What this covers

This is a conversation between Patrick and Gavin Uberti, the 21-year-old Harvard dropout behind the chip startup Etched. It begins with the basics: what a transformer is, how pretraining on next-word prediction forces a model to store concepts, and how a smaller second stage of RLHF turns that base into an assistant. It then moves to hardware. Uberti explains why inference is memory-bound, why batching users amortizes the cost of loading weights (about 2,500 users on a transformer ASIC versus about 64 on GPUs), and why the time to first token is slow when a chat's history is recomputed each turn.

The range is wide. On the modeling side, Uberti invokes The Bitter Lesson, predicts that text data will run out and that video and self-play will take over, and weighs tree search and the rumors around Q*. He also argues that transformers will persist because of ecosystem lock-in. Robotics comes up as a place where response time and scarce training data hold things back.

On the silicon side, he covers memory costing more than the logic on chips like the H100, a few suppliers sitting at critical nodes, and chip design cycles of four to five years. He also describes automated layout on older nodes and hobbyist tape-outs on shared photomasks. Language models are weak at Verilog because hardware code is proprietary. The contested claims include a gradual rather than sudden path to superintelligence, most firms fine-tuning rather than training from scratch, and ever-larger data centers, where he expects hardware failures to become a central obstacle.

Sharpest takeaway

Uberti argues AI progress is driven by compute and scale, so purpose-built transformer chips and a few mega data centers will capture the value, with latency rather than raw intelligence as the binding bottleneck.

  • Each GPT generation requires exponentially more compute, so superintelligence arrives as a gradual spectrum rather than a fast takeoff
  • Burning the transformer architecture into silicon, as Bitcoin ASICs did, can yield order-of-magnitude gains in compute and latency
  • Semiconductor and AI economics consolidate into a few firms that can afford $10B-scale training and fabs, with fine-tuning on top

The claims · ranked72 claims · weighted by value

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.

0.84

On chips like the H100 or AMD MI300X, memory costs more than the silicon itself by more than a factor of two, and memory is made by only two or three firms, Samsung, SK Hynix, and possibly Micron, making them critical supply chain nodes.

factualhigh valueestablishednovelty 4/4durability 3/4· Gavin

“the memory actually costs more than the silicon itself by more than a factor of two.”

0.81

AI compute will centralize into a few very large facilities because bandwidth between adjacent GPU racks is about double that of an entire undersea submarine cable, creating strong incentives to co-locate.

causalhigh valueestablishednovelty 3/4durability 3/4· Gavin

“the GPU racks have double the bandwidth of a literal undersea cable.”

0.80

Pretraining on next-word prediction across trillions of tokens forces a model to store many concepts, and a second stage of RLHF on helpful, honest examples turns that base into an assistant, with the second stage using far less data and compute.

causalhigh valueestablishednovelty 2/4durability 4/4· Gavin

“you have to have a huge number of concepts stored within that machine.”

0.80

Rich Sutton's Bitter Lesson holds that hard-coded human knowledge yields short-term gains but never scales, while methods that leverage cheaper compute and large data, like Deep Blue search and large neural networks, are the only techniques that routinely win.

causalhigh valueestablishednovelty 2/4durability 4/4· Gavin

“the only techniques that have routinely worked are those that levered the fact that compute is getting cheaper.”

0.79

The computational effort to lay out a chip grows superlinearly with transistor count, so going from roughly 100,000 transistors to over 80 billion on modern GPUs is more than 100,000 times harder, requiring a data center rather than one server.

factualhigh valueestablishednovelty 3/4durability 4/4· Gavin

“is super linear in the number of transistors”

0.79

Unlike classical models, where more parameters overfit and generalize worse, transformers do not fall into this trap as much, so adding more money and compute keeps making them better.

causalhigh valuecontestednovelty 4/4durability 3/4· Gavin

“They are, to quote one of my teachers in college, the first capitalist AIs. Adding more money makes them better.”

0.78

The first token is slow because the model must reprocess the entire prompt, and for multi-turn chats the whole message history is recomputed, since storing the key-value state in RAM between turns is more expensive than recomputing it.

causalhigh valueestablishednovelty 3/4durability 3/4· Gavin

“It's usually cheaper just to recompute it.”

