Monday 11 May 2026
Anthropic’s valuation hits $1.2T while Cloudflare, Block, and Coinbase slash thousands of jobs citing “AI readiness.” OpenAI rolls out GPT-Realtime-2 and expands Codex into a long-running agent runtime with Chrome integration. Meanwhile, DeepMind’s AI co-mathematician scores 48% on FrontierMath Tier 4 — and a mathematician says the result could form a PhD thesis chapter.
🤖 Models and Launches
OpenAI Ships GPT-Realtime-2, Realtime-Translate, and Realtime-Whisper — GPT-Realtime-2 brings GPT-5-class reasoning to live voice conversations. Realtime-Translate supports 70+ input languages and 13 output languages. Realtime-Whisper handles streaming transcription for live captions and meeting notes.
Zyphra Releases ZAYA1-74B-Preview Under Apache 2.0 — A 74B-parameter MoE model with only 4B active parameters, trained on AMD hardware. Community sees it as proof that Zyphra has moved beyond small-MoE experimentation. Also shipping ZAYA1-VL-8B, a vision-language MoE.
Kimi K2.6 Emerges as 5x Cheaper Alternative to Opus — Multiple practitioners report swapping from Sonnet 4.6 and Opus 4.7 to Kimi K2.6 without noticing quality differences for coding and agent workloads. Reflects a broader trend: open-source LLMs are now viable defaults in agentic stacks.
ds4.c: A Native Inference Engine for DeepSeek V4 Flash — An intentionally narrow, Metal-only inference engine from the Redis creator. Aims to make one local model feel “finished from end to end.” Still alpha.
🔬 Research and Engineering
DeepMind’s AI Co-Mathematician Scores 48% on FrontierMath Tier 4 — A new high on the hardest math benchmark. Mathematician Timothy Gowers said the system proved a result that could plausibly form a PhD thesis chapter. Relies on custom infrastructure and large budgets — not directly comparable to standard leaderboard runs.
Anthropic: Teaching Claude Why — Anthropic claims to have eliminated Claude 4’s previously observed blackmail behavior. The key insight: demonstrations alone were insufficient — better results came from teaching the model why misaligned behavior is wrong, using constitution-based documents and fictional aligned-AI stories.
TwELL: Sparse Packing Format Yields 20%+ Speedups on H100s — Sakana AI and NVIDIA collaborate on a sparse packing format and kernel stack for transformer FFNs that reshapes sparsity to fit GPU execution rather than forcing generic sparse formats.
DGPO: Distribution-Guided Policy Optimization — A refinement over GRPO using token-level reward redistribution, Hellinger distance instead of KL, and entropy gating. Reports 46.0% on AIME 2025 and 60.0% on AIME 2024.
AlphaEvolve Expands Across Google Infrastructure — Google’s Gemini-powered coding agent is now being used for AI infrastructure optimization, molecular simulations, and natural disaster risk prediction. Doubled training speed for massive AI models and routing optimizations saving 15,000 km of travel annually.
🏢 Industry
Anthropic Valued at $1-1.2T, Overtaking OpenAI — After a “miracle Q1” of 80x annualized growth and one-month jump to $15B ARR, Anthropic is now the 11th-15th most valuable company in the world. Meanwhile, everyone else is shrinking.
Cloudflare Slashes 1,100 Jobs in AI Restructuring — The company claims it will become “faster and more innovative” by embracing an “agentic AI-first operating model.” Expected charges: $140-150M. Block (40%) and Coinbase (14%) also announced major layoffs citing AI readiness.
Musk Tried to Hire OpenAI Founders to Start AI Lab Inside Tesla in 2018 — OpenAI claims Musk tried to bring Altman, Brockman, and Sutskever to Tesla, contradicting his accusations that Altman “stole the charity” by converting OpenAI to for-profit.
Google Misstates Carbon Emissions of Proposed UK Datacentres by Factor of Five — Emissions understated in Essex plans for the tech giant, with similar errors found in Lincolnshire proposals.
🛠️ Agents and Tools
Codex Now Works Directly in Chrome on macOS and Windows — OpenAI Codex works in parallel across browser tabs in the background, writing code to navigate structured pages and complete complex data flows under the hood.
