Wednesday 20 May 2026
Karpathy to Anthropic, Gemini 3.5 Flash drops, and Cursor trains its own model with SpaceXAI — a busy day before Google I/O.
🤖 Models
Qwen3.7 Preview Lands on Arena — Qwen3.7 Max Preview ranks 13th overall in Text Arena (#7 Math, #9 Expert, #10 Coding), while Qwen3.7 Plus hits 16th in Vision. Alibaba now sits at #6 lab in text and #5 in vision.
Gemini 3.5 Flash — Google’s new Flash model drops ahead of I/O. The Gemini app is also rolling out an “Extended” thinking level option for Fast and Gemini 3.1 Pro, plus new third-party app integrations (Canva, Instacart, OpenTable).
Cursor Composer 2.5 — Cursor’s strongest coding agent yet, trained with targeted RL and synthetic data. Bigger reveal: they’re training a much larger model from scratch with SpaceXAI using 10x more compute and access to Colossus 2’s million H100-equivalents.
NVIDIA Vera CPU Arrives at Top AI Labs — First Vera CPUs hand-delivered to Anthropic, OpenAI, SpaceXAI, and Oracle. Features 88 custom Olympus cores, 1.2 TB/s memory bandwidth, 50% faster per-core performance. Host processor for the Vera Rubin NVL72.
HRM-Text: 1B Model Trainable for ~$1,400 — A 1B text model based on the HRM architecture that needs 130-600x less compute than foundation models. The 0.6B version trains on 8 H100s in ~50 hours for ~$800.
🏢 Industry
Karpathy Joins Anthropic — The biggest talent move of the day. Andrej Karpathy announced he’s joined Anthropic. 414 HN comments and counting.
Musk Loses OpenAI Lawsuit — Jury dismissed all claims, finding Musk waited too long to file. OpenAI now has a clear path to a public listing. Musk plans to appeal.
Meta Reassigns 7,000 Workers to AI — Before mass layoffs, Meta is moving employees to four new AI-focused organizations with fewer managers and AI-native design structures. Capex expected to reach $135B this year.
Anthropic Acquires Stainless — Anthropic bought the SDK automation startup whose platform was used by OpenAI, Google, and Cloudflare. Signals continued vertical integration around developer ergonomics.
OpenAI Quietly Bought Weights.gg — Acquired the six-person voice-cloning team and its IP, then shut down the site and dispersed the team across OpenAI groups.
Google and Blackstone Create AI Cloud Company — New joint venture aims for 500MW of capacity online in 2027, scaling substantially from there.
🔬 Research
What Political Censorship Looks Like Inside an LLM’s Weights — Deep analysis of Qwen3.5-9B showing political censorship is a small circuit that can be read and turned off. The factual knowledge exists in pretraining; censorship is layered on top.
Lighthouse Attention from Nous Research — Selection-based hierarchical attention offering up to 17x faster forward and backward passes at large contexts, with 1.4-1.7x pretraining speedup.
LLM Architecture Developments: KV Sharing, MHC, Compressed Attention — Comprehensive survey of architecture tricks reducing KV-cache size, memory traffic, and attention cost as reasoning models keep more tokens around longer.
Notes on Pretraining Parallelisms and Failed Training Runs — Why training is such a precarious operation. Key culprits: breaking causality and adding bias.
Meta’s AIRA: Agentic Neural Architecture Discovery — Beats Llama 3.2 at 350M, 1B, and 3B scales within a 24-hour compute budget by splitting search into planning (AIRA-Compose) and implementation (AIRA-Design) agents.
🛠️ Agents and Tools
Codex to Control Other Desktop Devices via Computer Use — OpenAI working on letting Codex operate macOS apps even when a laptop is locked or asleep. Currently requires unlocked, awake sessions.
Claude Code in Large Codebases: Best Practices — Anthropic shares patterns for successful adoption in monorepos with millions of lines, legacy systems, and microservices across separate repos.
The 62.5-Minute Rule for Claude’s Cache — If you expect to need a cache before 62.5 minutes, refresh it. Otherwise let it expire. The decision point is the same regardless of cache size or model.
Devin Auto-Triage — Cognition launched an always-on “first responder” for bugs, alerts, and incidents with long-term memory, manager/subagent structure, and PR generation.
Headroom: Compress Agent Context Before It Hits the LLM — Compresses everything an agent reads before it reaches the LLM, producing the same answers at a fraction of the tokens.
Zero: A Systems Programming Language for Agents — Experimental language providing small native tools, explicit effects, predictable memory, and structured compiler output.
📊 Culture and Trends
The Last Six Months in LLMs in Five Minutes — Simon Willison’s rapid-fire recap. 545 HN comments — clearly resonated.
AI Eats the World — Big tech pouring ~$700B into AI capex this year. Foundation models commoditizing fast; real value shifting to applications, agents, and workflows. Adoption is wide but shallow.
Meta’s Giant AI Data Center Reshaping Rural Louisiana — 5 gigawatts of compute capacity, $200B+ in spending. The project has deeply entangled Meta in Louisiana’s politics, culture, and economy.
Samsung Union Strike: 48,000 Workers Threaten Walkout — 18-day strike threat amid fears of global memory chip shortages.
🔒 Security
CISA Admin Leaked AWS GovCloud Keys on GitHub — The US cybersecurity agency’s own admin exposed sensitive credentials. 146 HN comments.
314 npm Packages Compromised in Supply Chain Attack — “Mini Shai-Hulud Strikes Again” — another massive npm supply chain compromise. 271 HN comments.
Tesla’s Lithium Refinery Discharges 231,000 Gallons of Polluted Wastewater Daily — The Texas lithium refinery is dumping polluted wastewater at industrial scale. 154 HN comments.
🔥 Hacker News
Karpathy Joins Anthropic (414 comments) — The top story of the day.
The Last Six Months in LLMs in Five Minutes (545 comments) — Simon Willison’s rapid-fire LLM recap.
Apple Unveils New Accessibility Features (280 comments) — New features with Apple Intelligence integration.
Gemini 3.5 Flash (304 comments) — Google’s latest Flash model ahead of I/O.
I’ve Built a Virtual Museum with Nearly Every OS (105 comments) — Virtual OS museum with near-complete OS collection.
Show HN: Gaussian Splat of a Strawberry (179 comments) — 3D Gaussian splatting demo.
OpenBSD 7.9 (256 comments) — New OpenBSD release.
🌍 World and Australia
US “Locked and Loaded” on Iran as Talks Stall — VP Vance says US remains ready to restart military campaign; Iran warns of “new fronts.” Trump claims planned Tehran attack postponed for talks, but no deal imminent.
US Warns Russia After Moscow Threatens Latvia — “NATO membership will not protect you” threat from Moscow escalates as Latvia’s UN rep dismisses Russia’s “pure fiction and pure lies.”
CSIRO Job Cuts Will Gut Australia’s Climate Projections — Science agency planning to sack a third of the national climate modelling team, leaving Australia unable to contribute to global climate reports.
SpaceX Starship v3 Test Flight This Week — Major redesigns make it bigger, more powerful, and a step closer to full reusability. First test in seven months.
Magic Mushrooms Effective for Cocaine Addiction — Single-dose psilocybin participants more likely to abstain than placebo group.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 5 newsletters processed, ~45 stories distilled