staff revolt killed Meta's purge ✊, Nvidia guarantees buyer revenue 💸, a third of new compute is OpenAI's 🧊

Zuckerberg's team hatched the two-wave plan at a January retreat, then cancelled the November wave on 19 May, hours before Meta laid off 10 per cent of staff. Internal data showed the autonomous agent technology behind the plan was not delivering hoped-for productivity gains, and employees were in open revolt. By June at least 11 units had adopted the small pods designed to eliminate layers of middle management.

A Molniya drone carrying a Jetson Orin module chose its own aim point, then struck an apartment wall near a Zaporizhzhia gas station, killing three civilians. The recovered modules carried no encryption, so investigators could read the terrain imagery for visual navigation and the code listing which object classes it was trained to attack. Data-centre AI accelerators fall under US export controls; edge modules starting at $249 do not.

OpenAI built Jalapeño with Broadcom from a blank slate for LLM inference, and SemiAnalysis ran its InferenceX suite on the chip in OpenAI's lab. On output tokens per megawatt it passes Vera Rubin's July figures, and on tokens per dollar the two are roughly level, though Jalapeño ran without speculative decoding and Rubin did not. Vera Rubin systems already reach customers, while OpenAI has only engineering samples.

Open-weight models took 54 per cent of token volume on Vercel's AI Gateway on Tuesday, against 46 per cent for proprietary ones. DeepSeek-V4-Flash led, with Step 3.7 Flash, GLM-5.2 and the 0731 update second, fourth and fifth, and OpenAI's GPT-5.6 Luna third. Ramp puts Fable 5 at 6 per cent of Anthropic tokens businesses bought in its first month, and 11.4 per cent of spending on Anthropic's models.

Nvidia's backstops could put it on the hook for some $230 billion: $105 billion on an OpenAI lease in Ohio, and up to $125 billion of residual-value support under a $500 billion financing plan. In Australia it made itself customer of last resort for two cloud companies' unrented compute, taking a share of revenue above an agreed level. Nvidia calls that a new business model, and reports earnings on Wednesday.

Chris Malone, hired in March 2025 from data centre roles at Meta and Google, has left OpenAI, which says it reorganised its infrastructure organisation. He follows revenue chief Denise Dresser, Brad Lightcap after eight years, Fidji Simo, and four other executives who left in April. OpenAI is working to justify an $852 billion valuation, filed its IPO prospectus confidentially in June, and told staff it will be public in 2027.

Portable Computer runs the orchestrator, planner, tool router, scheduler, task queue and local search index on an NVIDIA DGX Spark, using Qwen 3.8 27B or the post-trained PPLX 27B. Local work carries no per-credit charge, and escalation to the web, connected apps or one of 15-plus frontier models needs the user's permission. It reaches Pro and Max subscribers on Linux first, with Windows and RTX GPU PCs to follow.

One-click root exploits now circulate for fully patched Google Pixel phones, and with one David Buchanan signed an AI-generated image as a genuine photo from the Pixel Camera app. C2PA camera apps lean on Android's Key Attestation and Play Integrity to stop tampering, and rooting defeats both without the attestation noticing. Google closed his report as infeasible to fix, and every Android C2PA app he examined breaks the same way.

GitHub Next has prototyped Knowledge Compressor, which rewrites documentation over and over to strip repetition while keeping the knowledge usable by a model. In one evaluation it cut a technical article from 996 tokens to 480, and it reports that typical documentation halves without meaningfully hurting how useful models find it. For builders packing docs into agent context, the same budget now carries twice the material.

Dylan Patel told Dwarkesh that lab economics have turned: Anthropic began profiting in Q2, and revenue from serving models has run as high as $50 million per megawatt. Base compute costs near $10 to $15 million per megawatt, so serving profit now funds training. Patel says that spread lets the labs pay the highest price for new capacity, which is why new builds get leased to them first.

Koetsier priced Apple's M5 Macs: a 128GB Mac Studio runs about $5,000 over three years, $140 a month, against the $200 Claude Max tier. What fits in that memory is a tier down, Qwen3-Coder-Next scoring 30.9 per cent on Terminal-Bench 2.0 against Opus 4.5's 53.9 per cent, and Kimi K3's 1.4TB of weights fitting nothing Apple sells. He concludes the buy earns its keep on privacy and compliance, not economics.

Paul Dix argues Bun 1.4 marks the end of programming as we know it, after its Zig to Rust rewrite landed as more than a million lines. Jarred Sumner drove it alone with a pre-release Fable 5 and a near-unlimited token budget, and over 11 days agents produced 6,778 commits worth roughly $165,000 at API pricing. Dix expects humans to review only the finished software, leaving the code itself unread.

AI companies facing wrongful death suits are running a striking defence: their models' outputs are legally speech, protected by the First Amendment. A public database counts 19 deaths connected to AI use across 2025 and the first seven months of 2026. A judge has already rejected Character.AI's free-speech defence, and the company settled multiple mental health lawsuits in January, leaving liability for model output unsettled.

Vercel released Run SDK, which executes agent-written JavaScript and TypeScript in a locked-down sandbox with no route to your application's secrets or the network. The application exposes only the operations it chooses, and a run can stop midway for authentication or a person's approval, then resume without repeating finished work. That turns a bare eval call into a boundary an approval step can actually sit on.

LatticeDB is an embedded database that keeps a graph, vector similarity search and keyword search in a single portable file and one query language. On the project's own published benchmark, finding the ten nearest matches among a million vectors returns in 0.83 milliseconds while still finding every correct result. For builders prototyping retrieval or agent memory locally, that replaces three separate systems with one file.