
What this covers
This is a 28-minute presentation from Bridgewater's AIA Labs, recorded at the Bridgewater-GIC anniversary event in Singapore. The speaker (named in the description as Jas Sekhon, though the transcript labels him unidentified) argues from the history of connectionist methods, which take raw data and use an algorithm to predict outputs, with more compute and data reliably improving results. He walks through benchmarks that fell faster than experts expected: handwriting recognition took about twelve years, image classification arrived years ahead of Stanford's generational estimate, and competition math went from 6% to above 94% of human level using scale alone.
The middle of the talk is a run of concrete results. Large language models (LLMs)) are reported scoring 83% on a neurosurgery oral board exam and in the 88th percentile on the LSAT. AI systems reached silver-medal level overall at the International Mathematical Olympiad, with gold-standard performance on geometry. He cites Terry Tao's "mediocre grad student" remark, and an antibiotic search that screened 30 billion molecules and found six candidates that killed drug-resistant bacteria. He frames the scientific gains as tasks humans could not do at all, rather than replacements for existing work. He also covers the cost side: compute scaling at roughly 4–5x a year, and data centers moving from hundreds of millions of dollars toward reports of $100 billion projects. He claims the San Francisco Bay area matters more to AI innovation than Europe and Asia combined.
The contentious turn is the gap between science and business. Scientists can work around the limits of LLMs, which have high knowledge but weak step-by-step reasoning and so cannot act as autonomous agents. Businesses lack that expertise, and capital allocators are disappointed by demand. He describes two bets by OpenAI, Anthropic and other labs: more scale, and built-in algorithmic reasoning, which draws on symbolic methods. Two outcomes follow. In one, agentic planning arrives and deployment is easy. In the other, adoption is a long grind like electricity, which took 25 years to pay off because factories had to be reorganized.
AI progress appears paradoxical because scientific applications are achieving breakthroughs while business adoption lags, but this resolves when understanding that current language models lack planning ability—either labs will solve this through algorithmic reasoning to enable rapid economic transformation, or adoption will follow the slow historical pattern of general-purpose technologies like electricity.
- Connectionist methods have achieved above-human performance on language, math olympiads, and scientific discovery while traditional ML extracts most current economic value from planetary-scale systems
- Large language models are knowledge-rich but not intelligent at step-by-step reasoning and cannot function as autonomous agents without expert-directed scaffolding
- Two paths forward: scaling + algorithmic reasoning (bullish, near-term transformation) vs. gradual deployment like historical GPTs (slow, 25+ year institutional rewiring)
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
A group at McMaster University and Stanford collaborated to use AI to screen 30 billion possible molecules for novel antibiotics, which humans could never do before, generating 56 candidate molecules of which 6 were shown in lab testing to kill drug-resistant bacteria, achieving a hit rate far above human drug discovery rates.
“a group at McMaster which is a Canadian University at Stanford got together and said hey mant why do we get AI to go through 30 billion possible molecules unimaginable for humans to do this right and see if it can do something so if you think about the Innovation effect of AI it's not like replacing a task for humans could do humans have no way of doing this before and you're like okay we're going to do something humans cannot do and see if generative AI can do it and and it's they don't have Google Deep Mind Capital they just have a lot of expertise so they did it right and they they created a system it goes through 30 billion possible molecules it proposes 56 they take it to a lab six of them work and actually kills the drug resistant bacteria right and a hit rate of six wins out of 56 proposals is way above human drug Discovery rates”
GPT-4 scored 83% on the neurosurgery oral board exam despite this exam not being in the training data, where neurosurgery is the longest-trained medical specialty in the US, requiring undergraduate degree, medical degree, residency, and surgical subspecialty fellowship, with no human scoring above 90% in a decade, and this exam tests differential diagnosis and reasoning rather than memorization.
“the most impressive test result here is the neurosurgery board exam because this was not in the training data so what are the neurosurgery oral board exams first neurosurgery is the medical specialty in the United States that has the longest training time so these are people who have had a undergrad a medical degree a residency a Surgical subsp Specialty fellowship and this is their last exam and it scored 83% no human scored about 90 in a decade and this is a high pass and when this paper came out that showed this it was in a paper from one of the labs it's in a meeting with the Deans of Harvard and yo med school and they just turned white they're like medicine's never going to be the same because this is not a memorization task this a test to differential diagnosis and reasoning”
In 2021, AI researchers concluded that solving middle and early high school competition math problems would require new science because AI could only score 6% of human-level performance, but by early 2024 these systems score above 94% of human-level performance on the same benchmark using only scale without breakthrough science.
