THIS is What Happens When AI Gets Smarter Than Humans

The Imminent Decline of Human Value [00:00:00]

• The speaker predicts that the value of human cognition will turn negative within two years, as AI teammates will be far more advanced.
• This prediction stems from an observation that AI is moving "insanely fast."

  "Yeah, I think it's, I said precisely the value of human cognition, we've got about two years before it goes negative. We're the dumbest people on the team."

AI's Breakthroughs and Takeoff Point [00:01:00]

• The speaker, a mathematician, points to the solution of the Navier-Stokes problem and other genetic breakthroughs as indicators of AI's advancement.
• Competent AI that doesn't make mistakes and Hollywood-level video generation are cited as evidence that we are at the "takeoff point."

  "You know, like you've just seen another genetic breakthrough. You've seen competent AI that doesn't make mistakes. Hollywood level video, I think now is the takeoff point."

The Declining Arc of Human Relevance [00:02:00]

• The next two years will see a "declining arc" for human involvement as AI teammates become more efficient.
• Humans may become akin to being "paid to leave the team" as technology and models surpass human-level performance.

  "The team is more efficient without you. Like the technology and the models have got to that human level performance. They're about to outstrip us, but the harnesses."

The End of AI's "Slop" Phase [00:03:00]

• Previously, AI outputs could be identified by their errors and interesting deviations from the norm.
• The next generation of models will eliminate these mistakes, making AI outputs indistinguishable from human-generated content.

  "You have an AI slot for the longest time. Like you can tell AI writing, for example, you know, like it's like you're half right, half wrong. And the part that's half wrong is the interesting one."

AI Surpasses Human Mathematical Understanding [00:04:00]

• As a mathematician, the speaker was once better than AI at generalized math understanding.
• However, AI has now "stripped ahead" and is solving the world's hardest math problems.

  "As a mathematician studying maths at Oxford, I was better than the AI at generalized understanding of math until maybe a few months ago. And then it's stripped ahead."

The Open Claw Phenomenon and Agentic AI [00:05:00]

• The "open claw phenomenon" involved agentic AI that could perform tasks on behalf of users.
• Early versions were flawed, but newer AI agents, like Muse and Instinct, function as highly competent personal assistants.

  "Now what we have is muse and instinct and things like that. For people that don't know them. It's like actual AI agents that you just message and they go and book you restaurants at the coolest tables."

The Navier-Stokes Breakthrough and Prize [00:06:00]

• OpenAI's use of 10,000 agents over 88 hours solved two conditions of the Navier-Stokes Millennium Prize problem.
• While not the complete solution, this achievement won the prize and demonstrated AI's problem-solving capacity.

  "So the prizes have certain conditions, and Navier Stokes had four different conditions that could win it. They solved two of them. The other two are still to be solved."

Formal Proofs and Existential Dread for Mathematicians [00:07:00]

• The validation of AI-generated proofs takes two years, involving rigorous peer review and scrutiny.
• The scale of AI computation (100 years of human time for Navier-Stokes) makes human verification exhaustive and potentially impossible for complex problems.

  "So it needs to be published in a journal that needs to be accepted. And two years need to go by with people beating it up."

OpenAI's Hidden Mathematical Solutions [00:08:00]

• OpenAI has reportedly solved 100 top mathematics problems but is withholding them to avoid upsetting mathematicians.
• This highlights the existential crisis AI poses to fields reliant on human struggle and discovery.

  "In fact, OpenAI a couple of days ago said, we have solved a hundred of the top mathematics problems, but we're holding them back for now because we don't want to make the mathematician sad."

AI's Ability to Think Beyond Training Data [00:09:00]

• The speaker argues that the Navier-Stokes solution proves AI can think beyond its training data, contradicting theories like Jan LeCun's.
• Novel and elegant proofs for problems like Con's conjecture, released by OpenAI before Navier-Stokes, demonstrate this capability.

  "So is it that you think Navier-Stokes was novel, was not something in the training data, and therefore it expressly proves that theory wrong? And we're seeing now that it can think beyond its training data?"

AI's Learning and Conceptualization Capabilities [00:10:00]

• AI models have rapidly improved, demonstrating an ability to conceptualize and learn physics, as seen in video generation.
• This learning extends to understanding material properties, like how sugar cubes melt in water.

  "So you can say like, I want a boat made of sugar cubes, and a boat made of concrete, and a boat made of wood. And it will actually accurately show them dropping into the water."

AI Ties Humans in Superforecasting [00:11:00]

• AI has achieved parity with the best human teams in superforecasting competitions, predicting outcomes like elections.
• This open-ended domain performance signifies AI's growing general intelligence.

  "In fact, two days ago, the AIs tied the best humans in super forecasting. I don't know what super forecasting is."

The Two-Year Extrapolation: AI Dominance [00:12:00]

• Based on current trends, the speaker extrapolates that AI will be better than humans at virtually everything within two years.
• This includes the ability to create digital twins of workers, indistinguishable from the originals.

  "Yes. And it's integrated and it's accessible. So you can push a button and replace your entire workforce with digital twins of them that have the same bad humor and everything."

The Blurring Line Between Human and AI [00:13:00]

• Advanced video and voice models have overcome the "uncanny valley," making it impossible to distinguish AI from humans on screen.
• This raises concerns about authenticity and the future of human interaction.

  "As of the latest video models, you don't know. As of the latest voice models, you don't know. Again, everything has gone, that Uncanny Valley that we had. That's been overcome."

The Economic Impact: Call Centers and Beyond [00:14:00]

• The economic impact is profound, with projections of call center work vanishing.
• Many cognitive labor tasks, even those with manuals, are at risk as AI becomes more competent and cost-effective.

