Jeff Dean最新判断：AI的下一个拐点 | Google | TPU | AI芯片 | 大模型推理 | 上下文工程 | 创业机会 | 知识蒸馏 | 神经网络 | AI硬件 | 科学方法自动化

Jeff Dean discusses AI progress, hardware innovation, agentic systems, and startup opportunities at YC Startup School.

Key Takeaways

AI capabilities have rapidly surpassed earlier expectations, especially in agentic and coding tasks.
Hardware specialization is essential for efficient, scalable AI inference and lower latency.
Automated experimentation and problem decomposition will drive progress in AI and other scientific fields.
Entrepreneurs should challenge assumptions and think from first principles to find new opportunities.
The TPU exemplifies how targeted hardware innovation can transform AI performance.

What the video covers

Jeff Dean reflects on his previous assessment that AI had reached junior engineer level and admits models have improved faster than expected.
Agentic AI systems are now capable of running complex, long-duration tasks by decomposing problems and automating experiments.
The 2026 'everything-fits-into-memory' moment is anticipated, analogous to Google's 2011 shift to in-memory search, enabling faster AI processing.
Hardware innovation, especially high-performance, low-power inference chips, is critical for scaling AI and reducing latency.
Custom chips like TPUs, designed for dense linear algebra, have revolutionized machine learning inference and will inspire future hardware.
Entrepreneurs should focus on first-principles thinking and hardware specialization to create breakthroughs beyond current mainstream solutions.
Jeff emphasizes the importance of automated scientific experimentation and problem decomposition across fields, not just AI.
Speed and energy efficiency in inference hardware are key to enabling widespread AI agent deployment.
The TPU's origin story highlights the value of rough order-of-magnitude 'napkin math' in innovation.
Cultivating good judgment in AI and engineering requires experience and re-examining assumptions regularly.

Chapters

Has AI reached the level of a junior engineer as Jeff Dean previously stated?

Jeff Dean admits that while AI models have improved rapidly and are quite capable in agentic and coding tasks, the earlier assessment was somewhat an overstatement by general junior engineer standards.

What is the significance of the 'everything-fits-into-memory' moment for 2026?

This moment refers to a breakthrough where AI systems can run entirely in-memory, similar to Google's 2011 search upgrade, enabling much faster processing and longer-running agentic AI tasks.

Why is hardware specialization important for AI's future?

Hardware specialization, such as custom low-power inference chips like TPUs, is crucial for reducing latency and energy consumption, allowing AI systems to scale efficiently and serve more users.

Full Transcript — Download SRT & Markdown

Speaker A

Hello everyone, this is Best Partner. I'm Da Fei. Jeff recently had a conversation at YC's Startup School, and the content was very dense—covering everything from the boundaries of AI capabilities to the essence of hardware, and from agentic systems to startup strategy. Today, we’re going to talk about this video. The host started with a very direct question: last May, Jeff said on Agent that AI had reached the level of a junior engineer. A year later, was that assessment accurate? Jeff candidly

Speaker A

startup strategy. Today, we’re going to talk about this video. The host started with a very direct question: last May, Jeff said on Agent that AI had reached the level of a junior engineer. A year later, was that assessment accurate? Jeff candidly

Speaker A

replied that models have improved very quickly in agent-based tasks and coding tasks with longer runtimes. Looking at it now, they are already quite capable.

Speaker A

But if defined by the general standard of a junior engineer, he feels that assessment was basically an overstatement. However, he admitted one thing: he underestimated how fast models would grow in handling complex tasks; it’s faster than he anticipated. And it’s not just coding

Speaker A

But if defined by the general standard of a junior engineer, he feels that assessment was basically an overstatement. However, he admitted one thing: he underestimated how fast models would grow in handling complex tasks; it’s faster than he anticipated. And it’s not just coding

Speaker A

the machine learning systems themselves . What does that mean? It means letting machine learning systems run massive experiments on their own, breaking problems into sub-problems, putting those sub-problems into tight automated experiment loops, and stitching the results back together to improve their

Speaker A

; agentic systems are starting to shine in other fields as well. So, what about 2027? The host followed up by asking, what is the next bold prediction? He thought for a moment and said he expects to see much more automation of

Speaker A

interesting historical analogy. In 2011 , Google Search was still running on hard drives. Jeff and his old partner Sanjay did the math and realized that one day the entire search index could fit into the RAM of all their machines.