“the meth this math benchmark is on competition level math questions for middle and early high school so not the most difficult things that we can imagine but they're above basic arithmetic they're like algebra and geometry in 2001 AI can only score 6% of human level performance and the authors of the who put out the Benchmark former colleagues of mine to Berkeley they had a very disappointing conclusion they said well I think to do well in this Benchmark we're going to need new science and the reason that's disappointing is you can't Time new science you don't know when Von's going to come around you don't know when Einstein's going to have a 1905 year and publish a whole bunch of transformative papers new science you can't time you can time Technology and Engineering right but but the beginning of this year on this Benchmark that we thought we need new signs for you're above 94% of human level performance on this Benchmark so and all it took was just scale just give more compute more data and just use the usual tools of the connectionist method”
Speech recognition took longer than handwriting recognition to reach human-level performance, and then image classification (cat vs. dog) arrived as a benchmark that Stanford researchers thought would take a generation to solve but was solved in only a few years.
“then came speech recognition right which took us even longer and then things started changing quite rapidly the thing was F Fay at Stanford put out a benchmark saying image classification tell me is it an image of a cat or a dog and they thought this would take a generation to get to human L performance it only took a few years”
Connectionist methods work by taking raw data (words, pixels, or macroeconomic statistics) and using an algorithm to predict desired outputs, where more compute and data reliably lead to better performance, and sufficiently scaled induction produces intelligence.
“they start with a very simple idea they say give me some data doesn't matter what that data is it could be a bunch of words it could be pixels from an image it could be macroeconomic statistics and just tell me what you want to predict you could want to predict the next word you could want to predict the next macroeconomic print or is it a picture of a cat or dog and in the middle the algorithm will do its thing it it could be a black box it might be hard to interrogate and it will and you'll make a prediction and if you give me more compute and more data I'll get better and better at it and if I keep doing this enough I'll get intelligence”
Foundation labs (Google, Anthropic, OpenAI, Mola, Meta) know that LLMs lack effective tool-use and agentic behavior, and they are pursuing two strategies to solve this: scaling (giving models more data and compute, which has been shown to work), and algorithmic reasoning (building in step-by-step reasoning inspired by symbolic AI methods).
“all the foundation Labs know this Google knows this anthropic knows this open AI knows this mol knows this Mena knows they all know this so they're like saying hey we're going to try to make these tools easier for people to use and there's two ways they can do this the first is just scale right the thing about this revolution is the stupidest thing works what's the stupidest thing I'm going to give you more data and more compute now it turns out not it's not trivial to do that but it's not like you need Albert Einstein you can just do it right so we're going to do this and much like chat gpt2 couldn't count to 10 and chat G pt4 can it's not like we invented a new algorithm we just gave it more data and more compute so we'll just keep doing that right so that's one bet the second bet is no no okay now we're going to actually try to be clever and we actually have to have an idea and our idea is we're going to build in reasoning kind of like from that symbolic school from AI that we kind of forgot maybe we shouldn't have forgot about it and we got to put in this step-by-step reasoning”
Innovation in AI is happening up and down the supply chain, almost entirely within the United States, and the San Francisco Bay area is more important for AI innovation than all of Europe and Asia combined.
“we're getting better at the algorithms so there's been Innovation up and down the supply chain and this Innovation is almost entirely happening in the United States like San Francisco barrier by itself is more important than the rest of the world out outside combined in this field right the people just don't fully get that right if all of Europe and Asia disappeared it's less important than the San Francisco Bay area for the innovation in this space”
In the year 2000, AI handwriting recognition was at 0% performance and achieved human-level performance in about 12 years, which was roughly one academic generation and received little cultural notice outside the US Postal Service.
“so the beginning of the century when the start of my scientific career things were like pretty normal so you wanted to do handwriting recognition right so in the year 2000 you're like 0% performance and after 12 years you've reached human level which is like in an academic generation and no one really noticed other than the US Postal Service”
Within the scientific community, AI progress in the past 12 months has exceeded expectations, with scientific problems previously considered science fiction 24 months ago now being solved across diverse fields including climate science and structural chemistry.
“at the same time within the scientific Community there's above expectation progress in Ai and the expectations are already very high and this isn't like people in computer science or or AI itself talking their book this is from scientific Fields as diverse as climate science to structural chemistry problems are being solved that were considered science fiction uh 24 months ago”
Electricity adoption initially created negative productivity when factory owners replaced steam dynamos with electric motors because the steam dynamo had more force than the electric one, requiring factory operators to discover that electric motors could be distributed in parallel across the factory floor and building, a realization that took 25 years.
