Why AI Research Acceleration Just Doubled | Eli Lifland & Brendan Halstead — Transcript
Full transcript
- 0:00this uh recent meter report that came
- 0:02out last week um where they did an uh
- 0:05assessment of kind of like anthropics uh
- 0:08internal speed of AI progress and
- 0:11[music] then they sort of like released
- 0:13their findings and their best guesses
- 0:15and they found that uh over the course
- 0:18of the last year there's been on average
- 0:19about a 2x acceleration of AI AI [music]
- 0:22progress and um it seems like at first
- 0:25it was like mostly not due to AI
- 0:27speeding up AI research and mostly due
- 0:28to other But they estimated that of the
- 0:31current acceleration about 80% of it is
- 0:34due to um sort of like attributable to
- 0:36AI speeding up AI research.
- 0:38>> And [music] uh that was basically a big
- 0:42um update for us just because um for
- 0:44example in AI 2027 this sort of like
- 0:461.7x acceleration uh due to due to AI
- 0:50research doesn't happen until early
- 0:522027. So sort of just naively it seems
- 0:55like we might be ahead of the AI 2027
- 0:57timeline.
- 0:59All right, we are back. We are live with
- 1:02Eli Lifelin and Brendan Halstead who are
- 1:05researchers at the AI Futures Project,
- 1:07co-authors of 2040, and today co-authors
- 1:10of a new post on AI timelines, where the
- 1:13hook is AI timelines are much shorter
- 1:16than previously expected. And the uh
- 1:19members of the AI futures project now
- 1:21believe that we're going to get a
- 1:22superhuman AI researcher in 2027. So,
- 1:25Eli, Brendan, welcome to uh welcome back
- 1:27to MTS.
- 1:29>> Thanks for having
- 1:31>> So, for the audience, uh can you explain
- 1:34more about why you updated uh your
- 1:37timelines to much shorter?
- 1:40>> Yeah. So the main factor is this uh
- 1:44recent meter report that came out last
- 1:46week um where they did an uh assessment
- 1:49of kind of like anthropics uh internal
- 1:52speed of AI progress and then they sort
- 1:56of like released their findings and
- 1:57their best guesses and they found that
- 2:00uh over the course of the last year
- 2:02there's been on average about a 2x
- 2:03acceleration of AI AI progress and um it
- 2:07seems like at first it was like mostly
- 2:09not due to AI speeding up AI research
- 2:11and mostly due to other things but they
- 2:13estimated that of the current
- 2:14acceleration about 80% of it is due to
- 2:18um sort of like attributable to AI
- 2:20speeding up AI research
- 2:21>> and uh that was basically a big um
- 2:25update for us just because um for
- 2:27example in AI 2027 this sort of like
- 2:301.7x acceleration uh due to due to AI
- 2:33research doesn't happen until early 2027
- 2:36so sort of just naively it seems like we
- 2:38might be the head of the AI2020 27
- 2:40timeline. Um, and that's kind of the
- 2:43biggest sort of thing that led us to
- 2:44make this update. There's also been
- 2:46various other things like a few other
- 2:47pieces of evidence that Enthropic has
- 2:49shared and um, you know, AI is being
- 2:51super good at uh, seemingly you know uh,
- 2:54doing doing math solving uh, very hard
- 2:56math problems. But um, yeah, this kind
- 2:59of like meter report showing this um,
- 3:01acceleration is basically the the
- 3:03biggest um, input to our update.
- 3:06>> Why would math be a big update towards
- 3:07timelines?
- 3:09Um I think mainly just because uh you
- 3:13know we think capabilities across areas
- 3:15are are correlated and um I think seeing
- 3:19just how big the progress are in math um
- 3:21isn't math. It seems like it is some
- 3:23update about how uh how much progress
- 3:26you can make in the in some areas of AI
- 3:28research where there might be you know
- 3:30relatively uh verifiable um sort of like
- 3:33um targets that you want to optimize
- 3:35like maybe for example pre-training
- 3:37research. M
- 3:38>> sure. So you explored some hypotheses as
- 3:42to why the AI futures model was wrong.
- 3:44Um so some of these are near um
- 3:47automated coder skill hold on uh coding
- 3:52skills result in more a AI R&D uplift
- 3:54than we anticipated. Skills besides
- 3:56coding and research taste um are driving
- 3:58acceleration and conversion of uplift to
- 4:00pace of progress is too conservative. So
- 4:02what would be some of these skills
- 4:04besides coding and research taste?