Speaker A

the machine learning systems themselves. What does that mean? It means letting machine learning systems run massive experiments on their own, breaking problems into sub-problems, putting those sub-problems into tight automated experiment loops, and stitching the results back together to improve their

Speaker A

everything-fits-into-memory" moment for 2026? What should the 6,000 entrepreneurs in the room be focused on today? Jeff's answer pointed toward hardware. He said that while the situation isn't exactly the same, you will see more and more high-performance , low-power inference hardware systems.

Speaker A

own capabilities. He calls this fully automated problem decomposition and automated experimentation. He believes this doesn't just apply to machine learning; any scientific or engineering field where you can define a measurable goal could see significant progress. At this point, the host brought up an

Speaker A

said we have all become accustomed to waiting for models to return results, but waiting is not fun, and speed is king. The host asked, what if we didn't have to wait at all? He said, right, you can imagine what would happen if

Speaker A

interesting historical analogy. In 2011, Google Search was still running on hard drives. Jeff and his old partner Sanjay did the math and realized that one day the entire search index could fit into the RAM of all their machines.

Speaker A

Jeff's answer might surprise many people. He said, people haven't truly realized that it is already possible for agent systems not just to run for an hour or two, but to leverage sufficiently powerful underlying models to run for days or even weeks to handle

Speaker A

So, within a few days, they launched a brand new version of search that ran on memory instead of hard drives. That was essentially the key turning point that made Google Search fast. History always rhymes. So, what is the

Speaker A

He gave an example. You could have an agent implement a brand-new version of a piece of software, for instance, rewriting it in a different programming language for better security or performance, and the AI could actually carry this out quite seriously. At this

Speaker A

"everything-fits-into-memory" moment for 2026? What should the 6,000 entrepreneurs in the room be focused on today? Jeff's answer pointed toward hardware. He said that while the situation isn't exactly the same, you will see more and more high-performance, low-power inference hardware systems.

Speaker A

usable. He did the math: if every Google user just used three minutes of speech recognition on their phone every day, Google's entire server fleet would have to double, which would be absurd, so they ended up building a custom chip

Speaker A

Because everyone now realizes that inference is the key to allowing agent systems to serve more people, and latency is critical, hardware specialization is the essential path to making systems more energy-efficient and lower-latency than general-purpose computing devices like GPUs or TPUs. He

Speaker A

rate was cut in half. That was almost equivalent to achieving the total progress of the past 20 years of speech recognition in just a few short months.

Speaker A

said we have all become accustomed to waiting for models to return results, but waiting is not fun, and speed is king. The host asked, what if we didn't have to wait at all? He said, right, you can imagine what would happen if

Speaker A

was born, which is essentially a chip specifically designed for dense linear algebra, the very core of almost all modern machine learning algorithms. It can't run C, nor can it run Word, but it is exceptionally useful for machine learning inference. The initial chip

Speaker A

latency could be reduced by a factor of 50. The host asked again, what is an AI assumption that everyone still takes for granted but is actually wrong?

Speaker A

Interestingly, when they were invented, the Transformer architecture hadn't even been released yet. He couldn't have predicted how fundamental the TPU would become, but he noted that precisely for this reason, they built a general-purpose linear algebra system rather than hardware over-specialized

Speaker A

Jeff's answer might surprise many people. He said, people haven't truly realized that it is already possible for agent systems not just to run for an hour or two, but to leverage sufficiently powerful underlying models to run for days or even weeks to handle

Speaker A

Subsequent generations of TPUs were also able to allow more chips to work together efficiently to solve the same problem. So, what kind of "financial math" should today's entrepreneurs be calculating? Can they figure out the next thing as important as the TPU? He

Speaker A

very complex tasks. Some people have seen a glimpse of this, but not everyone has truly internalized this fact. And this is going to be a very big deal. So, in what situations would you need to let an agent run for weeks?