0.78

Inference is memory-bound because each parameter must be loaded from memory and used once, so batching many users to reuse each loaded weight amortizes the expensive load, and a transformer ASIC can batch around 2,500 users versus about 64 on GPUs.

causalhigh valueestablishednovelty 3/4durability 3/4· Gavin

“load that weight in once, pay the expensive price and then amortize that expensive price over a huge number of users”

0.78

Automated open-source back-end tools can complete layout for older-node chips in a day instead of nine months, and extending this to modern chips is possible in theory but has not yet been achieved.

factualhigh valueestablishednovelty 3/4durability 3/4· Gavin

“you can have the back end done automatically in a day instead of nine months by using one of these automated tools.”

0.76

Pretraining a model to predict the next word across trillions of examples produces a model that is only moderately accurate (around 18 percent) but must store a large number of concepts internally, which is why it can predict next words in science papers or web text.

causalhigh valueestablishednovelty 2/4durability 4/4· Gavin Uberti

“Maybe it gets right 18% of the time, but to even get it right 18% of the time, you have to have a huge number of concepts stored within that machine.”

0.76

Hobbyists can now design and tape out chips on older nodes in weeks or months by combining many projects on one shared photomask set, a model Uberti thinks can scale up to companies building new-node chips.

factualhigh valueestablishednovelty 3/4durability 2/4· Gavin

“They put them together on one mask set. You make one photomask that has all the chips on it.”

0.75

Robots need very fast response times because a one-second round trip to a server is too slow for physical control, so latency is critical, and this is why transformer robotics has not reached the mainstream despite Tesla using vision transformers.

causalhigh valueestablishednovelty 2/4durability 3/4· Gavin

“A second is too long if you're about to follow. So latency is so, so critical”

0.74

At 10,000 GPUs one failure is occasional and recoverable from checkpoints, but at 10 million chips a failure occurs every couple of hours, and no one has yet solved how to handle failures without stalling the whole training run.

forecasthigh valueestablishednovelty 3/4durability 3/4· Gavin

“You have 10 million GPUs or 10 million transformer ASICs. Then you're going to have one failing every couple of hours.”

0.73

Proprietary data like Bloomberg's can be learned later by fine-tuning a general foundation model, because pretraining is the expensive and domain-agnostic step, so most companies will fine-tune rather than train from scratch.

forecasthigh valuecontestednovelty 3/4durability 3/4· Gavin

“I think this data is very, very valuable, but it can be learned later.”

0.73

Superintelligence will not arrive overnight as a jump from human-level to 100x smarter; it is a gradual spectrum because each generation requires exponentially more compute, and progress only looks obvious in retrospect.

normativehigh valuecontestednovelty 3/4durability 3/4· Gavin

“super intelligence is much more of a spectrum than it is a binary thing. And eventually, we'll look and we'll say, wow, GPT-12 is obviously super intelligent.”

0.73

Text data may run out, but video is the most valuable untapped source, and transformers should be trained from day one on large video datasets rather than adding images as a side feature to text-trained models.

forecasthigh valuecontestednovelty 3/4durability 3/4· Gavin

“we haven't even begun to tap into the most valuable source of data, the video.”

0.73

Language models are strong at C and Python because abundant open-source code exists online, but they are weak at hardware description languages like Verilog because hardware code is proprietary, and opening or licensing old chip designs could fix this.

causalhigh valuecontestednovelty 3/4durability 3/4· Gavin

“While software loves open source, hardware really doesn't.”

0.73

Transformers will persist because the ecosystem favors them: NVIDIA is adding hardware support for transformer memory access patterns, and transformer libraries are so mature that any replacement architecture would need to be dramatically better to overcome the advantage.

causalhigh valuecontestednovelty 3/4durability 3/4· Gavin

“So for a new architecture to go and beat the transformer, it has to be so incredibly good”

0.71

The biggest driver of AI progress is scale, moving from GPT-2 at 1.5 billion parameters to GPT-3 at 175 billion to multitrillion-parameter GPT-4, with companies planning roughly 10x more training spend for next-generation models.

factualhigh valueestablishednovelty 2/4durability 3/4· Gavin

“the biggest, by far, in my mind, is scale.”