- 4:07I think the biggest one that we could
- 4:09think of uh is just like AI's uh like
- 4:13generating RL environments. Um this is
- 4:15something that we didn't model in AI
- 4:17features model. Um but I think possibly
- 4:19like this is like of a different
- 4:21category than other kinds of coding
- 4:22automation. Um because it just directly
- 4:24contributes to the training and you can
- 4:26imagine there might be some relationship
- 4:27between like oral environments and uh
- 4:32points on this uh entropic ECI index
- 4:35that meter was measuring. this
- 4:36acceleration in
- 4:38>> um then it's possible that automating
- 4:39that area of coding specifically um
- 4:41would lead to impact uh greater than
- 4:44what we simulated in the feeders model
- 4:46>> right yeah I'm also like kind of curious
- 4:50what your take on like data as a
- 4:53bottleneck was just talking to our
- 4:55previous guest about this
- 4:58>> yeah I think uh depends on what kind of
- 5:00data you mean um I think that uh for
- 5:03pre-training it's quite hard to tell
- 5:04from the outside um But it certainly
- 5:06seems like pre-training gains have uh
- 5:08continued if not sped up um since the
- 5:11early days. Um so it doesn't seem like
- 5:13you know whether it's because of
- 5:14synthetic data or improvements in data
- 5:16efficiency in pre-training then it
- 5:17doesn't seem like there's been um really
- 5:19much of a problem there. Um and we don't
- 5:21really uh or I don't anyway expect one
- 5:25uh in [snorts] the near future. Um I
- 5:28think that for uh RL generally then uh
- 5:33yeah I think like AI can creating RL
- 5:36environments um is a big deal probably
- 5:38um I think that for certain domains
- 5:41where uh each rollout kind of
- 5:43fundamentally takes a long time you
- 5:44might not be able to do very many of
- 5:46them um and so you need a lot more data
- 5:48efficiency um however I think that like
- 5:50coming up with data efficient learning
- 5:52algorithms is the kind of thing that uh
- 5:54is itself uh verifiable And that's the
- 5:57kind of thing that you could get a lot
- 5:58of reps on and superhuman AI researchers
- 6:00could speed up a lot.
- 6:02>> But
- 6:03>> I want to shout out our sponsor,
- 6:05Lovable. You know the app you've been
- 6:06meaning to build, the internal tool,
- 6:08side project, or product you'd otherwise
- 6:10lose a weekend to. With Lovable, you get
- 6:13software you can ship today. [music]
- 6:14That includes an editable codebase you
- 6:16can inspect and change, plus two-way
- 6:18GitHub sync. Lovable also handles the
- 6:21backend and infrastructure including
- 6:22manage Postgress, O storage, hosting,
- 6:25payments, and more. So you don't have to
- 6:27wire it all up yourself. And through
- 6:29Lovable's MCP server, you can create and
- 6:32deploy your project directly from the
- 6:33agents you already use. [music] Turn
- 6:35ideas into software people love with
- 6:37Lovable. Lovable.dev. [music] Now on to
- 6:40the episode.
- 6:43Right before you guys, we were just
- 6:44talking to Romesh Nom, who is, you know,
- 6:46something of a short timeline skeptic,
- 6:48an RSI skeptic. Um, and I asked him
- 6:51like, you know, after you, we're going
- 6:52to talk to the AI futures project
- 6:53people, what do you think your cruxes
- 6:55are? Um, what should we talk to them
- 6:57about?
- 6:58And he said basically uh he thinks that
- 7:02uh the uh open AAI time horizon on a R&D
- 7:07tasks uh as talked about in their recent
- 7:09report is like much lower than the
- 7:11implied time horizon from like the ECI
- 7:14that we've seen for uh Anthropic's
- 7:17latest models on like ordinary meter
- 7:19software engineering tasks. So he thinks
- 7:22like there's something of a gap between
- 7:24models get good at software to models
- 7:27get good at a d. Uh and then he thinks
- 7:29there's another gap between like models
- 7:31get good at a r&d and they can automate
- 7:34all of a including the very hard parts.
- 7:37And then he thinks that there are
- 7:38diminishing returns from there to like
- 7:41full-on intelligence explosion. So like
- 7:43where do you think your disagreements
- 7:44with him are on that?