Speaker A

magnitude. Don't be locked into current mainstream solutions; think from first principles. If you were to solve it from scratch, what would you do? This often leads to directions others haven't thought of. The host then followed up with an insider's question:

Speaker A

He gave an example. You could have an agent implement a brand-new version of a piece of software, for instance, rewriting it in a different programming language for better security or performance, and the AI could actually carry this out quite seriously. At this

Speaker A

almost a bible for distributed systems engineers. Approaching 2026, if this list were to be updated to an AI version, what should it include? He provided several key numbers. What is the bandwidth between storage and the accelerator's memory? What is the

Speaker A

point, the topic turned to the origin of the TPU itself; Jeff Dean is especially famous for his skill at "napkin math," which is quickly doing rough order-of-magnitude estimates on a napkin. Around 2013, Google's speech recognition started to become truly

Speaker A

How many chips can be connected at this bandwidth, and how does network bandwidth degrade when scaling from 500 chips to 10,000 chips? These numbers will truly influence how you think about solving certain types of problems . Then, he mentioned a figure that

Speaker A

usable. He did the math: if every Google user just used three minutes of speech recognition on their phone every day, Google's entire server fleet would have to double, which would be absurd, so they ended up building a custom chip

Speaker A

Even just moving data from an accelerator’s high-bandwidth memory into the processor to actually compute it comes at a much higher energy cost.

Speaker A

, and that was the origin of the TPU. He recalled that at the time, the deep learning-based speech system was actually of very high quality; compared to the old system, the computational cost was much higher, but the error

Speaker A

transmit many samples or tokens in a single batch to amortize the cost of moving them. This way, you are no longer paying a 1,000-fold energy penalty, but rather 1,000 divided by the batch size. However, if you want extremely low latency, batch processing

Speaker A

rate was cut in half. That was almost equivalent to achieving the total progress of the past 20 years of speech recognition in just a few short months.

Speaker A

inference truly requires very low latency, whereas training doesn't necessarily. He feels that inference hardware is currently far from being specialized enough. The core is to minimize data movement, and to consider extremely low-precision arithmetic; perhaps we don't need to support so

Speaker A

Tune a model, scale it up, feed it better data, and the results appeared. But he also started to worry that if speech really got better, people would use it more, and continuing to run it on CPUs wouldn't work; thus, the TPU

Speaker A

increasingly important is everything surrounding the model. It's not just the scale of parameters or training data, but also retrieval tools, memory tools, and so on. Jeff agreed with this , stating that the model is just one part of the entire system you are

Speaker A

was born, which is essentially a chip specifically designed for dense linear algebra, the very core of almost all modern machine learning algorithms. It can't run C, nor can it run Word, but it is exceptionally useful for machine learning inference. The initial chip

Speaker A

context. A key point here is that training data is like a hodgepodge, with trillions of tokens mixed together and kneaded into hundreds of billions of parameters; in contrast, the context the model sees directly in a current problem is much clearer. Furthermore,

Speaker A

they ultimately produced was 30 to 80 times more energy-efficient than CPUs and GPUs, with 20 to 30 times lower latency. This was an astonishing foundation from which the TPU later evolved into what it is today.

Speaker A

is effective. This is the orchestration capability of complex agent systems and multi-agent systems. What's particularly interesting about the field of context engineering is that anyone can actually do it. Previously, training models required massive resources, GPUs, and data, but for

Speaker A

Interestingly, when they were invented, the Transformer architecture hadn't even been released yet. He couldn't have predicted how fundamental the TPU would become, but he noted that precisely for this reason, they built a general-purpose linear algebra system rather than hardware over-specialized

Speaker A

models and toolchains to parse problems , and then you will often see exactly where the model goes wrong. Often, the way to make a model perform better is not to tune its parameters from the outside, but to give it better guidance

Speaker A

for a specific algorithm. Since machine learning algorithms continue to evolve, you cannot make hardware too specialized, yet it must be specialized enough to achieve massive performance gains, such as using very large multiplier units, high-speed memory, and high-speed interconnects.

Speaker A

gave an example from his own personal experience. A few weeks ago, he and Sanjay were working on performance optimization for some very low-level libraries. Within Google, they have a custom-built microbenchmark library that can measure how long many different operations take. Some data

Speaker A

Subsequent generations of TPUs were also able to allow more chips to work together efficiently to solve the same problem. So, what kind of "financial math" should today's entrepreneurs be calculating? Can they figure out the next thing as important as the TPU? He

Speaker A

see what improved. You might also need to run a broader set of benchmarks to test cache usage and so on. So, they wrote a skill that essentially teaches the model how to complete these tasks in a specific sequence, including

Speaker A

said that it's hard to generalize. But the core idea is to first look at the problems and bottlenecks in your field, then ask yourself if there is a completely different solution from today's that could improve key metrics by an order, or even two orders, of

Speaker A

utilize. A few months ago, they also published a document called " Performance Hints," which he and Sanjay wrote. It's about 30 pages long and covers many performance optimization techniques. People have already summarized this document and fed it to

