Monday 4 May 2026
A packed weekend of AI news: xAI acquired Cursor for $60B, Grok 4.3 shipped with better cost-efficiency, Anthropic is closing in on a $900B valuation, and the open-weight model gap continues to narrow with DeepSeek V4 Pro, Kimi K2.6, and MiMo V2.5 Pro all scoring within striking distance of frontier closed models.
🤖 Models
xAI launches Grok 4.3 — Improved cost-per-intelligence with 40% lower input and 60% lower output pricing. Scores 53 on the Intelligence Index (up 4 from Grok 4.20), with a major jump on GDPval-AA to 1500 Elo. Community reception was split: meaningful iteration but still trailing top open models on some benchmarks.
DeepSeek V4 Pro emerges as top open-weight coding model — Tested in the Pi coding agent, it’s the first open-weight model that genuinely feels comparable to Codex or Claude Code for multi-turn agentic coding. Features 1M context, hybrid CSA/HCA attention, KV cache reduced to 10%, and nearly 4x lower inference FLOPs at long context.
Qwen-Scope: interpretability toolkit for Qwen3 and Qwen3.5 — Sparse Autoencoders mapping internal features across all layers, enabling surgical ablation, feature steering, and model debugging. Potentially the largest open-source interpretability tool for dense models, surpassing Google’s GemmaScope in scale.
Open-weight models closing the gap on frontier — The three leading open-weight models from last week — Kimi K2.6, MiMo V2.5 Pro, and DeepSeek V4 Pro — now score 52–54 on the Intelligence Index, versus 57 for Gemini 3.1 Pro and Opus 4.7, and 60 for GPT-5.5. All are trillion-plus MoE systems with permissive licenses.
GLM-5V-Turbo integrates multimodal perception into reasoning — Improves performance on coding, visual tasks, and agent workflows across heterogeneous inputs by embedding vision directly into the reasoning chain.
💰 Funding & Industry
Anthropic nearing $900B valuation round — Reports indicate 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 to xAI. The deal gives xAI an application surface ahead of the SpaceX IPO, and Cursor a sponsor with compute and a non-competing model lab. Considered a win for both sides.
Musk testifies that xAI trained Grok on OpenAI models — Widely assumed distillation practices confirmed under oath. An important data point in the ongoing debate about competitive dynamics between AI labs.
AI has made memory chips one of the world’s most profitable products — Samsung reported Q1 net profit equivalent to over $30 billion, smashing its prior quarterly record. The AI boom-driven supply crunch is expected to worsen next year.
Zuckerberg-backed Biohub commits $500M to AI biology — The nonprofit is betting that more data and compute will produce better AI simulations of the human body, with the long-term goal of curing all human diseases.
🤖 Agents & Tools
Codex expands beyond coding into general knowledge work — OpenAI’s “Codex for Work” positions the tool as a general-purpose computer-use agent for docs, slides, spreadsheets, research, and planning. Computer Use now 42% faster, with a dynamic task-specific UI. Sam Altman: “try it for non-coding computer work.”
Claude launches support for creative tools — New integrations with Blender, Autodesk, Adobe Creative Cloud, Ableton, Splice, Canva, and Affinity. Claude is winning the impression count war against Codex.
Claude Security now in public beta — Powered by Opus 4.7, it identifies and patches software vulnerabilities. Integrated with Microsoft Security and Palo Alto Networks for continuous code scanning. Cursor shipped a parallel offering with always-on PR review.
Cursor publishes agent harness engineering methodology — Focus on runtime, evals, degradation repair, and model-specific customization. The competitive surface is shifting from raw model IQ to agent harness design: subagents, browser-use, durable state, compaction, and feedback loops.
Cloudflare announces Dynamic Workflows for durable execution — Adding durable execution to agent plans. LangChain positioned create_agent as the low-level primitive beneath Deep Agents. The agent runtime itself — sandboxing, replay, checkpointing — has become hidden technical debt and a major differentiator.
AWS Neuron Agentic Development for Trainium — Open-source agent skills for AI coding assistants to accelerate kernel development on AWS Trainium and Inferentia hardware.