- 7:50Uh yeah. So for the uh I think there's
- 7:52there's several different parts. Um for
- 7:54the time horizon one um I think that uh
- 7:58makes sense that time horizon on real
- 8:00OpenAI tasks uh is lower than on meter
- 8:04tasks. There's a couple different
- 8:06factors for why. One would be uh the
- 8:08baseliners being lower context in meters
- 8:11uh eval. So it's like you have people
- 8:13who uh you know did not have high
- 8:16context on the project just like kind of
- 8:18contractors that were hired to do this
- 8:19and they might take much longer than
- 8:20someone uh would if it was like embedded
- 8:23in their larger job as it would have
- 8:24been with the openi tasks. Um and I
- 8:28guess like just in general like the
- 8:30absolute level of time horizon varies a
- 8:31lot across domains. Um the growth rates
- 8:33do not seem to measure as much. This is
- 8:35something that meter found like in their
- 8:36research from um last year I believe. Um
- 8:40and yeah like in the a model we actually
- 8:42account for this by increasing uh the
- 8:44time horizon required. We like add an
- 8:46additional multiplicative factor
- 8:48specifically because of this. Um so yeah
- 8:50I mean I think that for that particular
- 8:52thing then uh we have it accounted for.
- 8:55Um uh Eli any thoughts on other things?
- 8:59>> Yeah I mean I guess in general like um
- 9:03yeah I don't think I have a super
- 9:04creative understanding of of of where we
- 9:06might disagree. I mean I think in terms
- 9:08of like accounting for like diminishing
- 9:09returns like so our model generally you
- 9:12know does count for like ideas getting
- 9:14harder to find in this phenomenon in
- 9:15kind of like the way that economists um
- 9:18often like to model it. Um and it seemed
- 9:20to have actually predicted the past
- 9:22pretty well up until this like potential
- 9:24anthropic acceleration. So if anything,
- 9:27we're kind of worried that in our model
- 9:28the diminishing returns are not dim
- 9:31they're not strong enough basically like
- 9:33or sorry they're they're too strong in
- 9:34the sense that like you know it seems
- 9:36like idea you know if you would expected
- 9:37ideas to get harder to find at kind of
- 9:39like the same rate we would expect to
- 9:40kind of see a steady pace of progress
- 9:42>> but meter kind of says that there was a
- 9:44sort of like increase in progress
- 9:46despite there being at first mostly not
- 9:49sort of like attributed to AI
- 9:50acceleration and so like our model can't
- 9:51explain that so if anything I think kind
- 9:53of we've updated towards diminishing
- 9:54returns kind of being less steep than we
- 9:57thought. Um,
- 9:58>> yeah.
- 9:59>> Yeah. And then I guess there's I think
- 10:00maybe just real quick, yeah, one other
- 10:01thing you mentioned is kind of like, you
- 10:03know, what if like the hard parts of A
- 10:04and D are like much harder to automate
- 10:07um than the easier parts. We do try to
- 10:09automate sort of in the A features
- 10:11model. We kind of model this using like
- 10:12coding. Coding is like a taskbased model
- 10:14where there's like kind of like a bunch
- 10:16of different uh tasks and like there's a
- 10:18sort of like spectrum of how hard they
- 10:19are to automate. We didn't explicitly
- 10:21modif model model this in the sort of
- 10:23like research taste aspect of our model
- 10:24which is the other aspect of like you
- 10:26know when AIs will be able to surpass
- 10:27humans at at like you know selecting
- 10:29experiments research directions etc. Um
- 10:32but you know like in our in our most
- 10:34recent post we tried to sort of like um
- 10:36I think each of us kind of had our own
- 10:37adjustments that we made sort of to try
- 10:39to account for the fact that it is
- 10:41potentially more multi-dimensional. So
- 10:42something we've thought about but
- 10:43probably we do have a disagreement about
- 10:44the sort of extent of the update we
- 10:46should make.
- 10:48Um, how likely do you guys find it that
- 10:50even though like a 1.7x pace of progress
- 10:52is ahead of AI 2027 that there will be
- 10:55like some other factor that maybe blunts
- 10:59progress or something like that or do
- 11:01you think that this is just an update
- 11:02towards uh shorter timelines like across
- 11:05the board?
- 11:06>> Um,
- 11:08I think that seeing Oh yeah, you can go.