0.71

Robotics AI lags because robots cannot afford to fail a million times to learn tasks like walking, and simulations fail to transfer well to the real world, so training data is scarce compared with image datasets.

causalhigh valueestablishednovelty 2/4durability 3/4· Gavin

“For robotics, you can't have it break the robot a 1 million times before it figures out how to walk.”

0.70

Taking a chip from conception to mass production currently takes four to five years, which is too slow for training new models, so the semiconductor design process itself must be redesigned.

factualhigh valueestablishednovelty 2/4durability 2/4· Gavin

“Right now, going from conception to a mass production of the microchip can take four or five years”

0.70

Language models may move toward tree search, generating many candidate continuations, evaluating them, and backtracking to the best path, as chess engines do, though this has not caught on yet.

forecasthigh valuecontestednovelty 3/4durability 3/4· Gavin

“the model makes a couple of hundred plans of, hey, here are the next 100 things I could say.”

0.70

When human-generated data runs out, models that generate their own training data through self-play and self-critique, as AlphaGo did, will keep scaling, and this technique is already showing promise in language.

forecasthigh valuecontestednovelty 3/4durability 3/4· Gavin

“I think that self-play and generating their own training data is the final frontier here.”

0.69

Language models acquire concepts like ethics by absorbing human-written text, so they can judge which of two responses is more ethical, and that learned concept can be used as the objective for self-improvement.

causalhigh valuecontestednovelty 3/4durability 2/4· Gavin

“So the AI knows what is good, what is ethical. And it got the understanding by reading all the texts we've generated as a species.”

0.68

Bitcoin mining moved from CPUs to GPUs to ASICs, and once a specialized chip existed GPUs became unprofitable for mining; the same pattern will hold for transformers, where specialized ASICs will displace GPUs for inference.

causalhigh valuecontestednovelty 2/4durability 3/4· Gavin

“I think the same exact thing is going to happen for transformers.”

0.68

Earlier general-purpose AI chip startups such as Graphcore, SambaNova, Cerebras, and Groq lost to GPUs because their software was immature compared with CUDA and their chips were not meaningfully better in memory bandwidth or FLOPS.

causalhigh valuecontestednovelty 2/4durability 3/4· Gavin

“And I think this is two reasons. First, the software. The second the chips generally weren't better.”

0.68

The transformer is not a fad that will be replaced but a foundational building block on which retrieval augmented generation, memory caches, and agent systems will be built.

forecasthigh valuecontestednovelty 2/4durability 3/4· Gavin

“the transformer is not just the next fad. This is a building block.”

0.66

Although transformers generate one token at a time, there is strong evidence they plan ahead internally, similar to humans who form internal plans before speaking, and pretraining helps develop this planning skill.

factualhigh valuecontestednovelty 3/4durability 3/4· Gavin

“there is very strong evidence that there is a little bit of planning going on”

0.63

Export bans on state-of-the-art AI chips are in the national interest because China lags in lithography after missing the EUV transition, and the US must not be passive about maintaining its lead in data centers and chips.

normativehigh valuecontestednovelty 2/4durability 2/4· Gavin

“So I think the export bans for state-of-the-art AI chips are really in the national interest here.”

0.63

Current beam search in language models is primitive, typically using at most four beams with shallow lookahead, and it requires roughly 4x more compute, which is a large overhead, so it has not caught on.

factualhigh valueestablishednovelty 2/4durability 2/4· Gavin

“It usually has, at most, four beams. They don't look very far ahead.”

0.57

As state-of-the-art chips became more expensive, only two or three firms such as TSMC, Samsung, and Intel can make leading-edge chips, and AI model development will similarly consolidate to a few mega-scale developers, with a tail of niche providers.

forecasthigh valuespeaker onlynovelty 3/4durability 3/4· Gavin

“Now only 2 or 3 firms, TSMC, Samsung and maybe Intel are able to make the world's state-of-the-art chips.”