“Factory owner said hey man I have electricity I'm going to replace my Steam Dynamo with an electric one cuz that's the thing I'm supposed to do they did it negative productivity head why because the steam Dyno Dynamo literally did more work because it had more Force than the electric one you got so actually all your machines didn't work well so it was a negative productivity hit and they're like man I spent all this money and my productivity's gone down and then someone got the bright idea but hey unlike the steam one where there's one big one the electric ones I can like hook them up in in parallel and I can move around my factory floor in fact they don't even have to be in the same building and I can get a lot more work done then okay that took 25 years to do”
On the International Mathematical Olympiad, GPT-4 scored 0% (could not solve a single problem), matching the median human performance; but by late 2024, AI systems achieve 100% on Euclidean geometry problems (gold medal standard) and silver medal performance overall (approximately 83% on geometry), progressing from 0% to gold in geometry in six months.
“by this year you have gold medal in the geometry portion of it 83% earlier the latest mod gets 100% of the geometry questions right right so now we can do ukian proofs without fail when no AI system on Earth did it six months ago got zero go from zero to 100 in six months and now it's in a silver level across all of the intellectual math Olympiad”
Compute given to AI algorithms has been scaling at approximately 4x per year, recently accelerated to 5x per year, yielding approximately 10x improvement in compute every 2 years.
“the amount of compute given to these algorithm is going up by about a factor of four a year uh it's sped up recently was closer a factor five so you're getting an order and magnitude Improvement 10x on compute every two years”
GPT-4 scored 88th percentile on the Law School Admissions Test, which is the test required for college graduates to apply to law school, placing it above the average score of almost all American law schools.
“you have the law school admissions test that is taken for people with the post-secondary with a college degree to go to law school it scores 88 percentile which is above the average score of almost all American law schools”
GPT-4 training cost approximately $400 million for data center construction (distinct from the lower-profile $100 million operational training cost), but companies are already building $10 billion data centers and there have been reports of a $100 billion 'Project Stargate' between Microsoft and OpenAI.
“chbt 4 right got trained in 2022 it got released 2023 it was trained 2022 the data center cost to build the data center to train chat GPT for was $400 million if you Google it the number $100 million usually shows up that's kind of the leaste cost of it for trade not the build cost the data center it's $400 million all right that sounds like a lot but we're already building 10 billion data centers there's already been leaks of say project Stargate between Microsoft and op and AI for hundred billion dollar data centers”
OpenAI's latest model, released last week, represents the algorithmic reasoning bet and does not have more scale than their previous model but instead adds step-by-step reasoning capabilities.
“the new model from open AI that just dropped like last week is about the second bet of algorithmic reasoning it doesn't have any more scale than their previous model they're working on that too it just has reasoning in it”
Terry Tao, arguably the greatest mathematician of the last 50 years, praised AI systems as performing at the level of a 'mediocre grad student' in mathematics—high praise coming from him, since a year prior this comparison was 'not even close'
“that's why you have Terry to arguably the greatest mathematici lie probably the greatest mathemati last 50 years looks at these models and says hey they're now getting as good to be uh almost a mediocre grad student right in math but that's High Praise You have an a systems a meteor grass math and you're also the person saying it is arguably the greatest mathematician in like half a century so that you know a year ago was like not even close”
Large language models have high knowledge (understanding both thermodynamics and Shakespearean characters like Rosencrantz and Guildenstern) but low intelligence at step-by-step logical reasoning, cannot solve extended logical puzzles, and therefore cannot function as autonomous agents.
“they have difficulty thinking step by step they have a lot of knowledge so they're very smart in the sense that they have a lot of knowledge these systems know a lot about the laws of thermodynamics and they also know a lot about Rosa CR and gilded Stern two characters in Hamlet the number of human beings who know both of those sets of pieces of information is exceedingly rare they have those so they're very smart in that sense but they're not very intelligent of the sense of saying I'm going to give you a logical puzzle you got to go think about it step by step and you got to remember what you're doing they fall apart”
International math olympiad problems are so difficult that if you gave the modal human a billion dollars in training resources, they would be no more likely to win a gold or silver medal than to win a gold or silver in the 100-meter Olympic sprint.
“these tests are like so hard if you took the modal human and you gave all the training you want want spend a billion dollars on the modal human if you want they are not going to get a gold or silver uh or Bronze in the international math olympiads they're as likely to do that as they are to win a gold or silver in the 100 meter Olympics”
Capital allocators like SoftBank and Goldman Sachs are disappointed in AI progress over the last 12 months because market demand and use of AI products is not matching capital expenditure expectations.
