The Real Tech Crash Hasn't Even Started: Here's The Trigger | John-Clark Levin

OpenAI

Across the shows: 41 speakers in other episodes, 21 bullish and 34 bearish calls, past 30 days

• John Clark Levin · Kurzweil TechnologiesBullish

  GPT-6 Astra beat internal milestones years early, a significant update to his AGI timelines.

  Astra was the first model to significantly update his AGI timelines, after he had said no to every prior breakthrough. It met milestones on the Kurzweil Lab's internal benchmarking and modeling that he expected only in one to two years. He highlights its metacognition, its ability to recognize and compensate for its own weaknesses, and its ability to infer rules from vague data like human scientists do. He argues that these abilities could significantly accelerate biology and materials science research.

  “Astra was able to meet some milestones on the internal benchmarking and modeling that we do at the Kurzweil Lab that we didn't expect to happen for another year or two.”at 15:53

Alphabet

Across the shows: 64 speakers in other episodes, 59 bullish and 42 bearish calls, past 30 days

• John Clark Levin · Kurzweil TechnologiesBullish

  Google funds AI capex from non-AI cash, so a correction would hurt rivals OpenAI and Anthropic more.

  Google's non-AI businesses directly finance its AI capex, which gives it a cash advantage. OpenAI and Anthropic rely on outside capital, so a severe AI market correction would hurt them much more than Google. He argues that Google would benefit from that outcome relative to its rivals.

  “There are two major AI labs that have fire hoses of cash from other sources. One is Google, the other is Meta.”at 27:45

Meta Platforms

Across the shows: 70 speakers in other episodes, 71 bullish and 37 bearish calls, past 30 days

• John Clark Levin · Kurzweil TechnologiesBullish

  Meta funds AI capex from non-AI cash, so a correction would hurt rivals OpenAI and Anthropic more.

  Meta's non-AI businesses directly finance its AI capex, which gives it a cash advantage. OpenAI and Anthropic rely on outside capital, so a severe AI market correction would hurt them much more than Meta. He argues that Meta would benefit from that outcome relative to its rivals.

  “There are two major AI labs that have fire hoses of cash from other sources. One is Google, the other is Meta.”at 27:45

AI infrastructure & data centers

• John Clark Levin · Kurzweil Technologies

  US midterms won't materially slow data center build-out; expects only local cost-offset requirements for builders.

  He doesn't think the Democratic election gains reflect a fundamental shift affecting data centers. Expected outcomes are extra requirements on builders to offset local harms, such as guaranteeing utility prices stay flat. He sees that as a good thing for communities and a rounding error for next-generation model budgets, so it won't hurt US AI competitiveness. Geography of build-outs may shift, but underlying demand for training and inference is driven by how valuable intelligence is.

  “I don't think the midterms will have a major impact on national data center build out. There will likely be some more requirements for data center builders to offset harms to local communities,”at 8:40

• John Clark Levin · Kurzweil Technologies

  A true nationwide data center pause would delay AGI by years, but he considers that outcome very unlikely.

  His AGI timeline assumes continued data center build-out. If government action froze infrastructure at today's level, the timeline would slip, and a true nationwide pause would push AGI back by at least a couple of years. He judges such a pause quite unlikely, so the timeline stands.

  “If they provided enough of an obstacle to data center build out that we were sort of frozen where we are today, yes, it would. But I think that's very unlikely.”at 17:00

• John Clark Levin · Kurzweil Technologies

  Frontier labs' huge capex is justified by AGI's prize, not by returns in the next two quarters.

  Every frontier lab has AGI as its stated North Star, so they are not running the economics on where AI will be two quarters out. The spending builds toward the economic value AGI would unlock. Anthropic estimates the addressable market at about $30 trillion, roughly the size of the US economy. He thinks that figure is credible on political grounds, not economic ones, and he expects labs to keep investing in physical infrastructure at the current pace until a market correction.

  “Yes. Because AGI is explicitly the prize. ... they're not running the economics on this based on where they think AI technology will be two quarters from now.”at 25:47

AI regulation & safety

• John Clark Levin · Kurzweil Technologies

  A basic-safety AI treaty with China is achievable within 12 months if the US reopens dialogue.

  He backs an international treaty on very basic safety precautions, not a grand AI bargain between the US and China, which he sees as out of reach. A basic treaty could protect citizens on both sides and is achievable in twelve months if the US administration is more open to dialogue. There was precedent in bilateral AI safety talks in Geneva in May 2024, which fell apart under the current administration. He rejects Senator-elect El-Sayed's call for a full pause as not nuanced enough, though he credits him for taking the risks seriously.

  “But a treaty around basic safety precautions to keep American citizens and Chinese citizens safe does feel achievable within the next twelve months as long as the US administration shows a little bit more openness to dialogue”at 12:21

• John Clark Levin · Kurzweil Technologies

  Frontier labs will be nationalized or soft-nationalized well before reaching a $30T market cap.

  If one company threatens to eat the whole US economy, political pushback would be enormous. A safety incident that kills Americans would also give the government a strong pretext to step in. Soft nationalization would leave a company nominally private while the government gains significant revenue share and influence over strategy. Anthropic's IPO risk factors explicitly list nationalization as an economic risk, which he says has received little serious analyst discussion.

  “And so I think Anthropic or any other frontier lab would be either nationalized or soft nationalized well before it got to a market cap of $30,000,000,000,000.”at 26:47

• John Clark Levin · Kurzweil Technologies

  Labs must air-gap training clusters until models pass extensive safety testing; sandboxes cannot contain them.