0.57

NVIDIA is stuck in the innovator's dilemma because its high-margin H100 sells well, so launching an H100 variant for inference or a stripped-down chip would cannibalize margins, leaving startups able to move first on specialized chips.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gavin

“But right now they are stuck in the innovators dilemma. The H100 is selling like hot cakes and has an enormous margin on it.”

0.57

CUDA is a major NVIDIA moat for flexible general-purpose chips, but if the world shifts to chips with transformers burned into silicon, the software needed is much simpler, so NVIDIA's CUDA moat largely disappears.

causalhigh valuespeaker onlynovelty 3/4durability 3/4· Gavin

“In a world where you burn the transformer architecture into the silicon, you don't need CUDA in the same way.”

0.53

A chip that hard-codes the transformer architecture into silicon avoids flexible memory access, so it can fit an order of magnitude more compute and sustain over 90 percent utilization, returning responses in milliseconds rather than seconds.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“you can fit order of magnitude more compute and use it more than 90% utilization.”

0.53

Progress will proceed in cycles: roughly two years to build GPT-5, then three to five years to build the data centers for GPT-6, with models iterating slowly toward human-level agentic behavior.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“we're going to spend two years building GPT-5. And then we're going to build new data centers for GPT-6, and that will take three, five years.”

0.53

Scaling from GPT-4 to GPT-5 class models requires 10 to 100 times more chips and power, so future data centers will move from 20 megawatts to two gigawatts, consuming the entire output of a large power plant.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“This is not going to be a 20-megawatt data center as a two-gigawatt data center.”

0.53

Chip design cycles could compress from four to five years to a couple of months by spinning up huge data centers for back-end optimization and training language models to help with front-end code, though the required work is substantial.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“I think the upper limit really is a couple of months instead of the 4, 5 years of today”

0.53

The target for a transformer-specific chip is more than 10x better than GPUs, achieved by devoting far more silicon to multiply-accumulate units, since a multiplier takes about 10,000 transistors and NVIDIA devotes only about 4 percent of its chip to them.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“Only about 4% of that chip is the metric multipliers because to feed them requires so much more circuitry.”

0.53

Chips with about 20x better latency than leading GPUs, by processing huge prompts in full parallel, would enable real speech-to-speech conversation and let models ingest the whole internet in seconds instead of hours.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“With chips that have 20x better latencies than the state-of-the-art GPUs”

0.53

The future of models splits into very small on-device transformers for low-latency tasks like autocomplete and giant hyperscale models, while mid-sized models around 100 billion parameters have no niche because they are too big for devices and too small to justify compared with data center models.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“I don't think it's a future for medium-sized models, things in 100 billion parameter range.”

0.53

The realistic upper limit for chip design cycle time with heavy automation and compute is a couple of months rather than four to five years, though the first chips made on faster processes will likely fail and few firms want to go first.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin

“I think the upper limit really is a couple of months instead of the 4, 5 years of today”

0.53

At ten million transformer ASICs, a chip will fail roughly every couple of hours, requiring new fault-tolerance methods beyond checkpoint restarts, and cooling gigawatt-scale heat output is an open problem that will spawn a new industry.

forecasthigh valuespeaker onlynovelty 3/4durability 2/4· Gavin Uberti

“You have 10 million GPUs or 10 million transformer ASICs. Then you're going to have one failing every couple of hours.”

0.52

Once a specialized chip enters a market, general-purpose devices become uncompetitive, and the first mover is almost guaranteed to win, as the Bitcoin ASIC market shows with a MicroBT and Bitmain duopoly, so Etched must move as fast as possible.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“The person who gets there first is almost guaranteed to win.”

0.52

If training a model costs $1 billion and a custom chip costs $100 million, a chip that is 20 percent better makes economic sense, and at $10 billion of training cost it is a no-brainer to build custom silicon.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“if you're 20% better, then, yes, it does make economic sense to build that custom chip.”

0.52

The first chips built on a faster new process will probably not work, and because of high costs companies avoid being first, so industry adoption will wait until a few firms take the risk.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“the first chips that are made with some faster process probably aren't going to work.”

0.52

Humans keep participating in chess, Go, and running even after machines surpass them, because the value lies in doing a task against other humans, so AI may not eliminate human roles in entertainment and sport.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“chess has become more popular today than it was back then.”