Speaker A

magnitude. Don't be locked into current mainstream solutions; think from first principles. If you were to solve it from scratch, what would you do? This often leads to directions others haven't thought of. The host then followed up with an insider's question:

Speaker A

When it comes to agents, a very realistic issue that everyone has surely seen is that an AI might do well for the first ten steps, but by the 30th, 40th, or even 50th step, it starts to go off track. He said that

Speaker A

Years ago, he wrote a famous list called "Latency Numbers Every Programmer Should Know." For instance, how long a cache miss takes, how long a disk seek takes, or how long a network packet takes to travel from California to the Netherlands. This list became

Speaker A

comfort zone, performance begins to drop—the further it gets from its familiar domain, the more likely it is to fail. Two approaches were suggested: first, provide the model with skills and hints to keep it on more familiar and reliable paths; second, use a

Speaker A

almost a bible for distributed systems engineers. Approaching 2026, if this list were to be updated to an AI version, what should it include? He provided several key numbers. What is the bandwidth between storage and the accelerator's memory? What is the

Speaker A

is a highly useful general technique: using inference-time computation to search for feasible solutions, making long-chain workflows much more reliable . Within Google, they have various harnesses and a whole suite of skills.

Speaker A

bandwidth to the multiplier units, and how much energy does one multiplication operation consume? What is the interconnect bandwidth between chips?

Speaker A

entrepreneurs present were concerned about: since Google co-designs almost every layer of the system from processors to products, what opportunities are left for small teams of two or three people? He said that Google and similar models are indeed striving to build models that are as

Speaker A

How many chips can be connected at this bandwidth, and how does network bandwidth degrade when scaling from 500 chips to 10,000 chips? These numbers will truly influence how you think about solving certain types of problems. Then, he mentioned a figure that

Speaker A

set of models and skills—or even a more specialized model—you can absolutely gain a distinct advantage.

Speaker A

might cause many to re-examine AI system design. He said that performing a calculation only takes about a picojoule of energy, but moving data once costs 1,000 times that amount.

Speaker A

thing you are building is a resilient direction. Will the model do it in the next six to twelve months, or is it something that can't be done for two or three years? So, how exactly should entrepreneurs judge which directions

Speaker A

Even just moving data fro

Speaker A

already ahead of the pack. The second step is to look at what today’s general-purpose models can already achieve on that specific problem. You can test it yourself. If they fail completely, it’s usually a good sign.

Speaker A

If they can already do a little bit, but not very well, it isn’t necessarily a good sign, because it usually means the capability has started to emerge, and with just a bit more training data and a larger scale,

Speaker A

the model will continue to evolve. So, what you are looking for are tasks where the model’s current success rate is 0%or 1%, rather than 20%.

Speaker A

Sometimes these problems are types that fall outside the training distribution. Other times, it’s because your product possesses certain data that the underlying general model lacks, such as a product that helps organize personal information. Since the general model itself doesn't have this personal data,

Speaker A

your product suddenly gains a significant advantage. Another situation is when dealing with a very difficult problem; if you can obtain the right training data to train a more specialized model, the cost may not be that high, but the precision will be

Speaker A

exceptionally high. For example, Fold is a highly specialized model for protein folding, which is extremely successful, but it is not a general-purpose model. Fields like materials science and chip design might also be suitable for a similar path. So

Speaker A

, how does one become an AI-native founder? He says that when defining requirements, if you can clearly articulate what you want, you will be much more successful. The clearer you are, the more boundaries and rules you have, and the better you know what you

Speaker A

need to accomplish. In fact, since the early days of computer science, we have been telling students to clarify exactly what the software needs to accomplish before writing it. Now, it has just been replaced by agents writing it for you, but the importance

Speaker A

of specifications has become even higher. Because in the past, you were handing tasks to a very smart person who naturally had context and could ask you questions; while agents might sometimes follow up, overall, clear specifications remain a very good

Speaker A

practice. He cited a currently highly effective use case for coding agents: translating software from one programming language to another.

Speaker A

Today’s models perform exceptionally well on this, and the reason isn’t a mystery; it’s because the specifications here are particularly clear. The original software itself serves as the specification. The model can use all tests from the original version to ensure the new version

Speaker A

passes them, and it can also translate the tests themselves, continuously comparing behavioral differences between the two until there are none left. Precisely because the specifications are clear enough, this type of task is extremely effective.

