Friday 29 May 2026
A monster funding day in AI land. Cognition and Anthropic both announced massive raises, Claude Opus 4.8 dropped, and the protein biology crowd got an open atlas of 6.8 billion sequences. Meanwhile, Tesla’s breaking ground on a robot factory that plans to churn out 27,000 Optimus units a day. Business as usual in the singularity waiting room.
💰 Funding
Cognition raises over $1B at $26B valuation — Devin’s parent company is now the largest independent agent lab in AI, projecting >$1B ARR by year-end with $492M run-rate revenue. Enterprise usage up >10x YTD.
Anthropic raises $65B Series H at $965B valuation — Claude’s parent overtakes OpenAI as the world’s most valuable AI firm. The round underscores the vast sums still flowing into frontier AI, despite growing questions about sustainability.
🤖 Models and Launches
Claude Opus 4.8 — Anthropic ships its latest flagship model (812 HN comments — the internet has opinions).
BioHub releases world model of protein biology — ESMC, ESMFold2, and ESM Atlas: open tools for protein structure prediction and design across 6.8 billion sequences and 1.1 billion predicted structures. Freely available to researchers worldwide.
ElevenLabs Music v2 — Music generation model that can switch genres mid-track while maintaining vocal and compositional coherence.
Surya OCR 2 — 650M-parameter OCR model hitting 83.3% on OLMOCR bench, 87% across 91 languages, and 5 pages/s on an RTX 5090.
NVIDIA LocateAnything — Vision-language grounding framework that decodes bounding boxes in parallel, claiming 10x speedup for dense object detection.
🔬 Research
Sakana AI DiffusionBlocks — Reinterprets forward passes as diffusion-like denoising steps so deep nets can be trained one block at a time, dramatically reducing memory while matching end-to-end performance.
Delta Weight Sync in TRL — Transmit only changed parameters between RL steps, reducing async RL data transfer from gigabytes to megabytes via a Hugging Face Hub “bucket” approach.
DeepSWE benchmark — 113 tasks across 91 repos in 5 languages, requiring 5.5x more code and touching 7 files on average versus SWE-Bench Pro. Tests long-horizon, messy real-world coding workflows.
EAGLE 3.1 for speculative decoding — Stabilizes hidden-state feedback and reduces attention drift at deeper decode steps. Inference optimization is increasingly architectural, not just kernel-level.
🛠️ Agents and Tools
LangChain Deep Agents v0.6 — Delta Channels cut checkpoint storage for a 200-turn coding session from 5.3 GB to 129 MB. Also launched computer use in Fleet and Context Hub for versioned agent context.
Beyond the Prompt: Claude Code — Treat Claude Code as an autonomous agent that needs guardrails, not a chatbot. Tuning directory structure, skills, agents, plugins, and MCPs turns it into a teammate.
Trajectory launches with $15M — Platform for using product usage signals and agent traces to continuously post-train large agentic models. Design partners include Clay, Harvey, Decagon, and Mercor.
Ramp: 10,000 coding-agent sessions for security — Pointed ~10,000 Inspect coding-agent sessions at its backend in an 8-hour run with a minimal “find security issues” prompt. Found high-severity issues.
Building self-improving tax agents with Codex — OpenAI used Codex to build agents that self-improve by tracing reviewer corrections back into evals and fixes, resulting in AI that can prepare increasingly complex tax returns.
🏢 Industry
NVIDIA bets $150B on Taiwan — Jensen Huang is cementing Taiwan as the world’s AI manufacturing hub, building a new HQ to expand the TSMC partnership. US reshoring plans not going as Washington hoped.
DeepMind’s Hassabis: AGI is 3-4 years away — Google DeepMind CEO accelerates his timeline from 2030-35 to 2029-30.
Google expands Gemini for Business — Shareable Projects allow team members to collaborate within dedicated multi-surface workspaces.
Anthropic expanding Claude voice mode to 18 languages — Claude goes multilingual in voice, moving beyond English-only voice interactions.
YouTube expands automatic AI video labeling — Automatically applying labels to videos with significant photorealistic AI content, reducing reliance on creator self-disclosure.
🌍 Notable
Anthropic reaches $965B valuation, overtakes OpenAI — Guardian deep-dive on what the raise means for the AI industry’s power dynamics.
The SpaceX IPO and data centers in space — SpaceX’s $2 trillion valuation assumptions require Starship to work, sufficient chip supply, massive compute demand growth, and agentic inference to unbundle current architectures. Lots of ifs.
EU fines Temu €200M for allowing sale of illegal products — The EU cracks down on the Chinese marketplace. 206 HN comments debating enforcement vs. protectionism.
🔥 Hacker News
Claude Opus 4.8 (812 comments) — Anthropic’s latest flagship dominates the front page.
Disagreement among frontier LLMs on real-world fact-checks (333 comments) — Frontier models disagree more often than you’d think on factual claims.
Bricks and Minifigs stole a man’s $200k Lego collection (158 comments) — A cautionary tale about consignment and corporate overreach.
Just Use Postgres for Durable Workflows (74 comments) — You probably don’t need Temporal. Postgres is enough.
Dynamic Workflows in Claude Code (100 comments) — Anthropic adds dynamic workflow orchestration to Claude Code.
Continue? Y/N: A 60-second game about AI agent permission fatigue (93 comments) — A game that captures the absurdity of “Allow this action?” prompts.
bttf: a command line datetime Swiss army knife (87 comments) — From the author of ripgrep. Because date math on the CLI is still painful.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 6 newsletters processed, ~45 stories distilled