0.52

The AI stack has layers: model developers at the top such as OpenAI and Anthropic who depend on compute from hyperscalers, which buy chips from designers such as NVIDIA, which rely on fabs such as TSMC, with application builders forming a fifth layer above.

factualhigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“this supply chain of model provider on top, then hyperscale data center below that, then chip designer below that and chip fab below this. This is the fundamental stack of the future.”

0.52

Specialized chips can reach production-grade software faster than general-purpose chips by being less ambitious, since supporting only transformer workloads makes the software stack far simpler than CUDA's.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“the only way to solve this problem besides being old, I think, is to be less ambitious.”

0.52

Leaders must be technical enough to understand their team's problems because that is what allows them to recruit the right people and to know how far along critical blocks like design verification really are.

normativehigh valuespeaker onlynovelty 2/4durability 3/4· Gavin

“you have to understand that problem and that's what being technical gets you.”

0.52

Latency, measured as milliseconds to first token and time between subsequent tokens, is the key bottleneck for real-time AI applications such as conversation, translation, and robotics.

causalhigh valuespeaker onlynovelty 2/4durability 3/4· Gavin Uberti

“I strongly agree that real-time AI is the bottleneck here.”

0.49

To enable the next wave of uses such as tree search and speech conversation, inference must become roughly 100x cheaper, latency must drop by an order of magnitude, and systems should generate about 1,000 tokens per second so code writing takes seconds instead of tens of seconds.

normativehigh valuespeaker onlynovelty 2/4durability 2/4· Gavin

“bringing the cost of running these models down by a factor of 100.”

0.49

Custom ASICs only make economic sense once demand is large enough to justify roughly $100 million of design spend, and ChatGPT and similar models have now crossed that threshold for transformers, whereas 18 months earlier they had not.

causalhigh valuespeaker onlynovelty 2/4durability 2/4· Gavin

“there wasn't enough demand to justify spending $100 million to burn that algorithm into the silicon. But now with ChatGPT”

0.48

Chip development proceeds through architecture and blocks, RTL coding in a hardware description language, front-end verification with test benches and coverage, netlist synthesis, back-end layout and timing closure, mask creation from GDSII, and finally fabrication with photolithography.

definitionestablishednovelty 1/4durability 4/4· Gavin

“then you'll take these blocks. Number of blocks a chip has varies widely, but it can be as high as hundreds.”

0.42

Pretraining is like grade school: tasks that are not directly relevant but easy to grade build baseline skills, and only then does specialized training like RLHF, akin to college, add the target capabilities.

definitionestablishednovelty 2/4durability 3/4· Gavin

“this baseline stuff has to be easy to grade. It doesn't necessarily have to be super relevant.”

0.40

Training costs more than inference because it runs the model forward and then backward to attribute error to each weight, roughly doubling compute, and it requires different network connectivity than forward-only inference.

causalestablishednovelty 1/4durability 4/4· Gavin

“To do that you have to run the model backwards, which takes about double the compute.”

0.36

Robotics transformers such as Google's RT-1 benefit from pretraining on huge text corpora to build world understanding, then adding robotic sensor inputs as another modality, which transformers handle naturally.

factualestablishednovelty 2/4durability 2/4· Gavin

“they pretrained the model with a huge amount of text”

0.35

Latency is measured as the delay before the first token returns after a user submits a prompt, and the gap between subsequent tokens, with the first-token delay being much longer.

definitionestablishednovelty 1/4durability 3/4· Gavin

“the number of milliseconds between when you submit your response and when you get your first token back”

0.34

A good strategic leader sets the vision, gets the right people into the key seats, and makes sure the company does not run out of money, relying on technical depth to judge who to hire.

normativespeaker onlynovelty 2/4durability 3/4· Gavin

“The role of a good leader is to set the vision, get the right people in those chairs and not running out of money. That's it.”

0.33

China did not get on the EUV lithography train early and is now using older immersion technology, which holds back its chips, so despite producing some 7-nanometer chips it will need better technology.

factualcontestednovelty 2/4durability 2/4· Gavin

“They're now on this older immersion technology, and it's really holding them back.”

