accurate on predictingthe progress of large language models.
(01:31:39):
They're also, the people who havethe most experience in biosecurity.
They're also the people who get,highest accuracy on, low probability
questions that they, they askotherwise in, in the survey.
So.
That's kind of bad news, right?
The people who are the best at forecastingthis sort of stuff tend to assign the
highest probability to this ultimately,you know, leveling out at, around, you
(01:32:01):
know, one to 3%, something like that.
But, but still significant.
And so, this is something that doessuggest, hey, there is meaningful
uplift, at least accordingto these, these forecasters.
So take it with a grain of salt.
The last thing I'll mention is they do saythat, uh, mitigation measures are probably
gonna be enough to buy down this risk.
(01:32:21):
When they were asked about mitigationmeasures, like including mandatory
screening of synthetic nucleic acidorders, and just basic AI model
safeguards, they reduce their riskforecast back to close to baseline level.
So they basically figured if you do putin the right mitigation measures, you
should be able to essentially buy downall the risks that comes from this.
I'm personally reallyskeptical about that.
I think people very much, sortof overestimate the effectiveness
(01:32:46):
of a lot of these safeguardsfor reasons we could talk about.
But yeah, anyway, I think a reallyinteresting study, and again, a
fundamentally different angle, right,from these more empirical studies that,
that Rand has put out, that OpenAI hasput out, that Anthropics put out that
are in their own right, very useful.
And Rand gets credit for,for kicking off that trend.
so many, months, I, I wannasay over a year ago now.
(01:33:07):
Yeah, quite an interesting read.
And they do go into quite a bit of detail.
So they start out with thisunconditional question of what's the
probability of a hundred thousanddeaths due to a pathogen in 2028?
And then, then they then conditionon various hypothetical advancements
in n LMS to see what the changes.
(01:33:29):
So, they begin with this 0.3% baselinethat rises to 1.5 conditional on the
several hypothetical animal capabilities.
And then musically, they thenve check actually they have.
And the, uh, people respondingto the survey thought that
would not happen until 2030.
(01:33:51):
So, that is an interesting data pointsaying that maybe the forecasters are
underestimating the, uh, uh, degreeto which LLMs are moving and, and are
able to achieve these advancements.
Which could color your predictionor at least would mean that this
1.5% probability, is their actualprediction given the state of lms.
(01:34:17):
And obviously, you know,people listen to podcasts.
You're aware of my bias on this.
Like I, I do think that AI is movinga lot faster than most people realize
or want to admit to themselves.
And one tell is things like this, thishappens over and over and over again.
You'll have people say like, you know,oh, we're, we're not gonna have this.
When you actually get people togive you dates by which they think
(01:34:39):
certain capabilities will emerge,they tend to just like hilariously
predict like, oh, it, it'll beanother 10 years, another five years.
And the usual case is like, the thinggets done in like a month or two.
But in this case, this is sucha great example of it because it
literally had already happened.
It's so hard to keep up with thespace in fairness, like it does move
(01:34:59):
that fast and, and we ourselves aresurprised by things all the time.
But that's kind of partof the problem, right?
That you need to have some amount oflike, yeah, I guess, uh, epistemic
uncertainty, when it comes tothis stuff and, and factor it in.
If you find you continually getsurprised by how fast things are moving,
then you know, maybe that implies youshould just change your world model.
And, anyway.
I think a lot of people arebanging that drum these days.
(01:35:20):
Yeah.
And, and the specific hypotheticalshere are quite specific.
so we're talking about AIenabling 10% of non-experts to,
synthesize the DNA fragment.
I think of some influenzafrom 1918 in a laboratory.
They are looking at the virologycapabilities test, which came
(01:35:45):
out just a couple months ago.
So these are, you know, not sort ofjust like, oh, you do this well at, some
coding benchmark, very specific to wetlab work, virology work, things like that.
which is obviously quite relevantand I think does lend this,
more credibility as an analysis.
(01:36:06):
Speaking of predictions, nextwe have AI task length horizons
in offensive cybersecurity.
So this is, an adaptation really ofthe methodology of a matter we just
discussed about predicting the lengthof time, tasks that alarms can do.
