PDS

On the AI regulation panic. My 2 cents.. Capabilities and diffusion have already decoupled. The X community's reaction to Dario et al. calling for paced capabilities is arguing over a settled fact. This decoupling was inevitable, whether enforced by government action (Fable 5 getting export controlled) or by coordinated pacing from the private labs as Dario proposed today. Dangerous capability is entangled with general capability -> cyber is the most grindable dangerous domain and you can't strip it out surgically, so every frontier release re-opens the same gate. Whether that gate is Washington or a lab consortium is a governance detail, not a change in regime. So stop freaking out. This isn't your problem to solve. This is a lab-layer problem with a lab-layer solution. Not an ecosystem problem. Take a step back and think why the labs tend to lead the safety narrative: they are overweighting the capabilities they observe in verifiable domains. RLVR compounds only where outcomes are dense, replayable and gradeable, where the base model already puts mass on the solution and you can grind rollouts in parallel. This is a narrow ecosystem: Code, math and cyber. If you use frontier models to code, you probably feel these are AGI-like machines. Same for math. Same for cyber. But if you use frontier models to work through non-verifiable domains and you are at the frontier of your field, you appreciate how limited they are and how much work is ahead before they are ready for RSI. Non-stationary, reset-free domains, like markets, have sparse, delayed, self-erasing outcomes. Sample-efficient continual learning is unsolved, and scaling context doesn't substitute for it. So for all those claims arguing AGI is here (including the goat Jensen), I welcome you to show us where we reached general intelligence in macro investing. Or life sciences. Or anything reset-free. The labs and frontier users are sampling different regions of the same capability surface. Labs live in the grindable region and extrapolate; domain experts live in the ungrindable one and see the ceiling. Both observations are correct. Only one generalizes to RSI, and it isn't the labs'. So the pessimism from the labs makes sense BUT also feels trivial to anyone in the real world. Capability-diffusion decoupling is the natural state of the diffusion regime. It shouldn't be a point of contention. Labs pace their development so models don't become superhuman hackers. The rest of us keep pushing frontier work in domains where the models are still nowhere near that. Tip -> don’t freak out when the hyperscalers decide to pace themselves too. That is the most bullish market outcome out there Keep grinding…

Andrew, you're sampling the wrong ecosystem. At the frontier, across every industry, demand for tokens and compute is insatiable. If a company is rationing, it's because usage blew through budgets or they lack the talent to build with it and profit from it. Surprised to see you shorting American innovation 🚀

This is the way

In a capabilities-diffusion decoupling regime what matters to markets and the AI ecosystem is speed of diffusion not whether Anthropic releases a new generation Mythos a week or 6 month later.

Elon is right

Count me as someone who religiously mines Gavin’s podcasts for where to point research hours. Which is why this one is worth pushing on. As an AI-pilled investor and builder, I want the industry’s leading voices held to high and transparent sourcing standards. “I have been told by multiple people I trust” isn’t a source. It’s unverifiable secondhand / gossip. “Multiple” carries the entire credibility load here. Were they independent? In an ecosystem this small, several sources routinely trace back to one origin event. Are these sources unbiased? Anyone who wants AI to diffuse should want its leading voices held to a high standard. In an era this fluid, this stuff makes it into policy.