“certain Capital allocators are somewhat disappointed about AI progress in the last 12 months so you have places like seoa and Goldman Sachs writing where is the demand for all this capex expenditure going there's not that much Market uh uh use or interest for the AI products currently being generated”
The innovation effect of AI in scientific domains involves tasks humans cannot perform at all (like screening 30 billion molecules), not replacing tasks humans currently do—representing a genuinely new capability frontier
“if you think about the Innovation effect of AI it's not like replacing a task for humans could do humans have no way of doing this before and you're like okay we're going to do something humans cannot do and see if generative AI can do it”
The bullish case for AI is that labs will make LLMs agentic and capable of planning even without increased intelligence or knowledge, which would enable companies to deploy AI without reorganization, and productivity improvements would flow easily similar to managing external teams.
“the first one the bullish case will be will all the labs are working on it the entire scientific community that works on these FS working on it we'll just get large language models to be agentic and to be able to plan even if they got no smarter than they currently are right a they have above human level knowledge on a whole bunch of things that as I just showed you but if they could just plan this would be great we could add a lot of economic value and then companies wouldn't have to reorganize themselves you know much like you have to manage an external team in Spain or or India you would have some management task but it's not like your workers would have to be different it's not like you have to go hire an army of phds etc etc the technology would flow pretty easily and products would get get induced so that's the bull case we'll get there”
The vision 12 months prior was that adding tools (calculators, programming interfaces, other resources) to large language models would enable them to solve business problems, but this failed because LLMs cannot maintain sufficient attention to use tools effectively and solving this requires either significant scientific expertise and capital investment or simpler technology.
“the thought about 12 months ago in Silicon value was okay we know this but what's going to happen is we'll take the large language models we'll give them a lot of tools they actually can't do math which is not what anyone would have predicted but they can't but we have calculators we'll get them a tool of programming and we'll give them like all these other tools and then pretty soon they'll do back office stuff for me they'll write complicated code and maybe they'll solve a lot of my business problems the problem came that they don't their attention span is so limited kind of like a goldfish they can't even use the tools effectively right and in and if you want them to try to use the tools effectively you need a lot of scientific expertise and use the significant amount of capital to solve your problem but who has that mix if you're Google you can do this and if you're sign scientific lab and your problem is protein folding or your problem is I want to find a new antibiotic you can do this but if you're Ford or you're basically anybody in the economy you can't do this you're not set up for it”
The scientific community is excited about AI because scientists have sufficient expertise to work around current technical limitations and control these difficult-to-use models, whereas the business community is uncertain about value extraction because they lack the expertise to maximize current technology.
“why the scientific Community is very excited even the sci community community that doesn't directly do AI because they're like they have the expertise to do all that work to control these models that are very hard to control right and why also then the business Community is like well I'm not sure how to I use it I'm not sure I'm how to get the value out of it right now because the expertise is not there to make maximum use of the current technology”
The less bullish but still transformative case is that making AI technology flow through the economy will be a long grind comparable to general-purpose technological innovations in history, which were slow due to required institutional changes, using electricity adoption which took 25 years to become productive as the canonical example.
“see that doesn't happen then you have the less bullish case but that's still very transformative which is no to get this technology to flow through it's going to be a long grind like every other general purpose technological innovation human history has been a slow grind and it's going to be like electricity electricity came so Factory owner said hey man I have electricity I'm going to replace my Steam Dynamo with an electric one cuz that's the thing I'm supposed to do they did it negative productivity head why because the steam Dyno Dynamo literally did more work because it had more Force than the electric one you got so actually all your machines didn't work well so it was a negative productivity hit”
Bill Gates previously believed it would take a generation to achieve 85th percentile performance on the AP Biology exam; OpenAI accomplished this in six months
“Bill Gates thought would take a generation to get and it didn't they took open AI six months of work to do it”
Language models are culturally significant because humans believed language ability was unique to humans—machines now perform above human-level on language tasks, challenging the assumption that language is what makes humans uniquely intelligent.
“the thing that's culturally capture the most attention is language and I think culturally it's captur the most attention because we're used to like like other entities on the planet running faster than us being stronger than us but until these language models came out there was nothing on this planet that could speak language like we could we thought that was the thing that made us uniquely human well it turns out machines can now do it and they can do it above human level performance”
Different labs have different incentive structures: Google DeepMind prioritizes scientific applications (e.g., protein folding) and gave the results away to transform medicine rather than monetize them, while OpenAI and Anthropic must produce consumer and business models due to their business structure.