- 11:11>> Uh, yeah, you can add anything if you
- 11:12want but yeah, so I think
- 11:24is that you know like we're kind of
- 11:25deferring to the like meter authors a
- 11:27good amount on this. Um they did weren't
- 11:29able to like reveal a lot of their most
- 11:30important evidence. So like it's
- 11:31possible that there wasn't actually this
- 11:33acceleration in that case would be
- 11:34wrong. I think the other main thing that
- 11:36is sort of like potentially bearish
- 11:38update from this is if you think that
- 11:40you know maybe like if there was already
- 11:43significant AI attributable acceleration
- 11:45you know 6 months ago and now we have
- 11:481.7x but it hasn't gone kind of crazy um
- 11:51that is some update I think against like
- 11:53sort of like very fast takeoffs right
- 11:55because you might have expected that um
- 11:57kind of like in our model basically you
- 11:59conditional on this 1.7x acceleration
- 12:01recently and then you say you know how
- 12:03how much was acceler accelerating 6
- 12:05months ago, it usually says basically
- 12:06none. Um, so there's like potentially
- 12:08some update. The meter doesn't really
- 12:10give any details on exactly what what
- 12:12what the AI attribute acceleration was
- 12:14in the past. In fact, they basically
- 12:16don't say anything about it except a
- 12:17fake data graph. But um, but yeah,
- 12:19that's that's a potential up that like I
- 12:21think mostly it's a potential this whole
- 12:23thing is an update towards both shorter
- 12:24time and faster takeoff because the two
- 12:26are correlated. But there is like
- 12:29somewhat of an update towards longer
- 12:30takeoff from from this uh potential like
- 12:32more gradual acceleration than we
- 12:34thought.
- 12:35>> H can you explain that in more detail?
- 12:37There's somewhat of an update towards
- 12:38longer takeoff.
- 12:40>> Yeah. Um
- 12:42so essentially like um in order to get
- 12:47um sort of like a substantial AI
- 12:50attributable acceleration in the AI
- 12:52futures model at least you generally
- 12:55needed to get um significant gains from
- 12:57research taste. So significant so rather
- 13:00than just sort of like automating coding
- 13:02um generally you needed to get
- 13:04significant gains from like you know
- 13:06like being at least like you know around
- 13:08human level median at least in in
- 13:10research taste and and uh sometimes sign
- 13:13oftentimes significantly higher and so
- 13:16um yeah if if this is true uh if if this
- 13:21is sort of like an accurate model of the
- 13:22world that that you need like a
- 13:24significant research taste to get like
- 13:26significant acceleration at least sort
- 13:27of given our sort of like original
- 13:29guesses about the relative like
- 13:31abilities of coding and research taste
- 13:33then uh you might think well okay well
- 13:36six months ago we had to have had you
- 13:38know significant gains from research
- 13:39taste and it's been 6 months and yes we
- 13:41have 1.7x acceleration but it hasn't
- 13:43gone crazy or we haven't had this like
- 13:44crazy singularity up to like 10x
- 13:46acceleration or whatever it might be 5x
- 13:49acceleration something like that and so
- 13:52yeah so basically it's just an update
- 13:53that you know kind of um the research
- 13:56taste abilities of AIS might be
- 13:57improving slower than than we thought.
- 14:00And if so, that would um imply that take
- 14:04sort of takeoff would be slower than we
- 14:06thought because that's kind of the main
- 14:07thing at least in our model that
- 14:08determines that is basically how quickly
- 14:10as you sort of for example increase ECI
- 14:13or general capabilities, how quickly you
- 14:15increase research taste. And so yeah, I
- 14:17don't know if that makes sense, but
- 14:18that's basically kind of like the
- 14:20reasoning behind that. Of course there's
- 14:21like caveats like you know as you
- 14:23mentioned earlier like maybe the um we
- 14:26sort of talked about the hypothesis that
- 14:27actually you're getting a lot more
- 14:28automation from coding than we thought
- 14:30we would and maybe that's the
- 14:31explanation rather than sort of like the
- 14:33research taste improving much slower
- 14:34than we thought we would but um yeah
- 14:36basically that is one one the the sort
- 14:39of like relatively gradual acceleration
- 14:41is some update towards uh takeoff being
- 14:44slower because of sort of like key
- 14:45skills like research chase improving
- 14:47slower.
- 14:48>> Right.
- 14:49>> Right. So, if we do get to the end of
- 14:512027 or the end of 2028 and we clearly
- 14:54do not have like a superhuman AI
- 14:56researcher, what do you think the most
- 14:57likely reason for that would be?