It's been a while — so a long one, two subjects. With July on the books, worth looking back at the AI momentum crash and why Knightian uncertainty explains a cycle where fundamentals came second to market structure. Knightian uncertainty is when you can't assign probabilities to future states because you can't enumerate the states. No amount of additional data resolves it. To me there's no better description of what the AI complex just went through. June's export-control episode on Fable 5 seeded a dense cluster of disputed stories -> delay on OAI 5.6 release, agent containment failures, intensifying open-weight and China debates (K3), petitions to slow AI development pace, etc. Macro compounded it - Fed hike fears, Iran. Not one of those items moved a single cash flow of AI capex beneficiaries, but every one of them widened the set of end-states you could plausibly imagine. And none could be argued away. The AI complex trades on a blend of near-term cash flows, which skeptics read as unsustainable, and a belief about whether capability rents are capturable at all. That belief has no verifiable anchor outside of leading AI labs + cloud, and native-AI efforts are still nascent. The open question for the skeptics isn't when that revenue base shows up, it's whether it shows up, and at what capability level it gets unlocked. That's the mechanism. When the state space widens and nothing narrows it back, fundamentals take a step aside. They aren't wrong, they're just mute. And when fundamentals are muted, the marginal price-setter becomes a risk limit. In my view, one of those growing narratives carried more weight than the rest, because it went straight at the paradigm: if frontier capability is commoditizing, the capex thesis has no terminal owner - capability gets built, rents get competed away, nobody books the return. BUT this narrative rests on an incomplete argument, and mostly on a chart. I see Pareto frontier charts left and right making the point that open models are catching up to OAI and Anthropic. I get the excitement, I'm bullish on what NVIDIA is doing with Nemotron. But the discussion is tilted by voices of leading investors and analysts rather than builders, and the labs aren't ground truth here either. They optimize for training the next model, not for living inside the last one across long-horizon production loops. Quite different observation channel. Here is my issue with the Pareto frontier charts. The frontier measures the diffusible component of capability with a statistic that hides reliability, on axes chosen by the party with an interest in what's shown, in the domains where convergence was structurally guaranteed (code, math, cyber). Naive readers treat frontier proximity as substitutability. It's better read as a lower bound on what's been made public, over a subspace selected for fast convergence. What gets measured is what gets copied. A benchmark turns a fuzzy capability into a scoreable target, and once there's a score, anyone can train against it, distill toward it, and verify they got there. That's what collapses replication cost. So a published benchmark isn't a map of capability -> it's an announcement of which capability is about to get cheap. So my critique for the Pareto frontier aficionados is to focus less on the scoreable target and more on co-adaptation. That's the better way to support the AI narrative. Capability is delivered by a PRODUCT: model, training environment, runtime harness — co-adapted. The environment a model is rewarded in is the harness it's deployed in. Weights carry priors about tool schemas, failure recovery, when to stop searching, etc. Bolt an open agent stack onto open weights and you get a model reasoning about a harness it was never rewarded in. The flagship Pareto frontier chart, Artificial Analysis Intelligence Index vs cost, is a composite of nine evals, and it fails on both halves for different reasons. The static half (GPQA, HLE, SciCode) is the grindable subspace where convergence was guaranteed. The agentic half (GDPval, τ³-Banking, Terminal-Bench) runs every model through the same third-party harness, which is exactly what strips out co-adaptation. A closed model scored in someone else's scaffold is being measured with its main advantage switched off. Then all nine get averaged into a single number, which turns reliability into a mean of means. And the chart draws a tradeoff — pay less, get less. But building isn't a tradeoff, it's a threshold. Below the bar the output isn't worse, it's unusable, and the savings buy you nothing. A model five times cheaper that hallucinates graph edges via semantic retrieval isn't a cheaper option. It's not an option. The open side is building directly at this - OpenEnv, and OpenForgeRL out of MSR ~10 days ago - training open models inside the harnesses they actually run in. Early, not yet decisive, and the thing I'd watch. I think the quiet shift here is from capex beneficiaries to capability compounders. And I don't just mean the labs. If you run a real deployment, you own something the labs don't: the failure taxonomy of your own domain. That's what specifies the environment you train against, and it's what turns a rented model into one shaped to your workload. The moat isn't the weights. It's knowing which environment to build — and that only comes from running the thing. Analysts read the chart. If you are building, you read the failure logs. LET'S GO

Demis is the 🐐 but he keeps calling AGI is around the corner yet models still don’t generalize in non-verifiable domains. If RLVR is our way to AGI (which appears to be for OAI and Anthropic), the AI race is a marathon not a sprint.

💯

Dwarkesh’s “AIs learning on the job” sharpened something I keep circling: progress now gates on grindability, not just verifiability. A domain advances only if you can stand up deterministic, replayable simulators and run rollouts in parallel - code, math and cyber qualify. Reset-free, non-stationary domains like markets, judgment or open-ended work, don’t, so that’s where models likely stay non-AGI convergent. Buried underneath stale bottleneck market narratives, two inevitable paradigm shifts are surfacing. The capability wall: RLVR compounds where outcomes are dense and replayable, flat where they’re sparse. Blowing up the KV cache to hold ever-growing domain-specific knowledge doesn’t fix the underlying weakness, the lack of sample-efficient continual learning. No surprise Fable 5 is this good at coding and cyber, that’s where the gradient lands directly, while everything else improves only second-order through transfer, and not where the reward is reflexive. This takes me to the second paradigm shift and one I heard treated lightly at a CIO call today. Diffusion: permissionless broad release is now negotiated with DC, precisely because the grindable capabilities are the ones on the path to AGI. Individual models get whitelisted, Fable likely comes back, but you can’t surgically strip capabilities out of a model, and they’ll keep getting better at exactly these. So expect the Trump EO to evolve from voluntary review into a real throttle on capability diffusion -> breaking the capability/diffusion coupling that’s carried us this far. @dwarkesh_sp

This is one way to flatten the exponential

Interesting -> .html vs .md