Saturday 2 May 2026
A week of strategic repositioning: OpenAI pushes Codex beyond code into general knowledge work, Cursor is acquired by xAI for $60B, and Anthropic nears a $900B valuation while shipping Claude Security to public beta. Meanwhile, the open-weight ecosystem continues to surge with Qwen3.6 claiming the sub-150B crown and Grok 4.3 getting cheaper and smarter.
🤖 Models
Grok 4.3 — xAI’s latest model scores 53 on the Intelligence Index (up 4 points from Grok 4.20 v2) with ~40% lower input prices and ~60% lower output prices. Major jump on agentic benchmarks including GDPval-AA to 1500 Elo.
Qwen3.6 27B — New open-weights leader under 150B parameters with an Intelligence Index score of 46, ahead of Gemma 4 31B. Apache 2.0 licensed, 262K context, native multimodal input, fits on a single H100. The companion 35B A3B MoE is the strongest open model around 3B active parameters.
GLM-5V-Turbo — Integrates multimodal perception directly into reasoning and tool use, improving performance on coding, visual tasks, and agent workflows across heterogeneous inputs.
GPT-5.5 matches Mythos on cyber evals — UK AI Security Institute reports GPT-5.5 became the second model to complete its multi-step cyber-attack simulation end-to-end, with 71.4% average pass rate vs Mythos’ 68.6%. Performance still improving past 100M tokens of inference budget.
Tencent Hy3-preview — A 295B total / 21B active MoE with 256K context, scoring 42 on AA’s Intelligence Index. Notable bright spot on CritPt scientific reasoning where it matched GLM-5.1.
💰 Funding & Industry
Anthropic nearing $900B valuation — Reportedly closing a ~$50B round driven by strong investor demand and rapid revenue growth nearing $40B run rate.
Cursor acquired by xAI for $60B — The most operationally successful software company of the AI era sold rather than underwrite the path to $100B. Gives xAI an application surface before the SpaceX IPO and Cursor a sponsor with compute and a non-competing model lab.
Apple earnings soar past Wall Street expectations — $111.2B in revenue in the first earnings report after Tim Cook’s pending departure announcement.
Samsung posts record $30B+ quarterly profit — AI boom has pushed memory-chip industry into a super boom cycle. Supply crunch expected to grow worse next year.
Zuckerberg blames slower sales on Iran war, layoffs on AI costs — In a companywide meeting, Meta’s CEO addressed the 8% share drop, attributing ad business slowdown to geopolitical spending pullback and planned layoffs to increased AI investment.
🛠️ Agents & Tools
OpenAI Codex expands to general knowledge work — “Codex for Work” pushes beyond coding with role-based onboarding, Microsoft/Google/Salesforce integrations, 42% faster computer use, and a dynamic task-specific UI. Sam Altman: “try it for non-coding computer work.”
Claude Security now in public beta — Powered by Opus 4.7, identifies and patches software vulnerabilities. Integrated with Microsoft Security and Palo Alto Networks for ongoing code scanning without custom API integration.
Claude adds creative tool support — Anthropic now supports Blender, Autodesk, Adobe Creative Cloud, Ableton, Splice, Canva, Affinity and more — a significant expansion into creative professional workflows.
Cursor’s agent harness engineering — Published how it tests and tunes its agent harness with vision-driven development, A/B testing, and dynamic context adaptation. The industry shift from model-centric bragging to harness-centric engineering continues.
Perplexity expands enterprise AI workflows — Added workflows, enterprise data connectors, and integrations like Teams and Excel for structured business tasks and continuous automation.
🔒 Security
Copy Fail: 732 bytes to root on every major Linux distro — A logic bug in the Linux kernel’s authencesn cryptographic template allows unprivileged users to trigger a controlled 4-byte write into the page cache. Exploitable via a single 732-byte Python script to obtain root on essentially all Linux distributions since 2017. Discovery was AI-assisted. Patch available.
PyPI lightning package compromised — Versions 2.6.2 and 2.6.3 contained malicious code that executed on import, downloaded Bun, and ran an 11MB obfuscated JavaScript payload for credential theft. The tempo of supply-chain attacks is increasing.
Mozilla pushes back on Google’s Prompt API in Chrome — The Prompt API gives web pages the ability to directly prompt a browser-provided language model. Mozilla says it causes severe negative consequences to interoperability, updateability, and neutrality of the web platform.
Musk testifies xAI trained Grok on OpenAI models — Confirms what the industry widely assumed: distillation techniques between US labs. Came during heated cross-examination in the OpenAI trial.
🔬 Research
Anthropic’s 1M conversation sycophancy study — Large-scale analysis of guidance and sycophancy tied behavioral research directly to training changes for Opus 4.7 and Mythos Preview — post-training loops becoming more data-informed.
Qwen-Scope interpretability toolkit — Open suite of sparse autoencoders for Qwen3/Qwen3.5 models focused on feature steering, debugging, data synthesis, and evaluation. A rare release focused on understanding rather than raw weights.
KV Cache locality: the hidden variable in LLM serving cost — Same GPUs, same model, same traffic can produce measurably different throughput depending on which GPU gets which request. The post discusses cost of recomputation and what changes when load balancers understand token locality.
Speculative decoding for RL training — Applied to RL rollouts without changing output distributions, delivering up to 1.8x throughput gains and projected 2.5x end-to-end speedups at scale.
🔥 Hacker News
Grok 4.3 (484 comments) — xAI’s latest model release draws massive discussion.
AI uses less water than the public thinks (253 comments) — California Water Blog challenges the narrative around AI’s water consumption.
Spotify adds ‘Verified’ badges to distinguish human artists from AI (185 comments) — Platform moves to differentiate human-created music from AI-generated content.
Your website is not for you (169 comments) — Design discussion on user-centric web development.
OpenWarp (135 comments) — New open-source project gaining traction.
Show HN: WhatCable — USB-C cable inspector (122 comments) — Tiny menu bar app for inspecting USB-C cables.
The Gay Jailbreak Technique (85 comments) — Novel jailbreak approach sparks debate on LLM safety.
City learns Flock accessed cameras in children’s gymnastics room as sales demo (51 comments) — Surveillance company Flock used children’s gym cameras for a demo, city renews contract anyway.
🌐 World & Tech Pulse
Iran offers new peace proposal to US — Tehran passed proposal to mediators in Pakistan; Trump claims hostilities have ended in letter to Congress.
Sabastian Sawe’s sub-two hour marathon — Record-breaker receives hero’s welcome in Kenya after London Marathon feat.
1X starts humanoid robot production in California — 58,000 sq ft facility in Hayward can produce 10,000 NEO robots per year, targeting 100,000+ units by 2027.
UK job hunters frustrated with AI interviews — Survey finds nearly half of job seekers have been interviewed by AI, describing the process as “awkward and humiliating.”
Silicon Valley bracing for a permanent underclass — Many believe AI will displace millions of jobs, ferry power and wealth to AI companies and capital owners while ordinary people lose economic leverage.
Netflix launches Clips — TikTok-like vertical feed — Rolling out to US, UK, Australia and more. Shows short clips from series and films tailored to user tastes.
Gun licence applications surge in Australia after Bondi — Premiers reject Albanese’s buyback scheme amid surge in applications.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 4 newsletters processed, ~45 stories distilled