In partnership with
In Todayโs Issue:
๐ A serious plan to delay superintelligence to 2040
๐ง Richard Sutton's new lab bets against today's LLMs
๐ Oracle craters and SpaceX hits a record low
๐ค The Altman vs. Amodei feud, next round
๐ค A humanoid that learns its job by trial and error
โจ And more AI goodnessโฆ
โก The Signal
The same researchers who war-gamed a runaway 2027 have published their answer to it, and it is a brake, not an accelerator.
The AI Futures Project's new "AI 2040: Plan A" argues the world should deliberately push superintelligence out to 2040 through a US-China transparency deal and verified limits on compute, instead of sprinting there by 2030. It lands on a day when the market is asking the same question from the opposite side: Oracle has lost nearly half its value since June, Larry Ellison has slid from the world's #2 fortune to #8, and freshly public SpaceX just hit an all-time low as its "AI play" story wobbled. The bet on an endless AI build-out is meeting a safety argument and a valuation reality check at the same time. Whether the brake comes by treaty or by selloff, "faster" is no longer the only story anyone believes.
All the best,
Kim Isenberg
(Richard Sutton, via the-decoder)
๐ง Sutton's New Lab Bets Against Today's AI
Richard Sutton, the Turing Award-winning "father of reinforcement learning," has left John Carmack's Keen Technologies to launch Oak Lab, a startup betting that today's large language models are a dead end. With former student Khurram Javed, he is building the OaK architecture ("Options and Knowledge"): agents that learn continuously from real-time experience, without storing or replaying data, on far less compute. His wager is blunt: real intelligence comes from doing, not from reading everything humans have already written.
๐ tl;dr: The most decorated name in RL just bet his new company that the LLM era is a detour.
(Starship on the pad; AP via NPR)
๐ SpaceX Sinks to an All-Time Low
Six weeks after the biggest IPO in history, SpaceX (SPCX) has fallen to a record low near $139, down 38% from its $225 peak and below its debut price. The trigger was a mix of reusable-rocket milestones from China's Long March 10B and a Japanese vehicle, plus growing doubts about the orbital data-center vision that let investors treat SpaceX as an "AI play." The company still carries a roughly $2 trillion valuation despite about $5 billion in losses last year.
๐ tl;dr: The market's favorite new "AI stock" just learned a narrative can deflate as fast as it inflated.
(Larry Ellison; via Forbes)
๐ธ Oracle Craters, Ellison Loses $125B
Oracle has crashed nearly 50% since its June 1 peak, wiping out roughly half a trillion dollars in value and knocking co-founder Larry Ellison from the world's #2 fortune to #8, now behind Nvidia's Jensen Huang. Investors are balking at Oracle's plan to spend up to $95 billion on AI data centers for customers like OpenAI, and S&P cut its credit rating on fears the build-out could strain the company's finances. Ellison's net worth has fallen about $125 billion in six weeks.
๐ tl;dr: The clearest sign yet that Wall Street is rethinking who actually pays for the AI build-out.
Turn any AI model into a devil's-advocate analyst before you believe a hot take.
Why it helps: Whether the claim is "the AI bubble is bursting" or "superintelligence by 2027," the loudest takes are rarely stress-tested, and a model can argue the other side in seconds.
Try this: "Here is a claim I keep seeing: [paste the claim]. Steelman it in three bullet points, then give me the three strongest counterarguments and the single piece of evidence that would most change my mind. Be concrete, and name what would prove each side wrong."
๐ฌ Watch This
The Hidden Bottleneck in AI Factories
Everyone obsesses over GPUs, but in real AI infrastructure the binding constraint is often the boring layer underneath: data.
In this episode, Kim Isenberg sits down with Sven Breuner of VAST Data to unpack why costly GPUs so often sit idle waiting on storage, how AI workloads differ from classic high-performance computing, and what a genuine "AI operating system" and "AI factory" actually demand. It is a grounded tour of the least glamorous, most decisive part of the stack, and a fitting reality check on the thread running through today's issue: compute alone was never the whole story. Superintelligence Exclusive interview.
(Google Cloud CEO Thomas Kurian and Nvidia CEO Jensen Huang; NVIDIA)
"[We have] a multi-decade partnership with Nvidiaโฆ GPUs are the most popular chips on Google Cloud, used by our mutual customers, ranging from large enterprises like Toyota to cutting-edge startups like Thinking Machines."
The decade-old rivalry between OpenAI's Sam Altman and Anthropic's Dario Amodei keeps finding new stages.
After a year that reportedly included a Super Bowl ad war, a frosty non-handshake at the India AI summit, and Anthropic's standoff with the Pentagon over safety "red lines," the two now appear to be circling each other on the world stage. At a recent G7 gathering, reports suggest Amodei urged democracies to cooperate on AI and, unusually, drew public backing from Altman, even as both firms race toward blockbuster IPOs and trade jabs over safety, pricing, and government contracts. The feud, it seems, has matured from personal to structural, and it shows little sign of cooling.
The Forecasters Who Warned About 2027 Now Want to Slow It Down
The Takeaway
๐ The AI Futures Project, the team behind the viral "AI 2027" forecast, has published "AI 2040: Plan A," a roughly 100-page proposal to deliberately delay superintelligence by about a decade.
๐ The core mechanism is a US-China agreement by 2029 on full AI research transparency, backed by verification and what the authors call "mutually assured compute destruction."
๐ It maps five options, from Plan A (verified slowdown) to Plan D (keep racing to about 2030) and Plan S (shut it all down), and argues A is the least-bad path.
