Friday 22 May 2026
Google I/O dominated the cycle this week, with Gemini 3.5 Flash going GA, a new Omni video model, and Antigravity evolving into a full agent platform. OpenAI quietly disproved an 80-year-old math conjecture for under $1,000 in compute, Karpathy jumped to Anthropic, and the industry’s IPO and compute arms race continued to accelerate.
Models and Launches
Gemini 3.5 Flash is GA — Google shipped immediately at I/O, skipping preview. 1M-token context, 65k output, 4 thinking levels, 4x faster than comparable frontier models. Artificial Analysis: Intelligence Index 55 (+9 vs Gemini 3 Flash), 5.5x costlier. Arena #9 overall.
Gemini Omni: Multimodal Video Creation — New family merging Gemini reasoning with generative media. Text/image/audio/video in, video out. Rolling out in Gemini app and Flow for paid users, YouTube Shorts this week, APIs coming.
Cohere Command A+ Open Release — Apache 2.0, 218B MoE with 25B active parameters, 48 languages, runs on 2x H100s at W4A4. Unusual architecture: parallel transformer blocks, LayerNorm, large shared experts.
Stable Audio 3.0 — Open-weight models generating music and sound effects up to 6+ minutes.
Cerebras Running Kimi K2.6 — Trillion-parameter model hitting approximately 1,000 tokens/sec, fastest frontier inference ever measured by Artificial Analysis.
Funding, Compute, and IPOs
Anthropic-SpaceX $45B Compute Deal — $1.25B/month through May 2029 for capacity at Colossus 1 and a second data center. 90-day exit clause either side.
Anthropic Heading for First Profitable Quarter — Q2 revenue projected at $10.9B, more than doubling. Growing faster than Google or Facebook pre-IPO. May not stay profitable for the full year due to compute spending.
OpenAI Preparing for IPO — Filing in coming days/weeks, aiming for September listing. Still has to prove it can generate enough revenue to cover spending. Recently missed internal targets.
SpaceX IPO Filing and Cursor Acquisition — Planning June 12 Nasdaq listing under SPCX. Lost $4.9B last year on $18.7B revenue. Plans to acquire Cursor in July with a $10B breakup fee if the deal falls through.
Polymarket Launches Private Company Trading — Prediction markets for valuations, IPO timing, and secondary-market activity for companies like OpenAI and Anthropic.
Exa Raises $250M Series C at $2.2B Valuation — Positioning as a search lab organizing web data specifically for agents.
Industry
Karpathy Joins Anthropic — OpenAI co-founder and former Tesla exec joins Anthropic’s pretraining team. Will lead a new group using Claude to accelerate pretraining research.
Bezos’ Project Prometheus: $38B Startup — Building an “artificial general engineer” and next-gen CAD tools. $6.2B in funding, roughly 120 employees from OpenAI, DeepMind, Meta, and xAI.
Meta Lays Off 10% of Workforce — Zuckerberg warns “success isn’t a given” and assures no more cuts in 2026.
Spotify and Universal AI Remixes Deal — Licensing agreement lets subscribers use AI to create content on Spotify for the first time.
Alibaba Unveils Zhenwu M890 AI Chip — Designed for the heavy memory and communication demands of AI agent workloads. Part of China’s push for domestic chip alternatives.
Cheap AI Could Derail IPOs — American and Chinese labs producing cheaper models threaten OpenAI and Anthropic’s pricing power.
Research
OpenAI Disproves 80-Year-Old Erdos Problem — A general-purpose reasoning model (speculated GPT 5.6, under 32 hours, under $1,000) autonomously disproved the planar unit distance conjecture from 1946. 125-page proof. Timothy Gowers: “the first really clear example of AI solving a well-known open math problem.”
A Bitter Lesson for Data Filtering — Scaling studies show no data filtering may be optimal for large model pretraining. Large models tolerate and benefit from low-quality data when compute is ample.
InferenceBench: Agents Struggle with Infra — New benchmark for AI R&D automation. Frontier agents struggle with dependency management, underperforming simple vLLM/SGLang hyperparameter tuning. Inverse scaling effect observed.
MINTEval: Memory Systems Still Weak — Long-context memory evaluation: average accuracy only 27.9%, best 33.4%, despite 138.8k average token instances. Memory needs to be a dedicated learned subsystem.
Perplexity Context Compression — Productionized query-aware compression cutting context tokens by up to 70%, claiming 50x compression on SimpleQA at frontier-level performance.
Agents and Tools
Google Agent Executor — Open-source distributed agent runtime with durable execution, secure isolation, and connection recovery. Collaborates with GKE Agent Substrate for massive deployments.
Antigravity 2.0 — Agent-first desktop with multi-agent orchestration, dynamic subagents, scheduled tasks, artifacts, and live voice. Google claims a demo built a functioning OS in 12 hours using 93 parallel sub-agents for under $1K.
Gemini Spark: 24/7 Personal AI Agent — Runs on dedicated Google Cloud VMs, checking with users before major actions. Long-running tasks persist while devices are closed.
Making Monorepos Ergonomic for Agents — Basis made its codebase agent-friendly in 3 months using principles of verifiability, interoperability, and canonical context.
Google Adds llms.txt Check to Chrome Lighthouse — New “Agentic Browsing” audit category to enhance machine interaction with websites.
Security
GitHub Confirms Breach of 3,800 Repos — Via a trojanized VS Code extension. Also investigating separate unauthorized access to internal repositories.
Seattle Shield Surveillance Network — Intelligence-sharing network operated by Seattle police involving private companies.
Notable
SpaceX $1.75tn Flotation Could Make Musk a Trillionaire — Prospectus reveals sprawling empire plans and Mars colonization ambition.
New Weight-Loss Shot Outperforms Market — Eli Lilly’s retatrutide shows significantly more weight loss than other GLP-1 drugs in sizable trial.
Colossal: Artificial Womb at One-Yard Line — De-extinction startup claims near-complete artificial mammal womb with 100% development rate. Plans to open-source all technologies.
Disney Erased FiveThirtyEight — Roughly 200,000 human-hours of work wiped. The entire site is gone.
Cloudflare Replacing Employees with AI — Laid off 20%+ of workforce. Better measurement tools mean humans can focus on creating and capturing value.
Hacker News
Flipper One — We Need Your Help — (396 comments)
Google’s Antigravity Bait and Switch — (233 comments)
Throwing AI-Generated Walls of Text into Conversations — (268 comments)
We’re Testing New Ad Formats in Search and Expanding Our Direct Offers Pilot — (450 comments)
Vivaldi 8.0 — (218 comments)
Python 3.15: Features That Didn’t Make the Headlines — (133 comments)
The Letter S, by Donald Knuth (1980) — (53 comments)
Lost Images from the 1945 Trinity Nuclear Test Restored — (82 comments)
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 6 newsletters processed, ~40 stories distilled