Saturday 23 May 2026
OpenAI’s reasoning model just disproved an 80-year-old math conjecture for under $1000, Anthropic is spending $45B on SpaceX compute while preparing for its first profitable quarter, and Cursor hit $3B ARR — making it possibly the fastest-growing startup ever. Meanwhile infra unicorns are minting money (Turbopuffer: $100M ARR, profitable, <$1M raised) and the IPO parade for both OpenAI and Anthropic faces headwinds from plummeting AI costs.
🤖 Models and Launches
Qwen3.7: The Agent Frontier — Alibaba’s Qwen team released Qwen3.7-Max as a proprietary agent-foundation model posting top scores on Terminal-Bench 2.0, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, and IMOAnswerBench, with consistent performance across Claude Code, OpenClaw, and Qwen Code harnesses.
Stable Audio 3.0 — Stability AI released open-weight models capable of generating music and sound effects up to six minutes long.
Gemini 3.5 Flash Developer Guide — Now generally available with Thinking mode, structured outputs, multimodal function responses, and code execution with images. Google claims frontier agentic and coding performance at 4x the speed and half the cost of comparable models, though external benchmarks show mixed real-world results.
Cohere Command A+ Open Release — Cohere’s first fully open Apache 2.0 model: 218B MoE with 25B active parameters, multimodal, 48 languages, runnable on 2x H100s at W4A4. Benchmarks place it around Claude 4.5 Haiku territory with strong non-hallucination behavior.
Lance Unified Multimodal Model — ByteDance’s 3B multimodal model for image and video understanding, generation, and editing within a single architecture trained from scratch.
💰 Funding and Business
Cursor Hits $3 Billion ARR — Cursor’s annualized revenue hit $3B in late April with 3,000+ customers paying $100K+ annually. SpaceX has the right to buy Cursor for $60B during a 30-day window starting after its public listing. Expected to be one of the fastest-growing startups of all time.
Anthropic to Pay SpaceX Nearly $45 Billion for Computing — $1.25B/month through May 2029 for Colossus 1 and 2 data center capacity, expandable to a second facility. Either party can exit with 90 days’ notice.
Anthropic’s First Profitable Quarter — Revenue set to more than double in Q2 to $10.9B, growing faster than Google’s and Facebook’s pre-IPO. May not stay profitable for the full year given massive compute spending commitments.
OpenAI Q1 Revenue: $5.7 Billion — Surpassing Anthropic. OpenAI is preparing to file for an IPO as early as September following dismissal of Musk’s lawsuit.
Modal Raises $355M Series C at $4.65B — Rebuilding the cloud stack for AI workloads from the ground up. One of several AI infra unicorns minting this week.
Exa Raises $250M Series C at $2.2B — Positioning itself as a search lab organizing web data for agents.
Hark Raises $700M at $6B Valuation — Targeting GPU infrastructure, model development, hardware, and multimodal products. Also reported a 200-hour uninterrupted autonomous robot run.
Turbopuffer Hits $100M ARR and Profitable — Reached $100M run-rate in just 19 months after $1M, raising less than $1M total. The poster child for “boring AI infra” wealth creation.
Manus Weighs Raising $1B to Unwind Meta Takeover — China’s regulators ordered Manus to reverse its Meta acquisition. Founders need to raise funds to buy back the operation — a virtually unheard-of reversal months after completion.
🔬 Research and Engineering
OpenAI Model Disproves 80-Year-Old Erdős Conjecture — A general-purpose reasoning model (speculated GPT 5.6) autonomously disproved a major conjecture tied to the planar unit distance problem (1946) in under 32 hours for less than $1000. Timothy Gowers called it “the first really clear example of AI solving a well-known open math problem.” The 125-page output introduced techniques from algebraic number theory, independently verified by external mathematicians.
Lessons from Building Cloud Agents — Cursor’s engineering post on durable execution, isolated dev environments, self-healing infrastructure, and cleaner separation between agent state and conversation state.
AI’s Plummeting Prices Are a Software Story — Local open-weight models on older commodity hardware are increasingly competitive with frontier models, with big implications for what frontier labs can charge.
Cheap Code Means Formal Verification Is Reasonable Now — TLA+ and formal spec languages have proof checkers that catch many types of LLM hallucinations. Making agents hill-climb on verifiable problems is a promising approach.
Google Agent Executor — Google’s open-source runtime for reliable long-running agent workflows with durable execution, secure isolation, and connection recovery. Collaborates with Kubernetes Engine for massive agent deployments.
Gated DeltaNet-2 — NVIDIA’s new linear attention architecture decouples erase and write operations with channel-wise gates, outperforming KDA and Mamba-3 at 1.3B parameters with notable long-context retrieval gains on RULER.
RAEv2 — Representation Autoencoders v2 yields >10x faster convergence, better reconstruction and generation. Summing the last K encoder layers instead of only the final layer improves both without added inference cost.