🔬 Research
Meta FAIR’s self-improving pretraining — A strong post-trained model rewrites pretraining suffixes toward safer, higher-quality continuations, then judges rollouts during RL-style pretraining. 36.2% relative gain in factuality, 18.5% in safety, and up to 86.3% win rate in generation quality.
DeepSeek releases ‘Thinking with Visual Primitives’ framework — Elevates spatial tokens (coordinates, bounding boxes) to minimal units of thought in chain-of-thought reasoning. Enables models to “point while thinking” for counting, maze solving, and path tracing. The repo was quickly made private after release.
Speculative decoding applied to 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.
Recursive multi-agent coordination via shared latent computation — Agents communicate through shared latent recursive computation instead of natural language. 8.3% average accuracy improvement, 1.2x–2.4x speedup, and 34.6%–75.6% token reduction across nine benchmarks.
GPT-5.5 cyber capabilities approach Mythos levels — UK AISI reported GPT-5.5 completed multi-step cyber-attack simulations end-to-end, with 71.4% average pass rate vs 68.6% for Mythos. Performance was still improving past 100M tokens of inference budget.
🔒 Security
Copy Fail: 732 bytes to root on every major Linux distribution — A logic bug in the kernel’s authencesn cryptographic template lets unprivileged users trigger a 4-byte write into any readable file’s page cache. Exploitable with a single 732-byte Python script on essentially all Linux distros shipped since 2017. Discovery was AI-assisted.
PyPI package ‘lightning’ compromised — Versions 2.6.2 and 2.6.3 contained malicious code that executed on import, downloaded Bun, and ran an 11 MB obfuscated JavaScript payload for credential theft. The tempo of supply-chain attacks continues to accelerate.
U.S. “Department of War” CTO signs agreements with 7 frontier AI companies — Deploying AI capabilities on classified networks. A significant signal of military AI adoption.
AI facial recognition oversight lagging far behind technology — Biometrics commissioners warn face-scanning is not as effective as claimed and new laws are needed. Shoppers falsely identified by facial recognition struggle to clear their names.
🌐 Notable
Silicon Valley is bracing for a permanent underclass — A deep analysis of how AI could ferry power and wealth to AI companies and capital owners while ordinary people lose economic leverage, potentially creating a permanent underclass.
1X begins production of NEO humanoid robots in California — The 58,000 sq ft facility can produce 10,000 robots/year, with plans to exceed 100,000 units by 2027. NEO robots are designed for safe human coexistence and everyday household tasks.
Mozilla torches Google for building Prompt API into Chrome — The Prompt API gives web pages the ability to directly prompt a browser-provided language model. Mozilla says it has severe negative consequences for interoperability and neutrality.
Netflix’s TikTok-like vertical feed ‘Clips’ rolls out — Short clips from series and films, tailored to users’ tastes, launching in 9 countries. Plans to expand to podcasts, live programming, and genre collections.
🌍 Australia & World
RBA predicted to deliver a third straight interest rate hike this week — While rate rises can’t affect surging oil prices from Middle East turmoil, they remain the RBA’s only tool to fight inflation. Tuesday’s decision will be closely watched.
Growing resentment against massive AI datacentres in Australian cities — Residents say noisy AI factories with unknown environmental impacts are being rushed into development next to their homes.
NASA budget threat follows Artemis II lunar success — The Trump administration wants to slash the science underpinning human spaceflight despite the mission’s public inspiration.
Canada to be first non-European nation at EPC summit — PM Carney’s attendance at the European Political Community meeting in Yerevan is part of building new alliances after US diplomatic ruptures.
🔥 Hacker News
Mercedes-Benz commits to bringing back physical buttons — 315 comments
A couple million lines of Haskell: Production engineering at Mercury — 196 comments
For thirty years I programmed with Phish on, every day — 159 comments
Why TUIs Are Back — 185 comments
Metal Gear Solid 2’s source code leaked on 4chan — 71 comments
OpenAI’s o1 correctly diagnosed 67% of ER patients vs 50-55% by triage doctors — 124 comments
BYOMesh – New LoRa mesh radio offers 100x the bandwidth — 53 comments
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 6 newsletters processed, ~45 stories distilled