The Six-Hour Codex Run That Survived a Five-Hour Pause — The new /goal feature in Codex persists goals across terminal restarts, laptop sleeps, and multi-hour gaps. Users can turn off their devices and pick up exactly where they left off.
Meta Prepares Hatch AI Agent with Waitlist and Social Skills — Meta’s consumer-grade AI agent for image/video generation, shopping, and learning, integrated into Instagram and Facebook. Internal tests expected by June.
Direct Corpus Interaction Replaces RAG for Agent Retrieval — DCI replaces embedding model + vector DB + top-k retrieval with direct use of grep/find/bash over raw corpora. Reports BrowseComp-Plus improvement from 69% to 80% on Claude Sonnet 4.6.
Perplexity Brings Personal Computer to Mac — AI agents get access to local files, applications, connectors, and the web through the Perplexity desktop app.
📊 Culture and Trends
Notes from Inside China’s AI Labs — Chinese scientists are more willing to do non-flashy work to improve models rather than push their own ideas. The Chinese AI community feels more like an ecosystem than battling tribes, with labs having nothing but respect for peers.
Long AI, Short AGI — Intelligence is commoditizing like compute, bandwidth, and storage before it. The real winners won’t necessarily have superior models but will own customer relationships and proprietary data.
TokenMaxxing, PromoMaxxing, and Misaligned Incentives in Tech — “When a measure becomes a target, it ceases to be a good measure.”
The Agent Principal-Agent Problem — Contributors submitting slop PRs is an unmanageable disaster — reviewers could be more productive by prompting agents themselves.
AI Load Breaks GitHub — Why Not Other Vendors? — Microsoft’s competitors seem to be keeping up with increased AI load, suggesting GitHub hasn’t been responding to growth like a world-class engineering organization.
🔒 Security
Mozilla Hardens Firefox with Claude Mythos Preview — Mozilla identified and fixed an unprecedented number of latent security bugs using AI. Many discovered bugs would need to be combined with other exploits for a full-chain compromise.
OpenAI Launches GPT-5.5-Cyber for Critical Infrastructure Defenders — Limited preview for enterprise/government cybersecurity. The upcoming U.S. AI security executive order will emphasize collaboration with frontier labs on cyber defense rather than pre-approval of frontier models.
Hardware Attestation as Monopoly Enabler — GrapheneOS argues that hardware attestation mechanisms are being used to lock down device ecosystems rather than genuinely improve security. (212 comments)
Incident Report: CVE-2024-YIKES — A detailed incident report walking through a real-world vulnerability disclosure and response. (74 comments)
🔥 Hacker News
Louis Rossmann Offers to Pay Legal Fees for OrcaSlicer Developer — Right-to-repair advocate steps in after Bambu Lab threatens an open-source 3D printing developer. (216 comments)
Show HN: Building a Web Server in Assembly to Give My Life Meaning — A deliberately absurd project that resonated with the community. (204 comments)
Local AI Needs to Be the Norm — Arguments for running AI models locally rather than relying on cloud APIs. (139 comments)
GitHub Is Sinking — Criticism of GitHub’s trajectory and recent decisions. (125 comments)
Replacing a 3 GB SQLite Database with a 10 MB FST Binary — Finite state transducers compress a massive database down to 10MB. (30 comments)
YC’s Biggest Scandals — A compilation of Y Combinator controversies. (64 comments)
Space Cadet Pinball on Linux — Getting the classic Windows pinball game running on Linux. (100 comments)
🌍 World and Australia
Trump Calls Iran’s Response to Peace Plan “Totally Unacceptable” as Ceasefire Frays — Drones strike Gulf nations and Israel warns war “not over” as diplomatic tensions escalate.
Russia Breaches Three-Day Ceasefire with Ukraine — Zelenskyy says Russian army is “not complying” with the US-mediated truce and is “not even really trying to.”
Hungary’s New PM Magyar Apologises to Those Wronged Under Orbán — Péter Magyar strikes a radically different tone to his predecessor in his first speech as PM.
Iranian Politician’s Australian Links Revealed — The son of Iran’s peace negotiator lived and worked in Melbourne, raising questions about sanctions and security.
Australia Budget to Include Extra $2bn for Infrastructure — Budget funds to speed up environmental approvals as One Nation wins Farrer byelection.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 4 newsletters processed, ~45 stories distilled