Speaker A

When orchestrating hundreds of agents becomes the standard and all code is written by agents, what will become the scarcest ability? Jeff’s answer is taste—judging exactly which problems should be solved by agents. From his own research background, this is

Speaker A

essentially the core of choosing a research problem. A researcher may possess all the tools and techniques, but victory or defeat is largely decided by the problem you choose to invest your time in at the start.

Speaker A

Choosing a good problem and actually solving it is far superior to studying a boring problem in exquisite detail.

Speaker A

So, how do you cultivate that taste? He says it’s truly not easy, nor does it have clear, measurable goals like certain metrics. Part of it comes from experience; having worked on many different problems in the past teaches you what might be worth doing in the

Speaker A

future. One method is to write down things you think might be important over the next 12 months. You might only choose one to work on, but looking back 12 months later to see which predictions turned out to be important,

Speaker A

which were achieved by others, and which didn't happen, you are generating more samples to train your ability to form taste. The third method is to conduct some truly wild thought experiments, and don't take for granted what everyone assumes to be true. He

Speaker A

shared a thought experiment he recently did with a colleague: for the past 60 years, the entire chip design and manufacturing industry has been striving to make transistors smaller and smaller with lower error rates, because the default goal is that every

Speaker A

chip of the same design should be identical, with no bit flips, but at a macro level, that’s not how we build systems. Large-scale distributed systems are built to be reliable using unreliable components; a single disk might fail, but the data cannot be lost

Speaker A

. So we use triple replication and Reed-Solomon encoding. Yet at the foundational scale of transistors, it seems this line of thinking hasn't been taken to the extreme. What if you built a system where transistors might fail 20 times a day, rather than once every

Speaker A

few million years? That would be a completely different design point, one that might open up very interesting new possibilities in manufacturing. For instance, signal transmission might take multiple redundant paths to ensure at least one arrives; this sounds a bit

Speaker A

like neural computing and the human brain. Signal transmission in the brain isn't an inherently ultra-reliable method; when important information needs to be sent, it often takes multiple pathways. He says he isn't saying we should do this right now, but

Speaker A

such thought experiments are important; every so often, we should re-examine those premises we take for granted. In his career, the TPU itself is an example of a default assumption being thrown out the window, ultimately becoming an influential system. Back

Speaker A

when that problem domain wasn't as important as it is today, they started thinking about whether they could design something specifically for a niche problem. Another example is MapReduce. At that time, he and others at Google had written a lot of code for

Speaker A

crawling and indexing systems, along with many manual parallelization and fault-tolerance mechanisms, but the problem was that this code would get tangled up with the simple logic they actually wanted to execute. Later, they looked at the problem from a different

Speaker A

angle and proposed an abstract computing framework: write simple data processing logic on top, while the bottom layer handles all parallel scheduling and fault tolerance. It eventually became an incredibly successful method for Google to handle massive-scale computing, and it all

Speaker A

started with a simple experiment. Now, we have AlphaChip helping with chip floorplanning and systems like Alpha that can propose, evaluate, and retain effective solutions; it feels like we are building an entire ecosystem where AI is used to create better AI. He says

Speaker A

that more generally, this is essentially automating the scientific method—for example, proposing experiments, building the execution environment, running them, evaluating the results, and moving forward based on those results. Now, more and more problems can have this loop automated, and what's more, we can run a vast

Speaker A

number of experiments. Once you drive the latency of the loop down to be extremely low, it becomes very significant. The key is that you must first make the evaluator fast enough.

Speaker A

He gave an example: about ten years ago , colleagues studying quantum chemistry problems to understand molecular properties usually had to run a full density functional theory simulator.

Speaker A

But after collecting a lot of inputs and outputs, they trained a neural network-based approximate simulator that was 300,000 times faster. And its accuracy was almost on par with the full simulator, which completely changed how science is done. Previously

Speaker A

, screening 10 million candidates might have taken six months of computing resources; now, it's done before lunch.