0.31

Cooling data centers in the 200 to 300 megawatt range is manageable, but a two-gigawatt facility turns nearly all its power into heat, and no clear method exists for pumping that out, so a whole industry will need to evolve.

forecastspeaker onlynovelty 2/4durability 2/4· Gavin

“But what happens when the two gigawatts that go into that building are all turned into heat?”

0.30

After GPT-3 appeared in 2020 showing transformers were the best approach, earlier techniques such as convolutional networks and support vector machines were largely abandoned.

factualestablishednovelty 1/4durability 2/4· Gavin

“all this previous technology of convolutional networks and the support vector machine was thrown away in 2020 when GPT-3 came out”

0.30

Building a new semiconductor fab costs $10 to $20 billion, and TSMC buys lithography machines from ASML at about $300 million each, which in turn depend on Carl Zeiss optics, creating deep supply chain complexity.

factualestablishednovelty 1/4durability 2/4· Gavin

“that building a new fab costs $10 billion, $20 billion.”

0.29

Because high efficiency requires huge batches of users grouped together, inference will be centralized in large data centers in much the same way as training.

forecastspeaker onlynovelty 2/4durability 3/4· Gavin

“inference will be centralized in much the same way as training.”

0.29

Just as GlobalFoundries serves customers for reliability and American manufacturing despite not matching TSMC, smaller AI data centers will find niches without competing on raw intelligence.

forecastspeaker onlynovelty 2/4durability 3/4· Gavin Uberti

“You'll have a few smaller data centers that don't try to compete on being as purely smart as the big megaliths, but they'll find some niche.”

0.28

The chatbot text interface is not the ultimate use case, because token-by-token text is expensive relative to model cost, and richer uses like speech require much lower latency that current systems lack.

normativespeaker onlynovelty 2/4durability 2/4· Gavin

“this chatbot interface that I love as well is definitely not the end use case for these models.”

0.26

Etched's founders wrote a white paper and built a functional simulation of their transformer chip before approaching an industry veteran, Mark Ross, who then joined and helped raise funding, recruiting senior chip talent.

factualspeaker onlynovelty 1/4durability 2/4· Gavin

“We write our technical white paper. We do the original architecture for the chip. And we build our functional simulation”

0.26

Because AI is a young field without a century of accumulated work, low-hanging problems remain, so a single person can make a major impact, unlike in mathematics or chemistry.

normativespeaker onlynovelty 2/4durability 2/4· Gavin

“the field is so young that we haven't had 100 years to develop it, and we haven't had 100 years to solve these low-hanging problems. A single person can make a major, major impact.”

0.26

OpenAI and Anthropic's main moat is enormously good talent, since they do not run their own compute and instead rely on hyperscalers for it.

causalspeaker onlynovelty 2/4durability 2/4· Gavin Uberti

“I think that's the big moat, but they don't really run the compute themselves.”

0.25

The rumored OpenAI Q* project is allegedly a combination of Q-learning layered on top of RLHF with A* tree search, which would suggest OpenAI is experimenting with search, though this is only a rumor.

factualfringenovelty 2/4durability 1/4· Gavin

“it is the combination of Q-learning into A* tree search.”

0.21

Patrick notes that top H100 buyers are Meta and Microsoft with Amazon also large, and asks whether the buyer map changes when ASICs replace GPUs and Fortune 500 firms seek generative AI.

factualspeaker onlynovelty 1/4durability 2/4· Patrick

“Do you think that the market map and market structure looks a lot different”

0.20

GPT-5 and next-generation models from Google and Anthropic will be smart enough to escape loops and plan, enabling autonomous agent use cases that GPT-4 could not sustain.

forecastspeaker onlynovelty 1/4durability 1/4· Gavin Uberti

“GPT-4 is not quite smart enough to get out of loops and to go do the kind of planning needed to operating on its own.”

0.16

A high school teacher, Gunnar Mein, who had worked at Microsoft, led students in restoring a broken electron microscope over 1.5 years, an experience that sparked the speaker's interest in engineering.

factualspeaker onlynovelty 0/4durability 1/4· Gavin

“that got me interested in engineering, got me to Harvard in some ways”