Summary
A quiet morning for curated signal (no @bayesiansapien retweets in the lookback) and a loud one for Google's CVPR 2026 booth, which dominates the AI-handle feed. @GoogleResearch ran a nine-post demo blitz from Denver: the standout for this wiki is BlazeEdit, a 195M-parameter on-device image-editing diffusion model that drops text conditioning entirely and does a full inference pass in 290ms on a Pixel 10, a genuine efficiency result inside an otherwise vision-heavy lineup (TIPSv2 patch-text encoder, D4RT 4D reconstruction which won a CVPR Best Paper, Ego2Web egocentric web-agent benchmark, an enterprise agentic-RAG framework, Android XR). The morning's real-money story is compute and policy, surfaced by @ns123abc: Trump says he has spoken to OpenAI, Anthropic and SpaceX about a government stake and is "not that far apart" from Bernie Sanders's 50% public-private proposal. On the tooling side, Anthropic's @bcherny announced Claude Cowork usage limits doubled for a month, @cursor_ai shipped Design Mode (point/draw/talk to edit a running UI), and @kilocode kept hammering the post-June-1 GitHub Copilot usage-billing shift while noting Nemotron 3 Ultra is already topping its OpenRouter leaderboard. @nvidia spotlighted Sarvam AI's sovereign stack (100B+ MoE on 4,096+ H100s serving 1.4B Indians). The rest of the feed is personal/political noise (@brivael's French-politics and Elon-worship thread, @Scobleizer's Tesla FSD and Boardy AI anecdotes) and is skipped.
Posts
Google CVPR 2026 demo blitz (cluster of 7) (@GoogleResearch). Google ran its CVPR booth as a research showcase, and several items are substantive. BlazeEdit (arxiv 2605.28067) is the efficiency highlight: a generalist image-to-image diffusion model that consolidates object removal, outpainting, relighting, tone correction and sticker generation into a single 195M-parameter model by eliminating text-conditioning, completing a full pass in 290ms on a Pixel 10 for private on-device editing. D4RT (d4rt-paper.github.io, CVPR Best Paper) is a single feedforward transformer that jointly infers depth, spatio-temporal correspondence, and camera parameters from one video, with a querying mechanism that probes any 3D point in space-time instead of dense per-frame decoding. TIPSv2 (gdm-tipsv2.github.io) is a vision-language encoder whose headline finding is that distillation unlocks better patch-text alignment than standard pretraining, so distilled students beat their larger teachers at dense alignment, a distillation result that echoes the wiki's knowledge-distillation thread. Ego2Web is a benchmark requiring web agents to ground evidence from egocentric video before acting online. Plus an enterprise agentic-RAG framework (iterative multi-agent search before generation) and Android XR / Gemini-on-glasses demos. The vision/3D items are largely Tier 3 here; BlazeEdit and the TIPSv2 distillation finding are the parts worth keeping.
Trump in talks to nationalize AI labs (@ns123abc). @ns123abc relays Trump saying "the American public essentially becomes a partner with the AI companies," that he has spoken to "ALL of them" (OpenAI, Anthropic, SpaceX) and they are "all coming to the White House next week," and that on Bernie Sanders's proposed 50% public-private partnership "we're not that far apart." This is the social-feed leading edge of today's digest Industry Pulse item on OpenAI negotiating a direct US government equity stake via a proposed "Public Wealth Fund." See the daily digest Global View for the compute-becomes-sovereign-finance synthesis.
Claude Cowork usage limits doubled (@bcherny). Anthropic's Boris Cherny announced Claude Cowork usage limits are doubled for the next month (applied to 5-hour rate limits), pitching it for work "too big for a chat": research across dozens of accounts, recurring reports, inbox triage. A demand-pull move the same week the broader industry is repricing AI-coding consumption (see the Kilo/Copilot item below).
Cursor ships Design Mode (@cursor_ai, blog). Cursor's Design Mode lets you click an element, draw on the page, or describe a change by voice from the Cursor browser, and the agent gets the selected element, its code, and the surrounding layout as context for the edit, tightening the notice-to-edit loop for UI work. Co-founder @mntruell framed the broader bet: working with agents should feel like working with a colleague, gesturing at a shared screen and talking live, not just typing chat. A concrete instance of multimodal, spatial agent interfaces.
The Copilot bill came due, and Nemotron 3 Ultra tops Kilo's leaderboard (cluster of 2) (@kilocode, blog). Kilo kept pressing its thesis that fixed-fee AI subscriptions were never sustainable: GitHub Copilot moved to usage-based "access-plus-consumption" billing on June 1, and the scarier datapoint is that Uber reportedly burned its entire 2026 AI-coding budget by April, mostly on Claude Code and Cursor. Kilo's pitch is model-portability (500+ models, bring-your-own-keys, swap when economics change). Separately, Kilo notes NVIDIA's Nemotron 3 Ultra (the 550B/55B-active open-weights model from this week) is already topping its OpenRouter usage leaderboard, a real-adoption signal for the open frontier model.
NVIDIA on Sarvam AI's sovereign stack (@nvidia, story). NVIDIA spotlighted Sarvam AI building a "Made in India" full-stack platform: 100B+ parameter MoE models trained across 4,096+ H100 GPUs, delivering millisecond multilingual voice inference for Aadhaar and powering KYC/sales/support over telephony and WhatsApp for Tata Capital and Infosys, serving 1.4B people in their own languages. A concrete sovereign-AI-at-population-scale deployment, the demand side of the compute-scarcity story running through today's digest.
AI worm carries its own LLM (@Scobleizer, Schneier). Scobleizer surfaced Schneier's note on a prototype AI-powered internet worm that carries its own LLM and runs it on machines it has broken into, the closest yet to Brunner's 1975 conception of a computer worm. Relevant to today's agent-security theme alongside the digest's SABER coverage.
Scobleizer personal/product anecdotes (cluster, light) (@Scobleizer). Tesla FSD enthusiasm (163,000 miles), the Boardy AI networking agent ("Her for founders"), AI glasses from Chinese vendors. Color, not signal. Noted and skipped.
Skip: @brivael (French political commentary, Elon/Milei/Thiel reposts, no AI research substance), @WHFraudTF and @SeanParnellASW (US government fraud/policy, not AI), @spencerpratt and @magicsilicon one-liners.