on "Pacing the Frontier" (My view on the state of play.) Internally, Anthropic and OpenAI (and emergingly SpaceXai) are 1 to 2 generations ahead of what's available to external parties. They are banking a lead against the openweight model providers as well as the fine-tuners developing for specific verticals. Recursive self-improvement is a myth, but having access to wildly performant internal tooling, plus being able to access that tooling at cash cost rather than paying margin on top, gives their staff accelerating advantage versus the rest of the market. Hence this year's wild performance gains, many of which delivered to customers. The recent series of model releases and optimizations that have accelerated cost-performance improvements to 99.99% per year were partly catalyzed by OpenAI trying to reel in Anthropic commercially but are also driving continued market expansion. Demand is skyrocketing, the frontier lab companies see amazing margins since they are forced to price against a scarce compute resource. Additional releases could drive even better margins, but the capability of this next set of systems could well result in accidental botnet infestations being unleashed by unwitting customers. Such an occurrence would impose huge political as well as civic liability on the closed labs, so they need to co-agree not to jockey for position as they clean up that liability. If botnet infestations are an inevitability, much much better that the first notable large-scale attack comes from an openweight model. Better still if the closed labs then have the tool available that can clean it up and help to protect the rest of the world's vulnerable infrastructure. If, on the other hand, any of the US labs are the cause of the economically costly disruption, it will be very difficult for even a more blameless US lab to charge for the solution without facing severe political (and potentially regulatory) blowback. They've built a lead, better now to let their common enemies cross the dangerous threshold first while internally compounding their own advantage against the rest of the field. Moreover, the safe solution that they propose--the tool to fix the botnet infestations--will be an AI tool paired with an AI monitoring tool that will probably have to be run on their own infrastructure. I could imagine zero-data-retention-pledging mechanisms by which they could try to launch such a product but it would certainly be more expensive--also possible that palantir could provide more enterprise-friendly monitoring attachments, but those will almost certainly be less performant than vertically integrated solutions. In effect, just like the harness provided the closed labs with a higher degree of defensibility in terms of their business, the safety harness will similarly expand their entrenchedness. If there is some semi-co-opted regulatory apparatus that has to approve and monitor the safety harnesses, even more so! Net, I would expect that the blistering pace of cost declines will dampen somewhat. In 2025 we were experiencing performance doublings every 1.5 months. In 2026, on the agentic benchmarks, doublings were coming every 4 weeks(!) If every other doubling is now eaten up by safety investments we could move to a 2 month doubling pace, a paltry 98% annualized cost decline. Net the effective productivity delivery per GW could slow slightly, but considering the blistering pace of growth of the market overall, I don't think that it will that meaningfully impact the trajectory of corporate uptake and topline growth. If the botnet infestation concern does manifest, the closed lab providers will be extremely well set up to provide hardening and protection. Datacenters are far and away the most likely unintentional attack target though there could also be infestations of end-point infrastructure (your smart home stereo is burning more power because it has also been co-opted to run a VM for a little bot doing some sort of inane classification task at massive scale). (The more obviously politically salient targets are much more likely to be intentionally attacked rather than submit to unintentional infestation and so will almost certainly be performed on open weight models.) I do think that space supercompute will prove more robust to botnet parasitic drag than will terrestrial datacenters due to its modularity and infrastructural simplicity (you won't see a botnet penetrate a satellite system via its air-conditioning system; whereas for a terrestrial datacenter that sort of thing may be the most likely penetration vector.) I do think its wise for the closed model providers to seek a way to coordinate a slowdown, so they can let one of the openweight guys Wile E. Coyote off a cliff. On the margin, SpaceX is probably most advantaged, as they get a little more time to pull fully abreast at the frontier and likely have the more hardened infrastructure stack go-forward. Most regulatory outcomes would favor the incumbents with a lead, though by allowing them to pad their margins rather than drive top-line as aggressively. There remains a tail-risk regulatory clampdown outcome, which would be terrible for humanity, denying all of the benefits while exposing us to essentially the same level of risk. I don't think that is the likely case, however, in part because the tools are already proving so useful to so many.