(01:36:26):
this is less formal just FYI, it's,it's a blog post by just one person
of a kind of estimated length oftasks for various, benchmarks by the,
sorry, this, this is less formal.
In a space where it's like, now the,the formal version, of course, is a
pre-print slapped together and thrownon the archive without peer review.
(01:36:47):
Right.
Just formal.
It's just funny that that'slike Yeah, yeah, yeah.
I, I'm just, uh, outlining it because theblog post itself itself Totally, totally.
I had this double take where I agreed withyou, but then I was like, wait a minute.
Like, what's the bar?
Yeah.
It's not like these things arebeing released in journals.
But anyways, in this slightlymore informal, analysis, they have
(01:37:12):
tasks ranging from 0.5 seconds to25 hours in human estimated times.
And they are seeingthat, still pretty early.
the current models can solve six minutetasks with 50% success rates, but,
as with matter, you know, you can doa little analysis showing that these
(01:37:37):
models are likely to double every sixmonths or, or sorry, four months or
so this is exactly the debate Yeah.
That we were talking about earlier.
Right.
How do you fit that curve?
Yep.
Some ways of fitting it.
You get four months.
Some ways you get six,some ways you get seven.
It's, uh, yeah, it's pretty unclear.
Yep.
So there you go.
And more empirical analysis, and obviouslyrelated to safety in a sense that
(01:38:02):
cybersecurity is a huge challenge and LLMsare kind of an obvious fit for hacking,
for things like that, where unlikevirology, for instance, where you need a
wet lab and you need to like work with ahuman here, you could very easily see an
agent going off and doing some hackery.
Yeah.
And, uh, with the launch of AI coding aswell, it's gotta be a lot of cybersecurity
(01:38:29):
stuff going on in the next few years.
It's fine.
Everything's fine guys.
Uh, yeah.
Yeah.
This benchmark, by the way, is, in myopinion, extraordinarily, badly needed.
Any threat model that you have that runsthrough, you know, ai, self replication,
loss of control, weaponization, right?
The cyber use case is.
(01:38:50):
Arguably, and, and you would have toargue this, but arguably, is the most
sort of real and present that you mightexpect impacts from in, in the near term.
And so you should be very frigginginterested in measuring how,
successful these models are at longhorizon tasks that look like capture
the flag challenges, uh, that looklike malware generation challenges.
(01:39:12):
You know, natural language to bashtranslation, you know, that sort of thing.
Uh, so they look at five differentbuckets of tasks that have different
time horizon characteristics, fivedifferent benchmarks, I really should say.
So there's like s bash bench, which isthe shortest timeline tasks, which were
actually created by the author of thispaper, presumably because there just
aren't tasks short enough that like GTtwo can do anything meaningful whatsoever.
(01:39:36):
So that's like one to 32nd tasks.
At least that's how, how he assessesand, and, you know, we could get into
this and at some point we may, but.
Uh, the open question is always, how doyou assess the amount of time that it
takes for humans to complete these tasks?
That's itself a very interestingquestion, especially as you get into very,
very short and very, very long tasks.
(01:39:56):
But anyway, NL two bash,enter code, capture the flag.
NYU capture the flag sidebench side bench, by the way.
Interesting because the task lengththere range from two minutes to 25 hours.
So you're really covering quitea wide, kind of temporal range
from models from 2019 to mid 2025.
And it's all the usualcurve fitting stuff.
Check out our first, podcast onthe meter evals where we did a
(01:40:18):
deep dive into their methodology.
Uh, that'll give you a good sense forlike how this is being assessed here.
It looks like a five month doublingtime here, so, it was six minutes today,
but it's doubling every five months.
so the five month doubling timesuggests that you would reach a week
long task within about five years.
There's a lot of caveats here.
I think one of the really interestingthings to note though, is that
(01:40:40):
the time horizon here is so muchshorter than the time horizon we
see with the meter evals, right?
The meter evals are showing ushour and a half for oh three
as we just talked, about.
And yet here we're talkingabout like six minutes.
So yeah.
What's the delta there?
Well, you know, part of it is thelabs are directly optimizing for
recursive self-improvement, right?
This is not even a secret.
This is just straight up what theywill tell you in their blog posts,
(01:41:01):
which itself is a super dangerousthing to do and shouldn't be done.