Speaker A

Many fields still have plenty of room to build faster validation models, and once achieved, it can redefine the entire experimental cycle. The direction he is most excited about is allowing models to improve themselves recursively by running massive numbers

Speaker A

of experiments, optimizing how many new discoveries can be gained per unit of compute input. As the conversation drew to a close, the host discussed setbacks in professional growth. In 2014, Geoffrey Hinton and Oriol Vinyals wrote a paper on knowledge distillation, with

Speaker A

large models acting as teachers and small models as students. This leads to cheaper, smaller, but still effective models that the entire industry uses today, but that paper was actually rejected by NIPS back then. Because the reviewers said this work was unlikely

Speaker A

to have a significant impact, Jeff said he actually disagreed with the program committee. Reviewers were perhaps thinking from a purely academic perspective about whether it was a fundamental theoretical breakthrough, while they knew it was a crucial problem because they wanted to serve

Speaker A

models to people and cover more scenarios. They posted the paper on the web, everyone read and used it, and everything went great. The Flash model today is partially derived from larger PR models. It is one of the best

Speaker A

performers on benchmarks for its model size, and the lesson he learned is that even if you are rejected, you must keep going. If the young Jeff from the 20- person Google startup in 1999 were transported to today, would he choose a

Speaker A

frontier lab or start his own company? Jeff said this is always hard to generalize. To him, the most important question is whether you are working on something you truly care about, working with people you like, and whether it

Speaker A

will change the world in a positive way . Organizations have ready-made structures and many brilliant colleagues, providing a natural platform to create an impact. Small startups are completely different; you must have a strong passion for the problem, be willing to take risks, but

Speaker A

it can also be very rewarding. No matter which path you choose, at least ask yourself: if this thing reaches the best possible outcome, will the world be better in some way because of it? If the answer is just "it's cool" but

Speaker A

nothing more, then it's not something you should be investing your time in. Regarding how to work with smart people and find them, he said you certainly want to find people who are strong in key areas and have complementary skills

Speaker A

, but you also want to find people you enjoy being around, because you will spend a lot of time "grinding" on hard problems together. He has always felt the most interesting thing about small teams is that others know things you

Speaker A

don't, and you have strengths they might lack; together, you can build things that might not be possible individually. You should treat your engineering or research career like a tool belt; you have to keep adding new tools to it because you never know what

Speaker A

problems you will encounter, and the more tools you have, the more problems you can solve. The host asked one final question: what problems does he most hope future creators will work on? Jeff said there are too many interesting

Speaker A

problems in the world, but he is personally particularly excited about new hardware paths, as more efficient inference hardware is extremely important. Furthermore, the learning algorithms themselves could perhaps be entirely different, such as being much more data-efficient than today's

Speaker A

methods. Today's large models see 1,000 times more data by the age of 18 than a human does, yet humans at 18 often perform better or at least hold their own in many tasks. So, is it possible to build systems that are far more

Speaker A

data-efficient and can continuously learn from their own actions? Continuous learning itself is a deeply fascinating problem. The interaction with AI is also very interesting.

Speaker A

Looking a bit further out, how can we enable better conversations among people around the world? Is there a way to facilitate more civilized discourse and help people connect with others they were meant to know, based on shared interests? These are all very

Speaker A

intriguing directions. The world is vast, and there are many problems. After this conversation, my biggest takeaway is the way Jack Dean approaches problems; he always has a penetrating power that pushes upward from fundamental constraints. He doesn't speculate on when AI will

Speaker A

arrive, but instead starts from the hardest physical constraints like energy costs, data movement, and latency to derive how systems should be designed and where entrepreneurial opportunities lie. When everyone is obsessed with model parameters, he tells you that data movement is 1,000

Speaker A

times more expensive than computation, and that specialized inference hardware is the key to the next wave of efficiency. When everyone is anxious that general-purpose models will swallow up all startups, he tells you to look for fields with a 0%to 1%

Speaker A

success rate, to find niches with proprietary data and models; when everyone is learning how to write prompts, he tells you the scarcest skill in the future will be taste—the ability to judge what problems are worth solving. These judgments didn't

Speaker A

come out of thin air; they were distilled from decades of system-building experience, from countless complex calculations, and from thought experiments that challenged every assumption. Finally, I want to ask you: if you were to challenge an assumption that everyone

Speaker A

currently takes for granted, what would it be? Thanks for watching, and see you next time.

Topics:Jeff DeanAI capabilitiesagentic systemsTPUAI hardwaremachine learning inferencestartup strategyautomated experimentationlow-power chipsAI acceleration

Get More with the SozAI App

Transcribe recordings, audio files, and YouTube videos — with AI summaries, speaker detection, and unlimited transcriptions.

App Store Google Play

Or transcribe another YouTube video here →

Free tools: TXT to SRT · SRT Validator · Merge SRT · Subtitle to Text · All tools