Akshobya

Anthropic now plans to push for govts to require firmware level locks on Western GPUs that prevent "unapproved" (Chinese open-source) models from being run on them, in the name of "AI safety" and "national security." AMD/Google are on board, only Jensen opposed it 1/

A new breakthrough in Local AI I trained a DSpine speculative decoder for GLM-5.3-Flash on 4x DGX Sparks over a mere 8 hours. About a 20% boost so far, I will keep training it and it'll be the best GLM spec decoder out there in a week

My cold take on the Huangs Affair (pun intended) is that the MSS has played this near-optimally, better than Beijing writ large, and not just accidentally-in-retrospect. Here is my thinking. The Chinese are calibrating their moves to facts on the ground. 2024-2025 was an investigation into just how much the US is willing to budge on export controls under Trump, who is a consistently transactional merchant, a deal junkie and haggler respecting backbone and brute force above all. To this end they've retaliated harshly to Liberation Day, refused H20s and B30s, responded to BIS 50% with the MOFCOM threat and to Nexperia seizure with cutting off European supply, secured an eventual Busan deal (Oct 2025), and leaked news on their EUV light source progress (made public by Reuters in Dec). Then Trump went to Beijing with Jensen (China-sympathetic and desperately interested in getting back into their market before it closes forever) at his side and industrial/finance CEOs in tow; and Xi went to Washington; and so on and so forth. The Chinese have achieved the maximum rapproachment possible under the most cooperative administration they could hope for. And all they've got for it was a permit for max 750K H200s per year, 75K per entity (about $17B worth of compute at alleged prices), under the pretext of marginally more tariff bux for Trump; plus some lenience with regard to overseas buildouts or small-scale smuggling. Hawks catastrophised about that with bullshit charts where those chips would be strategically significant, but really, this was a polite rejection of Chinese advances. Huawei can make H200-grade systems themselves; on a cluster level, 950DTs are arguably better. It's worth at most a few months of their current throughput. It wouldn't have moved the needle or rescued then from crippling supply scarcity. And by March, I think it also became clear to Beijing that the GOP will most likely crushingly lose the midterms, having abandoned domestic control in favor of Israeli geopolitical needs; it wasn't then clear to me, but they must have had better intel about Iranian resilience and Pentagon's irrationality. Thus, Democrats would have come back. And Democrats are a) negatively polarized against the AI-pilled tech right, of which Jensen is the kingmaker and Trump the patron; b) capable of legitimately coordinating international effort against Unfree Countries; c) dogmatically committed to "this is what annihilation looks like!!" turbo-aggression against Chinese development, even though the current China, unlike the China of 2018 or even 2022, is very capable of retaliation. Plus they have rapport with the doomer/EA community and the Dario side of export control advocacy. Trump can still rule by EOs, but he'll have to ration this power, and he has many larger priorities. Additionally, technical arguments for resuming compute trade have been unraveling, chiefly the idea that availability of Nvidia can strangle internal R&D and keep the Chinese "addicted" or dependent, thus subject to future throttling and treats and more afraid of confrontation. I used to argue this too, because I didn't observe a near-term positive feedback loop: it was too costly for any top-tier Chinese group to switch away from Nvidia, scarce though it is; the cost measured in months of stagnation in a very cutthroat and neurotic market. The CCP, contra American fantasies about planned economy, seemed unwilling or incapable to address this strategic death trap, and mere subsidies wouldn't have sufficed. A big factor here is, again, DeepSeek. I am still not sure how legit the rumors about the failed R2-on-Ascend in 2025 were (logicallt speaking, there never was the R2 project), but much of late 2025-early 2026 was indeed dedicated to transition, and in the end the Whale has provided a nucleus for crystallization of the Chinese ecosystem, building the default Huawei-TileLang stack (as Wenfeng intended to do from the start). They create captive demand and feedback for Huawei's roadmap, which I now expect to accelerate and pull up second-fiddle players too. I'm not sure if H200s stranded in Hong Kong could have changed any of it. Given how successfully DeepSeek had been using CUDA until now, it is possible that unrestricted sales would have postponed this evolution for another year or so. But they were restricted anyway, and extra 75K legal H200s are just… not enough for what Whale intends to do. So, on balance, Washington had made this an inevitability. Finally, Nvidia's perceived importance for the American ecosystem has been undermined by the very progress in AI. OpenAI can just up and make Jalapeno that is at the very least competitive with Blackwell, at Gen 1; Anthropic will also design a great chip; TPUs are awesome, Meta and Amazon are improving. Broadcom et al. are enough to maintain the supply chain integrity. AIs can write superhuman kernels (Wenfeng also makes this point) and at this rate the CUDA moat will not last except for bespoke hobbyist experiments (every serious lab is now capable of building a better stack for their large scale workloads). On top of it all, Rubin ramp is slower than expected, troubled, and worryingly dependent on Chinese suppliers. And it is important to realize that Rubio's memo on Huangs had been drafted well over a year ago. Things don't just "take a while", they take as long as it takes for the action to become expedient. Under the trajectory of the first two years of 47th, the Huang operation was relatively costless. By mid-2026, the writing was on the wall that the crackdown is coming before the next generation of Nvidia hardware is broadly available. What, then, was the rational move? I think it was to keep milking Jensen, collecting intel, letting companies utilize overseas compute for current opex, facilitating smuggling of components that can't be produced at scale domestically and systems that actually make some qualitative difference (NVL72 Blackwells right now), and rapidly developing domestic supply chain. I think the MSS played this very well. It's probably still not enough, but I'm not sure they could have done much better.

Introducing Inkling 2 Small. Scoring 31 on the AA Intelligence Index, it's a big jump over its predecessor. Inkling 2 Small is available for fine-tuning and inference on Tinker today, with full weight release planned soon. thinkingmachines.ai/news/ink…

Sir Demis Hassabis talking to Reid Hoffman about RSI at the Masters of Scale Summit:

PSA for those of u vibecoding naked on your laptop with no containers: tell your agents to DELETE Playwright NOW. Even the latest (1.64.1) bundles early Chromium.156 with CVE-2026-118306 CVE-2026-118331 CVE-2026-118348 CVE-2026-118344 And the Playwright SANDBOX IS OFF BY DEFAULT