But that's part of the roadmap.
so there, you know, there'soptimization pressure pointed
directly along that axis here.
You can think of the cyber capabilities,the offensive cyber capabilities, at
least for these models that are kindof public and publicly used, as being
a side effect of optimization againstthe kind of core a GI benchmarks.
(01:41:23):
And so that's one reason whyyou're not seeing necessarily
the same impressive uplift here.
But you're still seeing the doubling time.
That's quite interesting, right?
It suggests some robustness to the broadertrend of exponential coherence length
increases in, in these, uh, AI agents.
Yeah, I think as a matter, likesimilar caveats, even more so here
in terms of like the themselvesof, of time to complete by a human.
(01:41:50):
be reliable.
And here, there's only oneperson able to do that.
The task distribution is also moreheavily, leaning towards the short side.
So, uh, yeah, quite like a largemajority of the tasks are between
one second and 10 minutes.
And, you do get a decent amount goingup to one hour, but then outside of
(01:42:15):
like two hours, you get very few tasks.
Yeah.
So, yeah, very impressiveeffort by a single individual.
But, do take it with a grain of salt,regardless, clearly the case that LMS are
able to do a bunch of cybersecurity stuff.
All righty.
Just a couple more things.
Moving on to the policy side.
(01:42:35):
and we start with, the US where we'vebeen dealing with the saga of the one
big beautiful bill, which just passed,uh, yesterday through the house and is
going to, president Trump's desk to sign.
The one big beautiful bill is primarilyabout the budget, about various
(01:42:56):
tax cuts, for the rich and variouscuts to services by the government.
But in it tucked away, there was, asection that we covered previously
that would have banned regulationof AI by the states for 10 years,
I believe, which, uh, became a bitcontroversial once it was highlighted,
(01:43:20):
it was, then removed from the bill.
The Senate voted nine nine to one toremove this proposed 10 year moratorium
on state level AI regulations.
So it's out.
and this article goes into the aggressivelobbying for the moratorium led by A 16
Z and Meta and others, and, and, uh, itmust be said seemingly open AI as well.
(01:43:47):
this is quite interestingbecause the case has been made.
So there, there's this, guy, Adam Terrier,I wanna say, who's kind of famous in DC
for having come up with the idea of a 10year state moratorium on on AI regulation.
His big claim has been to trot aroundthis number where he says that there are
over a thousand state level bills thatare imposing regulations on ai and that
(01:44:11):
this would create an untenable mishmashof like state level like regulation
that you had to have to adhere to.
And then of course there may be a federal.
Package that comes through at some point,and this makes it impossible for small
companies as Andreesen Horowitz calls themLittle Tech, uh, to compete in that space.
There's just a little problem.
That 1000 figure is, for all intentsand purposes like embarrassingly made
(01:44:35):
up, it basically seems to come froma search of a database of like state
level regulation that just like uses ai.
So a lot of these things are justusing the term ai, defining it, or,
uh, even in some cases finding ways toadvocate for it, to get it to be used
in education and things like that.
And don't actually introduce anymeaningful constraints on company's
(01:44:56):
ability to use or develop ai.
And so it's sort of disingenuous,frankly, to talk about it as like
there's a thousand things likethis when you whittle it down.
It seems like the estimatesI've seen are around 40 of these
things that actually are material.
The majority of those willnot actually pass either.
So the, the number gets whittleddown pretty fast from by about
like two orders of magnitude,which I, I think is significant.
(01:45:18):
The other piece of this too is theargument was historically that like
rather than having the states regulatethis, we should regulate at the federal
level, pass a law at the federallevel, which sounds like a great idea
until you realize that the federalgovernment has been gridlocked on the
issue of AI regulation legislation for.
Forever, right?
It is, I mean, it's beenfive years since GBT three.
(01:45:39):
It's been three years since chat.
GBT.
We're still in this endlesscycle of having committee
testimony and hearings and, andinvestigations and all this stuff.
And it never really goes anywhere.
And this is a recognizedpattern in the Valley.
People know, lobbyists know that there'sthis gridlock at the federal level.
And so by saying, Hey, let's, preemptany state level legislation for 10
(01:46:00):
years, by the way, which is likeopen AI internally believes, super
intelligence gets hit within like,you know, five years tops, more likely
three years, something like that.