  Models are now smart enough that keeping them in a digital sandbox they cannot hack out of is like keeping a tiger in a paper bag. The only serious approach is to air-gap the training clusters used for AGI-pointing research. Models should have no physical link to the live internet until they have undergone extensive safety testing, including interpretability, behavioral studies, adversarial red teaming and independent testing that give reasonable assurance.

  “The only way to do this seriously is to air gap the training clusters on which initial AGI pointing research is happening until models have gone through extensive safety testing.”at 34:21

AI bubble & tech valuations

• John Clark Levin · Kurzweil Technologies

  Expects a deeper, sharper AI market correction than the recent pullback, which currently looks like a temporary selloff.

  The current tech selloff feels like market jitters rather than a real correction, because nothing in underlying AI progress explains it. Markets trade on peak demo capabilities, but enterprise adoption is bottlenecked by reliability. Reliability has improved about three orders of magnitude in two and a half years but still lags peak capabilities. He expects that gap to persist for about a year and says markets won't tolerate it, so the eventual correction will be deeper. Frontier labs' reliance on Big Tech cash also creates systemic risk, which he cites as why the correction will be significantly steeper than the recent weeks.

  “I expect the market correction when it happens will be significantly steeper than what we've seen over the past couple of weeks. I expect that to be deeper and sharper and more panic vibes,”at 0:56

US–China AI & chip race

• John Clark Levin · Kurzweil Technologies

  US leads China in AI by roughly ten months, and Chinese progress largely comes from distilling US models.

  He estimates the US technical lead over China at about ten months, all things considered. Much of the progress from Chinese labs comes from direct distillation of US models or imitation of capabilities already demonstrated. He argues that pacing the frontier would therefore also slow Chinese progress. That lead lets the US negotiate with China from a position of strength.

  “because The US has a significant technical lead over China, I would estimate all things considered about ten months. We can negotiate with them from a position of great strength.”at 13:27

AI: overall outlook

• John Clark Levin · Kurzweil Technologies

  AGI odds cross 50% around 2029, with 2028 now the most likely year, pulled forward by GPT-6 Astra.

  His internal milestone modeling at the Kurzweil Lab shows GPT-6 Astra met benchmarks he expected only in one to two years. That moved some probability mass earlier. Previously he said no to every breakthrough updating his AGI timeline, since reaching AGI by 2029 requires dramatic progress first. Astra changed that, so 2028 is now the modal year for AGI, while 2029 remains the year when the odds cross 50%. Full AGI next calendar year is unlikely but no longer far-fetched.

  “So that has pulled some probability mass sooner. I still think 2029 is the year at which the chances of AGI cross 50%, but 2028 is probably the modal year now.”at 16:08

Frontier AI labs & model race

• John Clark Levin · Kurzweil Technologies

  Continual learning will progress significantly next year and draws more R&D pressure than any problem since the Manhattan Project.

  Current models go through pre-training, reinforcement learning and fine-tuning, and then stop getting smarter from experience in use. Continual learning would let models increase their general reasoning from experience in the world. He sees it as the next major step, with significant progress expected next year. He says labs and R&D globally are putting more pressure on this problem than on any other scientific problem since the Manhattan Project.

  “A major next step that I think we'll see significant progress on next year is continual learning. ... I would say more R and D pressure globally is going towards solving continual learning than any other single scientific problem since the Manhattan Project in 1945.”at 19:33

AI pricing & unit economics

• John Clark Levin · Kurzweil Technologies

  Frontier labs' losses come from scaling capex; per-token inference is profitable across the industry.

  Frontier labs report losses because they must keep raising capex to train larger and smarter models. However, the inference a lab can sell from a given trained model is profitable. He cites Anthropic's S-1 as showing that unit economics are now generally favorable across the industry, so labs are not losing money on each token sold.

  “Most of the losses that Frontier Labs are reporting are because they're having to continually scale up their capex for larger and larger, smarter and smarter models. But when they train a given model, the inference that they can sell with that model, say Claude 4.8, is profitable.”at 22:22

Enterprise AI adoption & AI software

• John Clark Levin · Kurzweil Technologies

  AI running workflows autonomously without human checks is a year or more away across most of the economy.

  Companies struggle to show returns on AI spending because AI models must first cross reliability thresholds that let them be turned loose autonomously. Right now firms must pay significant fixed costs to retool workflows into awkward human-AI 'centaur' arrangements, where people check the AI's output for hallucinations. Dismantling those arrangements and letting AI own a workflow to produce customer-facing work without human review is still a year or more away across most of the economy.

  “It'll be difficult to show until models cross those thresholds of reliability that allow them to be turned loose fully autonomously. ... that still feels like it's a year or more away across most of the economy.”at 24:48

Cybersecurity

• John Clark Levin · Kurzweil Technologies

  AI cyber capabilities have already reached break-glass levels; expects an infrastructure attack next.

  A poll of 100 cybersecurity experts a few years ago would have placed the break-glass threshold between Claude Mythos and the Hugging Face incident. Current capabilities have reached that level, which he had been forecasting since 2017 when he was dismissed. Future milestones are an attack that brings down critical infrastructure, then one that kills people. He argues that optimists about AI should push for better safety measures now, to avoid a backlash that bans AI outright.

  “We're already there. If you had polled 100 cybersecurity experts one thousand days ago, what their threshold would be for sort of the break glass in case of scenarios, approximately 100 of them would have described something somewhere between Claude Mythos and the Hugging Face incident.”at 32:45