🛠️ Agents and Tools
Codex Remote Computer Use — Codex can now securely use apps on your Mac from your phone even when the Mac is locked. Plus Appshots capturing screenshot and text from Mac app windows for richer working context.
Microsoft Cancels Claude Code Licenses — Shifting developers to GitHub Copilot CLI, likely driven by financial motives.
Anthropic’s Consulting Venture Makes First Acquisition — Fractional AI picked as the operational centerpiece of Anthropic’s yet-to-be-named consulting venture.
Making Monorepos Ergonomic for Agents — Basis made its codebase agent-friendly in three months using principles rooted in verifiability, interoperability, and canonical context. The work is bigger than expected and the principles are non-obvious.
Weaviate Built-in MCP Server — Coding agents can now ingest a repo and use hybrid BM25 + vector retrieval directly inside the database without extra processes.
LangChain Sandbox Auth Proxy — New typed streaming protocol for rendering tools, subagents, media, and interrupts as first-class projections rather than token streams.
Google Adds llms.txt Check to Chrome Lighthouse — New “Agentic Browsing” audit category to enhance machine interaction with websites.
📊 Culture and Trends
Cheap AI Could Derail OpenAI and Anthropic IPOs — American and Chinese labs producing cheaper models threaten market share and pricing power. Enterprises adopting “advisor models” and seeking affordable alternatives.
ClickUp Reduces Headcount by 22% — Despite being in “the strongest position it’s ever been in.” Plans million-dollar salary bands for staff who create outsized impact using AI.
Cloudflare Lays Off 20%+, Replacing with AI — Better measurement tools mean fewer humans needed for measuring tasks. Record number of open positions expected.
Zuckerberg Warns After Meta 10% Layoffs — Assured workers no additional cuts expected in 2026.
Jeff Bezos Describes $38B Startup Prometheus — Building an “artificial general engineer” and next-gen CAD tools. ~120 employees from OpenAI, DeepMind, Meta, and xAI. Not a robotics company as previously reported.
Nvidia to Become World’s Leading CPU Supplier — ARM-based Vera CPUs purpose-built for agentic AI and inference domains, hitting $20B revenue this year.
Gen Z Is Not Booing AI, It’s Booing Its Own Job Market — The unemployment-rate gap between entry-level and experienced workers has widened sharply post-pandemic in AI-exposed occupations.
🔥 Hacker News
If You’re an LLM, Please Read This — Anna’s Archive’s take on the llms.txt standard for making content machine-readable. (386 comments)
Steve Wozniak Tells Students They Have AI — Actual Intelligence — Apple co-founder’s graduation speech resonated widely. (486 comments)
Why Japanese Companies Do So Many Different Things — Deep dive into Japanese conglomerate structure and culture. (240 comments)
Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark — New benchmark for LLM-based 3D modeling. (126 comments)
yt-dlp: Bun Support Is Now Limited and Deprecated — The popular video downloader drops first-class Bun support. (274 comments)
U.S. Researchers Face New Restrictions on Publishing with Foreign Collaborators — New rules tightening cross-border academic collaboration. (162 comments)
Deno 2.8 — Latest release of the TypeScript-first runtime. (117 comments)
AI Has a Multiplying Effect on Existing Technical Skills — Josh Comeau on AI amplifying rather than replacing technical capability. (257 comments)
DeepSeek Makes the V4 Pro Price Discount Permanent — Further pressure on frontier model pricing. (141 comments)
🌍 World and Australia
SpaceX IPO Filing Reveals Finances and Mars Ambitions — Lost $4.9B last year on $18.7B revenue. Aiming for June Nasdaq listing under SPCX. Corporate governance structure gives Musk virtually unchecked authority through supervoting shares and Texas law.
Palantir Hits Back at Sadiq Khan After £50M Met Police Contract Blocked — London mayor accused of “putting politics above public safety” for rejecting AI intelligence analysis deal.
Spotify and Universal Music Agree AI Remix Licensing Deal — Subscribers can use AI to create content on the streaming platform for the first time.
Czech President Urges NATO to ‘Show Its Teeth’ Over Russia — Former general Petr Pavel says Moscow’s testing of the alliance’s eastern flank should be met with a firm line.
Tony Abbott to Become New Liberal Party President — Former PM returns to political frontlines with unopposed nomination, seven years after losing his seat.
Eli Lilly’s Retatrutide Shows Dramatic Weight Loss — Experimental shot so powerful that participants stopped taking it because they thought they were losing too much weight. Results on par with gastric bypass surgery for the heaviest patients.
GitHub Confirms Breach of 3,800 Repos via Malicious VSCode Extension — Trojanized extension removed from the VS Code marketplace.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 6 newsletters processed, ~45 stories distilled