So the idea that like at the statelevel, there's no regulation.
10 years seems pretty insane.
And then there's a basic questionor fact of the matter that
states are different, right?
I mean, California, which has OpenAIin their borders, which has, you
(01:46:23):
know, a lot of big labs in theirborders versus Idaho, which, you know,
or Virginia, which has a bunch ofdata centers, but no Frontier Labs.
Obviously these states are gonna havefundamental differences in the way that
they need to regulate and legislate ai.
So it actually does make sensethat you should have some freedom.
like I'm old enough to rememberwhen states rights was, was a thing,
among kind of more libertarianleaning people, such as actually,
(01:46:44):
you know, myself in this space.
So, yeah, it, it kind of like, seemslike a weird play to try to strip
away state's rights, fortunately and.
And I think this is a, a reflectionof just, good sort of education
in the Senate on this issue.
This was thrown out, very recently.
O overwhelmingly voted against thisprovision on a 99 to one margin to
rip out this, state level preemption.
(01:47:06):
So I. That's a pretty remarkable defeat.
By the end of it, I, I think Ted Cruzwas sort of like championing this
thing forward using the, China scaryline, which I actually take as well.
Right.
You know, if you're tracking the workthat we've done, uh, nobody is more
on the China scary camp than we are.
but there's a, a, a kind of fundamentalmisunderstanding here of the role
that state level AI legislationcan play in a context where there's
(01:47:30):
nothing happening federally.
Like this is just being real about it.
That's the consequence.
And you have to imagine that's exactlywhat the, you know, the companies
that have been lobbying for this,especially entrees and Horowitz,
were thinking, and now the backlash.
You know, the problem is when you dosomething like this, it is so obvious to
people that what you are trying to do islike, lock in a competitive advantage for
(01:47:53):
the, you know, the open eyes of the world.
That yeah, now you'regonna get a backlash.
What this looks likeis exactly what it is.
It looks like Mark Andreessen steppinginto the federal level, trying to
backdoor his way into putting some prettyextreme legislation on the table that's
reactionary in, in just the same way as alot of like burdensome regulation that's
(01:48:14):
been proposed at the federal level wouldbe reactionary in the other direction.
And so.
I, I think that this could actuallybackfire in some pretty concerning ways.
You just gotta be more carefulabout this and, and especially
like, you know, taking the, thetemperature of the public on this.
People are, are interested in regulating.
And so this doesn't kind ofmatch the public perception.
So I'll get off my soapbox, but I,I just, I feel like this is a, a bit
(01:48:36):
of a, an own goal for the people whoare looking for this sort of thing.
Like, federal nullification of state lawsin this way without a federal framework in
place is straightforwardly unprecedented.
Like this would never have been done.
And so the bet here is literallythat like we are so confident that
we don't want any state level AIlegislation over the next 10 years as
(01:48:59):
super intelligence may come and go.
We are so confident that we won'tneed that, that, we'll, we're gonna
lock it in at the federal level.
Like, that's, that's some balls,dude, that's some real, real balls.
I, I wish that were the case.
I I think we need a little bit moreflexibility on this and, and, you know,
Ted Cruz is doing his best and everybodyis, but I think ultimately it just
reflects a misunderstanding fundamentallyof, of the trajectory of the technology.
(01:49:22):
Yeah.
And worth noting, this is particularlyimportant in the case of the US
because there is no real federalregulation, uh, at the national level.
And there is not going to be any,uh, least until this president is
out of office, just based on who hechose to be leading on the tech side.
(01:49:44):
so effectively the states are theones doing any sort of regulation.
Uh, for instance, by the way,I, I'm a little skeptical.
Like I think that the, the feds may wellcome in and regulate this, but the point
is that they retain the option to, right?
Mm-hmm.
They're not like saying, Hey, we'renot gonna put in any regulation
for 10 years blanket statement.
(01:50:05):
Like, that's what's insane to me, right?
Yeah.
It's like, let's enshrinethis in the law of the law.
Like what?
Like it just literally, like, let'stake options off the table, right?
Sorry.
Yeah.
It just seems so, it's also a bipartisanthing, like Marjorie Taylor Greene
famously came out and said, Hey, like.