“the different players have different incentives in mind for scientific applications by far Google Deep Mind is ahead that is just what its passion is it's the one that solve What's called the protein folding problem and then just gave the information away to the world because their ideas to transform medicine not to monetize the the protein folds themselves but they haven't produced their best models for consumers or businesses while open Ai and anthropic because of the way the business structure need to do that so they do produce the best models uh for people at the end”
Almost all economic value from machine learning has come from traditional machine learning systems created by American tech giants (Amazon, Google, Facebook, Nvidia) rather than from generative AI, operating at planetary scale to optimize supply chains, ad targeting, and chip design.
“almost all of the economic value for machine learning has come not at all from generative AI it's come from just traditional machine learning I'm going to predict things and it's mostly come from planetary scale machine Learning Systems created largely American Tech Giants so what you're Amazon you're using machine learning to create a Global Supply system right and you're going to optimize and estimate demand and Supply curves from your suppliers to Consumers so you can match prices right your Google and I am going to create machine learning techniques and models to Target ads to you and I'm Facebook and I'm going to do the same thing right so on and so forth and this is a giant these planetary scale systems which you've never had before or work around the world and they take that value and they bring it back to the United States in general and the San Francisco Bay area in particular”
Information about algorithmic innovations is being 'locked down' and papers are no longer published, even though information does flow through personal connections and company transitions; the National Security apparatus has 'woken up' and is requiring labs to restrict algorithmic innovation disclosure.
“the information's getting to be locked down and in this chart if you look at when you see the compute going up how much is from the hardware side and how much is the algorithm algorithmic progress it's still the hardware side so I think of Nvidia right that's the compute scale but that light but that blue is growing that's the algorithmic progress that's like no we're actually going to like get better ideas instead of just compute and now what's happened in the last few years that's now gone silent you're not publishing those papers now it's not like the scientists working on it are sitting there Los Alamos in the desert in Nevada and not like communicating with the Germans because you know you bump into them people talk people move around in companies so so the information is not super secret but it is somewhat secret no one's writing the papers and it's pretty quiet”
Foundation labs have been leapfrogging each other in releasing the best models, with Anthropic's Claude briefly outperforming OpenAI's GPT-4 before OpenAI released a new model, and these leapfrogging dynamics make it difficult to predict which lab will win because the biggest players are committing sufficient capital that compute allocation becomes the primary variable.
“that if you look at you know when a model's released and their capabilities they keep leapfrogging each other right until this latest model from openi got released last week the the best model was from thropic Claud it was better than the previous model from gp4 and then you know for brief time Google spun up there but they keep Le Fring back of course it's very hard to predict who's going to win here right because the biggest players are putting because they consider an existential risk to themselves are putting the capital to do it so much of this is about just the compute allocation you have so it becomes very hard to predict who's going to win now”
Early reports from OpenAI's latest model indicate it is much better at reasoning than previous models, suggesting progress toward the bullish case of solving the planning/agentic problem, but significant work remains before this is fully solved.
“early reports from this open AI model is that it's much better at reasoning than previous models so it is like a step trying to solve towards the bullish case which they're trying to work on but it's still a long way to go and it's very hard to predict which of those two will win”
Connectionist methods (based on induction) have dominated AI research publications overwhelmingly compared to symbolic methods (based on deduction), with symbolic methods only briefly dominating between the two schools of thought.
“you see a quick history of AI and the yin and yang between two different schools of thought about AI the first one in Orange is consider it's what's called connectionist Methods and it's basically based on the idea of induction the second one in blue is symbolic methods and symbolic methods are based on the idea of deduction right and then on the y axis here is the proportional log scale of the scientific Publications from one school versus the other as you can see other than for a brief time in between where symbolic methods dominated it is all connectionist methods now”
Symbolic methods require writing down formal logical systems and deducing from them, exemplified by calculators which cannot make arithmetic mistakes because humans explicitly programmed the rules of arithmetic into them.
“the symbolic method School says well there's life than just induction there's also deduction so what you Al what you got to do is write down formal logical systems and then and deduct from them right so a canonical example of this is your calculator your calculator cannot make a a mathematical mistake the reason is it knows the rules of basic arithmetic now it didn't learn those rules someone wrote down those rules a human being and provided them to a calculator and unless there's a hardware failure it cannot make a mistake”
Whether returns to scale will continue to improve AI performance as scale increases is unknown and will only be determined through trial, but capital is flowing into this bet and scaling will continue.
“now are we going to continue getting the returns to scale because as scale's gone up performance keeps getting better we don't know until we try but the capital is so far going in uh and so scale is going to continue”
The base case for AI adoption (slow, 25+ year timeline) is 'still really shockingly optimistic' and 'fast' compared to the complete failure case, representing what will happen unless the agentic planning problem is solved.
“and that's what most technological innovation's gone through human history that's going to be the base case unless we fix this agentic problem right and the base case is still really shockingly optimistic right and fast”