Summary
The afternoon's strongest signal is hardware economics, not models. Jensen Huang argued directly that CUDA is what makes NVIDIA silicon fungible across generations, so a 2020 A100 stays a rentable, financeable asset through 2029, anchored to CoreWeave's 2029 A100 commitment (a cluster of 2, with @ns123abc endorsing it). Second is M&A that lands squarely on the efficiency question: Anthropic is reportedly in talks to buy Decart, a world-model and inference-optimization startup, for roughly $6 billion, which is a frontier lab paying world-model prices partly for serving efficiency. Third is a pile of model-shop needling from @ns123abc, all of it worth 30 seconds and no more: DeepSeek shipped V4 Pro with no social announcement because it scores one point above V4 Flash, Reuters recycled April reporting on a Brin-led DeepMind recursive-self-improvement push, and an SSI test-time-training rumor he himself labels fake news. Everything else is thin. A solar-plus-battery datacenter power pitch at $100/MWh is the only other item touching this wiki's interests, and roughly 35 of the 64 captured tweets are French and US domestic politics from two handles with zero AI content. No curated reposts landed in this window.
Posts
Jensen Huang on GPU fungibility: the A100 is a 2020 chip that stays productive to 2029 (cluster of 2, @JensenHuang · @ns123abc · Business Insider). The argument chains four claims: CUDA gives one platform across Ampere, Hopper and Blackwell, that versatility makes compute fungible, fungibility drives utilization, and utilization makes GPUs "rentable, durable and financeable." It is a depreciation-schedule argument dressed as a software argument, and it lands two days after NVIDIA's $500B financing platform, which needs exactly this claim to hold. See NVIDIA turns AI compute into an asset class.
Anthropic in talks to acquire Decart for ~$6B (@Scobleizer, relaying @shiringhaffary · Bloomberg). Decart is described as a world-model and inference-optimization company, and the second half of that description is the part that matters here. If the price is real, a frontier lab is valuing a serving-efficiency team at world-model multiples, which is the clearest market signal yet that inference cost is a moat rather than a line item.
DeepSeek quietly shipped V4 Pro, one point above V4 Flash (cluster of 2, @ns123abc · follow-up). His claim is that DeepSeek made no announcement anywhere on social because the gap is embarrassing, and the follow-up is him defending the take against pushback from other open-model boosters. Unverified and snide, but a one-point Pro-over-Flash delta is a real thing to check against the V4 architecture page.
Grok 4.6 pricing restated: $2/M in, $6/M out (@ns123abc). A carryover of the morning's dominant story, with the cost claim stated hardest: 85% cheaper than the comparison point, beating GPT-5.6 Sol on knowledge work and tying Fable 5 Max. Numbers are the poster's, not a lab card. Full treatment in the Grok 4.6 summary.
SSI test-time-training rumor, flagged as fake news by the person relaying it (@ns123abc). The rumored claim is a small reasoning engine competing with much larger training runs via better meta-learning-oriented data curation, plus a shift to test-time training where gradient descent replaces the context window. Treat it as vapor, but note the direction is the same one the continual-learning market is already pricing.
Reuters "new" DeepMind recursive-self-improvement story is April reporting (@ns123abc). The underlying content is a Brin-and-CTO-led strike team aimed at turning coding models into full AI researchers, focused on long-context coding tasks and training on Google's private codebase. The useful part is the debunk: the first line of the article dates it to April, so this is recycled, not news.
Ambrosia Energy: solar plus battery at $100/MWh, 99.9% capacity factor, 12 months to power (@Scobleizer, relaying @FutureJurvetson · TechCrunch · ambrosia.energy). Two SpaceX alumni pitching off-grid continuous renewable power to datacenters, built faster than a gas plant, with a claimed 12x faster proprietary solar install as the actual differentiator. If contract-to-power really lands at 12 months, the binding constraint on buildout shifts from interconnect queues to install throughput, which is the cost-per-megawatt layer of inference as energy-to-token production.
X Square's WALL-B hits 1,816 parcels/hour at 98% accuracy (@Scobleizer, relaying @TheHumanoidHub). Roughly two seconds per parcel in a fully autonomous logistics livestream, about 45% faster than Figure's sub-three-second pace from May. Scoble's own caveat is the right one: it is a task-specific policy, not a generalized world model, and not a humanoid.
A clean taxonomy of why robots fail outside simulation (@Scobleizer, relaying @muskan_kalra24). Three named gaps: sim-to-real physics, the sensorimotor gap where robots get a much weaker version of pressure, texture and friction feedback than humans, and action representation, turning understanding into motor commands. Outside this wiki's attention range, but the framing is unusually crisp for a thread.
"Stop trying to control AI, we are growing these things" (@Scobleizer, relaying @cxgonzalez). The argument is that coercive framings in safety work will backfire as models get more lifelike, with a registered prediction that alignment researchers will rederive parenting from first principles. Pure vibes with no concrete claim attached, so it goes in the pile with the rest of the anthropomorphism discourse rather than into responsible-ai.
China's industrial capacity as the real war variable, and the yuan at a 3.5-year high (cluster of 2, @MarioNawfal · yuan). The first thread argues a prolonged conflict favors China on manufacturing capacity, rare-earth control and supply-chain depth rather than on cost, and the second notes the yuan hitting its strongest level since February 2023 despite tariffs and export bans. Tangential, but rare earths and supply-chain leverage are the semiconductor story's upstream.
Lex Fridman's Khabib episode, dubbed with human plus AI translation (cluster of 2, @lexfridman · Russian version). Recorded entirely in Russian and shipped with both audio tracks and subtitles. Only the production pipeline is on-topic, and only barely.
DHH ships Omarchy Quattro and the Aether theme builder (cluster of 3, @dhh). A 56-second install on a $699 Dell XPS 13, plus a tool for generating themes from a background image. Continuation of the morning's Omarchy thread, no new AI content.
Grok 4.5 usage "563% above average" on Design Arena, and Tesla ranked #1 in brand loyalty (cluster of 2, @MarioNawfal · Tesla). Both are agency-written promo posts with credited designers. Skip.
Off-topic political and personal volume (cluster of roughly 35: @brivael ×19, @MarioNawfal remainder, @spencerpratt, @SeanParnellASW, @JonasBadalic). French domestic politics, Iran and Hormuz coverage, Netflix documentaries, a Paramount merger lawsuit, and two tweets about moka pots. No AI signal. Skip.