I, I voted for this bill beforeI realized that this fucking
(01:50:26):
crazy thing was in there.
Now that I see it, I'm like, holy shit, Inever would've, here's their quote, right?
I'm not voting for the development ofSkynet and the rise of the machines
by destroying federalism for 10years by taking away states rights
to regulate and make laws on all ai.
Like that's, this is a bipartisan issue.
That's why it's 99 to one in the Senate.
This is an insane, like it to, tothink you can backroom your way.
(01:50:49):
I think this turns a lot ofpeople off the, the kind of, uh,
Washington and lobby, treadmill here.
it just kind of seems to be exposedfor an, an attempted, I won't just
call it like a fraud on the Americanpeople here, but this is like an
undemocratic play that was attempted.
and that's not good, youknow, I mean, anyway.
Yeah.
Yeah.
So, uh, also just.
(01:51:11):
Kind of weird that they try to sneakus in in the budget reconciliation.
Uh, there's like nuance where thestates would lose some federal
grants if they go against it.
Also notable because in hindsight orlooking back so far, the biggest effort
to regulate AI was last year with SB10 47 in California, as we've covered,
(01:51:32):
that was defeated in significantpart due to large lobbying by tech.
It went to, as far as the governor,the, the governor vetoed it.
So this would basicallyhave prevented that.
Right?
And, and so.
Likely it's gonna happenagain in California.
There are efforts to do a revamp10 47, and, and quite significant
(01:51:55):
in that context as well.
And, and, and by the way, just like to,to make the point, if anybody is clutching
their partisan pearls on this, right?
This, again, not a partisan thing.
SB 10 47 obviously passed the highlydemocratic, legislature in California,
but it was vetoed by Gavin Newsom, right?
Like the, the most liberal governor,basically in the entire country under
pressure from Nancy Pelosi, among otherpeople at the federal level writing in
(01:52:19):
and, and saying it should be scrapped.
So it like, this is a reallyfascinating issue in that it just
crosses nukes, all partisan lines.
I love issues like that, by the way,because they prevent us from just
seeing things through this lens thatwe, you know, we want to see it through,
you know, Republican versus Democrat.
The reality is that'snot what's happening.
but figuring out what the right playis, I, anyway, I genuinely feel bad for
(01:52:40):
anybody who's in the, in the hot seatof having to make the calls on this.
They're hard to make, but surelypreserving optionality ought to
be part of our, our basket here.
Yeah, we'll say, I dunno if GavinYusef is the most liberal, but in a
very democratic state, to be fair.
anyways, moving on.
One last story.
it's about Denmark.
(01:53:01):
They are going to tackleDeepFakes by giving people
copyrights to their own features.
So we're gonna be amending copyright law.
To give individuals rights over theirown body, facial features and voice.
Uh, it's one of the first, initiativesof its kind in Europe has brought support
(01:53:22):
and it is gonna take a little while.
It's still being submittedfor consultation and will be
formally submitted in the autumn.
kind relates to someefforts, uh, elsewhere.
Certainly in in Hollywood there's beennegotiations, but to my knowledge,
not too much on the law side as faras copyright over your own appearance.
(01:53:46):
Yeah.
How, how you actually define the bounds onthat too is gonna be really interesting.
Right.
That's always a AI has a way of likefuzzing the boundaries around everything.
And so, you know, how much can you modifya face until it's not your face anymore?
All this stuff.
But yeah.
Really interesting.
'cause we, we accept makeup, right?
We accept hairstyledifferences, all this stuff.
So at what point is AI an adornmentversus a fundamental change of appearance?
(01:54:10):
Uh, anyway, uh, some of theinteresting philosophical
questions we'll have to deal with.
And with that, we are finishedwith this episode as promised.
Kind of a long run, we lots discussed.
So hope you enjoyed that.
Thank you as always for listening.
In particular, if you make itall the way to the end, we also
appreciate it if you share thepodcast, if you review it, and so on.
(01:54:34):
But more than anything, do keep doing it.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Special Summer Offer: Exclusively on Apple Podcasts, try our Dateline Premium subscription completely free for one month! With Dateline Premium, you get every episode ad-free plus exclusive bonus content.