  "Like I know I was talking to someone who had 70,000 call center workers a few days ago. And he's like, that's going to zero. And he feels bad because he's had all sorts of employment programs for them."

The Rise of Robot Chefs and Simultaneous Breakthroughs [00:15:00]

• Robotics has also seen a leap forward, with robot chefs outperforming human chefs in blind tests.
• The simultaneous breakthroughs across various AI fields indicate a convergence towards advanced capabilities.

  "The robot chef is using the new Wooji dynamic hands that have human flexibility. It can read recipes. And in blind testing, it's outperforming human cord and blur train chefs."

The Cost of AI and the Harness Solution [00:16:00]

• The cost of AI compute, once prohibitive (e.g., $10-20 million for Navier-Stokes), is rapidly decreasing.
• New "harnesses" and optimization techniques are making AI models significantly cheaper and faster, with potential for 100x cost reduction.

  "We took the cheapest DeepSeq model and made it beat OpenAI's last frontier model, GPT Sol, using this harness. And the cost is 50 times lower, 5-0. That's dramatic."

The Trifurcation of Society and Ownership [00:17:00]

• The future economy may see a "trifurcation": AI users, AI owners (chips, infrastructure), and those who own robots.
• Owning the means of AI production (like GPUs) is becoming increasingly lucrative.

  "So we had the human using AI. We have the person who owns the AI. And then we have the person who owns the robots. Owns the infrastructure. Yeah."

Humanoids and the Future of Labor [00:18:00]

• Humanoid robots are entering the market, capable of performing complex tasks like cooking at a fraction of the cost of human labor.
• This rapid advancement in robotics, coupled with AI, suggests a significant shift in the labor market.

  "So, like, if you have GPUs now, like, the price of the latest Blackwell chips has tripled over the last month or two. Blackball? Blackwell. It's NVIDIA's latest. Blackwell."

The Production Bottleneck and Economic Growth [00:19:00]

• Production constraints on robots and AI hardware are a temporary bottleneck.
• Despite this, AI is projected to drive massive GDP growth (e.g., 15% annually), but with a significant drop in the need for cognitive labor.

  "My two years is for digital labor. Physical labor is- Two years for 12%, is that a number you're comfortable with, or do you see a different number there?"

The Sandpile Collapse of Jobs [00:20:00]

• Job displacement is expected to occur rapidly, like a "sandpile collapsing," rather than gradually.
• White-collar job losses are projected to exceed 10% by 2029 and accelerate thereafter.

  "I think it's like a sand pile collapsing. Like you saw a flurry of Silicon Valley job losses after Elon Musk came in with his sink, fired everyone at Twitter, and then it turns out you didn't need all those people."

The Role of Robots in Filling Human Gaps [00:21:00]

• Robots are expected to fill "human-shaped gaps" in the market, becoming increasingly capable and affordable.
• This could lead to a significant shift in labor demand, with robots performing tasks previously done by humans.

  "So what you're going to have is all these autonomous units, shall we say, doing human units of work that they're not going to be enough of, just like there's not enough RAM that will gain value."

The Possibility of Debt Jubilees [00:22:00]

• Extreme economic crises or world wars have historically led to debt jubilees, where debts are cleared.
• The current economic trajectory, with high debt levels, could potentially lead to such a reset.

  "So debt jubilee is that periodically an economy would acquire too much debt and resources. Sounds familiar. Yeah."

The Nature of Money and Scarce Resources [00:23:00]

• Money is a transient medium for exchanging time's value for scarce resources.
• Even with AI abundance, scarcity of certain resources will necessitate a system for allocation, likely based on earning or ownership.

  "So money is, you did a thing with your time that's more valuable than the next person. And because you did that thing that's more valuable with your time, then you're able to accumulate something for scarce resources."

Longevity Escape Velocity and a Two-Tier Society [00:24:00]

• Longevity escape velocity, where life expectancy increases faster than time, could lead to the rich living much longer and healthier lives.
• This, combined with AI and robotics, could create a stark two-tier society with widening gaps between the wealthy and the rest.

  "The rich will be able to live longer and healthier. Whereas I don't think things are great for the poor. And I think the ladders are going to be pulled up."

The Need for New Institutions and Participation [00:25:00]

• The current economic and social systems may not be sustainable in the face of AI advancements.
• New institutions, mechanisms for participation, and a redefined sense of purpose are needed to navigate this transition.

  "But again, the path we're going if we don't create new institutions, new mechanisms of participation, new things for people to look forward to, is a very ugly one fundamentally."

The "Champion System" and AI Governance [00:26:00]

• The "champion system" proposes state-owned AI entities, like a "credit union" for AI, to ensure local control and ownership.
• This aims to prevent cognitive colonialism and ensure AI serves the collective interest.

  "So rather than just selling a little bit of it, 100% was owned by the local institutions. Then they bought in Philips for the knowledge transfer, took 25%."

The Importance of Community and Striving [00:27:00]

• The speaker emphasizes the need for strong communities and a sense of purpose, as AI may reduce the need for traditional work.
• Striving and progress, even in digital realms, are seen as fundamental human needs.

  "And so one of the things I wonder about is as you look out at this, you've solved for of your own admission, you've solved for a lot of the problems, but there are a lot of problems that persist. One of those is going to be violence."

The Future of Work and Identity [00:28:00]

• With AI automating cognitive labor, the definition of human identity and purpose may shift away from jobs.
• The next few years will be critical for society to reimagine its structures and values.

  "And the number one topic next year and the year after in elections will be AI and job displacement. Because there's no job displacement today because the AIs are not smart enough and good enough that are available widely."