- 15:02>> Um,
- 15:04yeah, I think that's a good question. I
- 15:06think I'm not sure I have like one sort
- 15:09of like overriding sort of like reason
- 15:11as a candidate. I think I I think
- 15:14probably I would say the most or like to
- 15:15extend accounts to the reason I think
- 15:17maybe probably the most likely is that
- 15:18like meters like sort of meters team and
- 15:21also anthropic who has now also by the
- 15:23way announced that like they think
- 15:25they've crossed like a 2x acceleration
- 15:27threshold um in their in their like near
- 15:29future models um uh basically if both of
- 15:32them are like overrating the amount of
- 15:35the gain that is happening andor
- 15:38overrating the amount that is retrieval
- 15:39to AI. I think those to me are like two
- 15:43of the sort of like biggest features
- 15:45just because you know they were you know
- 15:46they did they tried to do their best but
- 15:47they only had two people they were there
- 15:49for six weeks or something like that I
- 15:50think and you know people make mistakes
- 15:52um and and I think it's hard to yeah we
- 15:55we we don't have a great sort of science
- 15:57of of doing this sort of um sort of
- 15:59estimation um and then otherwise I think
- 16:02yeah maybe maybe something around um
- 16:06something around there being like lots
- 16:07of different skills involved in A&D and
- 16:10some of them are lagging um sort of like
- 16:12not improving as fast as others I think
- 16:14would be my main other hypothesis.
- 16:16>> Right.
- 16:17>> Right.
- 16:17>> Well, we're coming up on time, but it's
- 16:19been great having you on. Um love the
- 16:21work that you do. So, Eli Brendan,
- 16:23thanks so much for coming on MTS once.
- 16:26>> Thanks for having
- 16:27>> Thanks for having us.
- 16:29>> A huge thanks to MTS sponsors.
- 16:31>> 11 Labs, AI [music] that communicates at
- 16:33human level across every channel and
- 16:34modality. 11 ele.io/mts.
- 16:39Scale your startup on Neon. Millions of
- 16:41developers and startups have already
- 16:43chosen Neon for their backend. Start on
- 16:45the free plan or get up to $100,000 in
- 16:47credits for your startup at
- 16:48neon.com/mts.
- 16:52Special thanks to our sponsor, Kong, the
- 16:53AI connectivity platform. Connect APIs,
- 16:56LLMs, agents, and systems with serious
- 16:58security and governance. Konghq.com.
- 17:02Finally, prove what your AI investment
- 17:04is worth with Laridan. [music] Head to
- 17:06lariden.com today to start measuring
- 17:08your AI's bottom line. MongoDB, trusted
- 17:11to run critical applications by 75%
- 17:13[music]
- 17:14of the Fortune 100, the intelligent data
- 17:16platform for the era. Apply today for
- 17:18access to up to $15,000 in Atlas
- 17:20credits, [music] dedicated technical
- 17:22support, and exclusive events and go to
- 17:24market opportunities. mongodb.com
- 17:28ChargeBilling and monetization
- 17:30infrastructure for the AI [music] era.
- 17:32chargebbe.com
- 17:35ClickHouse is the real-time database
- 17:37built to keep up with every query your
- 17:38apps and agents send its way and it now
- 17:41runs Postgress observability and AI
- 17:43tooling on one platform. [music]
- 17:45Learn more at clickhouse.com.
- 17:48Blitzy autonomous software development
- 17:50for enterprise [music] codebases ship 5x
- 17:52faster blitzy.com.
- 17:54Adqu make your brand a billboard out
- 17:56ofome advertising as easy to scale as
- 17:58digital. adquick.com.
- 18:01Arena, measuring AI performance in
- 18:03[music] the real world. Arena.ai.
- 18:06Support for the show comes from VCX, the
- 18:09public ticker for private tech. [music]
- 18:10The US stock market started history's
- 18:12greatest wave of wealth creation. From
- 18:14factory workers in Detroit to farmers in
- 18:16Omaha, anyone could own a piece of the
- 18:18great American [music] companies. But
- 18:20today, our most innovative companies are
- 18:22staying private longer, which means
- 18:24everyday Americans are missing out.
- 18:26Until now. Introducing VCX, a public
- 18:29ticker for private tech. Visit
- 18:31getvcx.com [music] for more info. That's
- 18:34getvcx.com.
- 18:37Carefully consider the investment
- 18:38material before investing, including
- 18:39objectives, risk, charges, and expenses.
- 18:41This and other information can be found
- 18:42in the innovation fund perspectus at
- 18:43getvcx.com. This is a paid sponsorship.
About this transcript
This page contains the full transcript of Why AI Research Acceleration Just Doubled | Eli Lifland & Brendan Halstead by MTS , generated from the public captions YouTube serves with the video. The transcript has 3,476 words across 516 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.
What you can do with it
Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.
Free YouTube transcript tool
YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.