๐ The authors give Plan A only a 3 to 15% chance of actually happening, which is the point: they wanted a concrete alternative to a race they fear ends in extinction or locked-in power.
The most striking thing about "AI 2040: Plan A" is who wrote it: the same forecasters whose "AI 2027" scenario spent last year convincing people superintelligence was arriving fast now argue the wisest move is to slow it down on purpose. Newly published by the AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo, Plan A is a recommendation, not a prediction: a detailed picture of how the US, China, and the rest of the world could reach superintelligence around 2040 instead of racing there by 2030. Its premise is stark. Left to compete in secret, labs and nations build systems they cannot control or fully understand, and the default ending is "extinction or irreversible concentration of power."
(AI Futures Project)
The plan's spine is a deal. By 2029, the US and China would agree to make frontier AI research transparent to each other and to independent monitors, so no one can secretly sprint ahead. Enforcing that means watching the one thing you cannot hide, the giant compute clusters that train the models, which is where "mutually assured compute destruction" comes in: the credible ability to disable a rival's data centers if it breaks the deal, an AI-era echo of Cold War deterrence. Under Plan A, expert-level systems arrive around 2035 and superintelligence around 2040, with humans kept in control at each step.
(AI Futures Project)
The authors are refreshingly candid that this is a long shot, putting the odds of adoption at 3 to 15% and even sketching the awkward middle years where AI systems are "adversarially misaligned but controlled." A skeptic will point at the obvious: persuading Washington and Beijing to open their most strategic labs to each other, and to treat mutual compute strikes as stabilizing rather than provocative, is a staggering ask. But landing on a day when Oracle is cratering and SpaceX is sinking, it reaches a market already wondering aloud whether the all-out build-out is wise.
Why it matters: "Slow down" is finally a detailed plan instead of a slogan. By spelling out the treaty, the verification, and the failure modes, the AI 2027 team has pushed the debate past "should we?" to "how, exactly?", handing policymakers something concrete to argue with.
Sources:
๐ https://ai-2040.com
๐ AI Futures Project: AI 2040 Plan A
๐ The original AI 2027 forecast
Cut Lead Review From Hours To Minutes
Sign up for a free trial of Attio, the agentic CRM.
Ask Attio to build a daily workflow that surfaces the deals that need your attention today, like anything with a stage change, a recent reply, or a new signal in the last 24 hours.
Review your pipeline in Claude, synced live from Attio via MCP.
That's it.
The chart: METR's "time horizons" measures the length of task, in how long a human would need, that an AI can finish at a 50% success rate, plotted against each model's release date on a log scale from about one second to four hours. The core trend (METR-HRS) climbs from a few seconds of autonomy in 2019 to over an hour by 2025, and METR now extends the same curve across domains: software (SWE-bench Verified, LiveCodeBench), reasoning (Mock AIME, GPQA Diamond, MATH), computer use (OSWorld, WebArena), and robotics (RLBench).
The lesson: The frontier is rising fast, but very unevenly. Reasoning and coding agents already handle tasks that would take a person 15 minutes to over an hour, while computer-use and robotics benchmarks still sit near the bottom, at seconds to a few minutes. Autonomy grows first where feedback is fast and easy to check, and lags where the world is messy, exactly the gap today's Featured plan and Daily Feature both circle.
The caveat: A "50% success rate" is a coin flip, not reliability; a system that finishes an hour-long task half the time is not one you would leave unattended. And each line covers different tasks over different spans, so the chart shows direction and spread, not a like-for-like ranking of models.
๐ค The Robot That Practices Until It Gets It Right
โก Bottom line: London startup Humanoid says its new KinetIQ Ascend lets robots learn factory tasks by trial and error, not months of hand-tuning.
๐ก Why it matters: It targets 99.9% reliability at human speed, the unglamorous bar that decides whether humanoids are a demo or a workforce.
๐ What it means: Reinforcement learning is quietly moving from game scores to real assembly lines, where a few points of reliability are worth millions.
For years, humanoid robots have been better at demos than at day jobs. The hard part is not walking or waving; it is doing the same fiddly task thousands of times without dropping, jamming, or fumbling. On July 5, London-based Humanoid (founded in 2024 by Artem Sokolov, now more than 250 engineers) unveiled KinetIQ Ascend, a reinforcement-learning system built to close exactly that gap.
(Humanoid, via The Robot Report)
Reinforcement learning is learning by trial and error: the robot attempts a task, earns a reward when it succeeds, and adjusts. Instead of engineers spending months collecting data and tuning each new job, KinetIQ Ascend starts a robot with rough behavior and lets it practice on the real task until it is deployment-ready, what Humanoid calls a "capability factory." The reported results are concrete: on a machine-feeding job, throughput rose 42% and the robot ran at 1.5x human demonstration speed; on a bin-picking handoff, success climbed from 80% to 98%; and on a two-handed tote task, failures fell about twentyfold, from roughly 22% to 1%.
(Humanoid)
The honest caveats: these are company-reported numbers on chosen tasks, not an independent benchmark, and "human speed" at one station is not a full shift. But the direction matches today's METR chart, where robotics is the slowest domain to gain autonomy, so any method that reliably grinds real-world success rates upward is close to the whole game. It also rhymes with Richard Sutton's bet in today's news: that capable machines come from doing, rather than from reading alone.
(Humanoid)
See the whole platform. No guided tour.
Skip the sales call. Walk through Gladly's interface yourself โ the AI suggestions, the unified customer view, the full conversation thread. 15 minutes, no installation, no commitment.
