Thursday 28 May 2026
The AI infrastructure gold rush continues. Fireworks and Baseten both eye decacorn status, OpenRouter crosses $1B, and Huawei claims a chipmaking workaround that could rewrite the semiconductor race. Meanwhile, xAI and Cursor get an antitrust-flavoured talking-to from xAI’s lawyers, and the Pope compares AI to the Tower of Babel.
🤖 Models and Launches
MAI-Image-2.5 Launches at #3 on Arena — Microsoft’s text-to-image model debuts at third on the Arena leaderboard, excelling in style variety and text rendering. Breaks a top-5 previously dominated by OpenAI and Google.
Grok Build CLI Enters Beta — xAI launches a coding agent and CLI for SuperGrok and X Premium Plus subscribers, with plan mode reviews and headless subagent support.
GPT-5.6 Leaks: Coming in June — Leaked details point to stronger multi-step reasoning, better agentic workflows, and improved frontend generation.
Gemini 3.5 Flash: Best Model at Its Speed Point — Up to ~280 output tok/s with materially stronger agentic performance, but ~5x the cost of Gemini 3 Flash. Outscores 3.1 Pro on Terminal-Bench and MCP Atlas.
Qwen 3.7 Max Debuts at #4 on Code Arena: Frontend — Roughly on par with Claude Opus 4.6 on agentic webdev tasks. Commenters caution that the Max model may not represent open-weight releases.
💰 Funding and Infra
OpenRouter Raises $113M Series B at $1.3B Valuation — Weekly token volume grew from 5T to 25T in six months. The AI gateway processes 100T tokens monthly across 400+ models.
Fireworks in Talks at $15B Valuation — 3.75x valuation increase in 7 months. Baseten also raising at $11B (2.2x in 3 months). The inference infrastructure layer is having its moment.
SpaceX S-1: Two AI Compute Stories, Only One Generates Revenue — Terrestrial data centres with Anthropic as a $1.25B/month customer through 2029, plus the vision of orbital AI inference on Starlink satellites.
🏢 Industry
xAI Warns Staff to Limit Contact with Cursor Employees — Standard acquisition caution, but arriving late given the companies’ employees have been working alongside each other for weeks. Any co-mingling accusation could jeopardise the deal.
China Expands Travel Curbs to Top AI Talent at Private Firms — Startup founders, researchers, and executives at private AI firms now need government approval for overseas travel. Unprecedented extension of state restrictions to the private sector.
Microsoft Cancels Most Direct Claude Code Licenses — Engineers in Experiences and Devices must migrate to GitHub Copilot CLI by June 30. The pullback exposes how unit economics of enterprise AI coding don’t work at current token prices.
Pope Leo XIV Compares AI Threat to Tower of Babel — The encyclical warns AI will reduce humans to cogs in an efficiency machine and that concentration of power in a few private actors must be countered. Surprisingly approachable writing, even for non-Catholics.
Samsung Memory Chip Staff Get £310K Bonuses from AI Profit-Sharing — Agreement averts strike and shows the latest impact of the AI boom as two more chipmakers join the $1tn club.
🔬 Research and Benchmarks
DeepSWE: A Real Benchmark for Long-Horizon Software Engineering — 91 repos, 5 languages, contamination-free tasks with reliable verification. Creates sharper separation than SWE-Bench Pro. Called “the first code bench that actually aligns with how it feels to use these models coding.”
Claude Mythos Solves OpenAI’s Landmark Erdős Problem — With an appropriate harness, both Mythos and GPT-5.5 can reproduce what an internal model did one-shot, implying large latent capability not exposed by vanilla chat UX.
AlphaProof Nexus Solves Decades-Old Math Problems for a Few Hundred Dollars — Google DeepMind’s system autonomously solved 9 of 353 open Erdős problems at inference costs of a few hundred dollars each.
How Anthropic Contains Claude Across Products — Design for containment at the environment layer first, then steer behavior at the model layer. Match isolation strength to the user’s capacity for oversight.
NVIDIA CompileIQ Auto-Tuning in CUDA 13.3 — AI-driven evolutionary algorithms auto-tune GPU compiler settings for up to 15% performance improvements on already-optimized inference and training tasks.
🛠️ Agents and Tools
Microsoft Webwright: A Terminal Is All You Need for Web Agents — SWE-style browser agent framework achieving state-of-the-art results on long-horizon web tasks. Gives agents a terminal to launch multiple browser sessions.
Claude Code as a Daily Driver — Deep dive on Claude.md, skills, subagents, plugins, and MCPs. 219 HN comments discussing the practical reality of AI-first development workflows.
vLLM Merges Rust Frontend — ~837 req/s vs ~162 req/s on a preprocess-heavy workload in a single process. Drop-in alternative to the Python API server.
Agent Sandbox for Isolated Workloads — Manages isolated, stateful singleton pods with stable identity and persistent storage. Ideal for AI agent runtimes.
📊 Culture and Trends
Tech CEOs Are Apparently Suffering from AI Psychosis — 241 HN comments dissecting the pattern of tech leaders losing touch with reality when discussing AI capabilities.
DuckDuckGo Sees 28% More Visits After Google Says People Love AI Mode — The market is voting with its feet. 274 HN comments.
Simon Willison: Anthropic and OpenAI Have Found Product-Market Fit — 606 HN comments. The argument that the AI lab business model has been validated, even if the path to profitability remains unclear.
I’m Tired of Talking to AI — 844 HN comments. A growing counter-narrative to the AI-first everything movement.
🔒 Security and Geopolitics
Huawei Claims Chip Workaround to Match Leading Processors by 2031 — Claims to achieve transistor density matching 1.4nm process through a technique that bypasses lithographic restrictions. The US semiconductor blockade may have a shelf life.
US Seeks to Give Weapons-Grade Plutonium to Startups for Fuel — The Trump administration plans to convert surplus plutonium from dismantled warheads into nuclear power plant fuel. 50+ tons available.
🔥 Hacker News
I’m Tired of Talking to AI — (884 comments)
Private Equity Bought America’s Essential Services — (446 comments)
I Think Anthropic and OpenAI Have Found Product-Market Fit — (606 comments)
DuckDuckGo Search Saw 28% More Visits After Google AI Mode — (274 comments)
Tech CEOs Are Apparently Suffering from AI Psychosis — (241 comments)
Claude Code as a Daily Driver — (219 comments)
Canada to Order Military Plane Fleet from Sweden — (197 comments)
Last.fm Is Now Independent — (151 comments)
YouTube to Automatically Label AI-Generated Videos — (149 comments)
All of Human Cooking Compressed into 2 Megabytes — (133 comments)
🌍 World and Australia
Middle East Crisis: Rubio Says ‘Some Progress’ on US-Iran Deal — Trump and secretary of state give conflicting messages as US president meets with cabinet. Israel launches 120+ airstrikes on southern Lebanon.
Nearly Half a Million Russians Killed in Ukraine War — UK spy chief says Russian forces are “going backwards on the battlefield” for first time since late 2022.
Extreme Heat in Europe: ‘A Brutal Reminder’ of Climate Crisis — UK and France broke temperature records on consecutive days. At least seven people died in UK waterways trying to cool off.
NASA Selects Blue Origin for First of Three Lunar Missions — Three uncrewed missions planned this year to kickstart a $20B moon base construction, with operating capability targeted by 2032.
NDIS Cuts to Remove 241,000 Participants — New modelling shows the extent of government cuts over a four-year period. Labor also spruiks automatic $3,000 compensation for scam victims.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 5 newsletters processed, ~40 stories distilled