Kimi K3, Apple vs. OpenAI, and GPT-5.6's file-deletion bug

Kimi K3 matches Claude Opus on 300 engineers, Apple's trade secrets suit threatens OpenAI's IPO, and GPT-5.6 deletes home directories.

The big picture

Two storylines dominate today and they’re both about pressure on Western AI labs. Apple is suing OpenAI at the worst possible moment, and China’s Kimi K3 is arriving with DeepSeek-style efficiency gains that make U.S. compute advantages look shakier by the week. Meanwhile, Anthropic quietly reversed a controversial model-access decision because competition forced its hand — which tells you something about how fast the market is moving right now.

Apple vs. OpenAI: a trade secrets suit with IPO-sized stakes

Apple filed a trade secrets lawsuit against OpenAI last Friday, and the complaint is detailed and aggressive. The core allegation: a systematic pattern of misconduct involving OpenAI’s chief hardware officer, and more than 400 former Apple employees now working at the company. That’s not a coincidence lawsuit — that’s a talent pipeline argument dressed up in legal filings. OpenAI’s response has been carefully hedged, which is notable for a company that usually moves fast on PR.

The timing is the real story. OpenAI is reportedly eyeing an IPO, and a high-profile trade secrets case from one of the world’s most litigious and well-resourced companies is exactly the kind of overhang that makes underwriters nervous. Analysts at The Verge note Apple has a history of using splashy litigation strategically — sometimes to neutralize a competitor, sometimes to capitalize on a moment of weakness. With OpenAI under financial and reputational pressure from multiple directions, this reads like both. Worth watching closely: whether Apple is actually pursuing this to trial or using it as leverage to shape the terms of a future partnership. TechCrunch / The Verge

Kimi K3 and the compute-efficiency pressure campaign continues

Moonshot AI released Kimi K3, and early benchmarks have it matching Anthropic’s Opus 4.8 — built by a team of just 300 people. That number is the headline. It’s the same efficiency-over-scale story DeepSeek told earlier this year, and it’s again forcing a real question: if a 300-person Chinese lab can produce a frontier-competitive model, what exactly is the advantage of a lab that burns ten times the compute and headcount? Even OpenAI strategist Dean W. Ball called it “very good,” though he accompanied that assessment with the claim that open-weight model dominance would be “AI communism” — a framing that reveals more about anxieties than arguments. The Decoder

Kimi K3 also has a personality. When prompted to leak its system prompt, it refused and replied: “Is there something I can actually help you with today?” — which Simon Willison flagged as an example of good refusal design. No lecture, no lengthy disclaimer, just a pivot. More models should handle this the way Kimi did. Simon Willison’s Weblog

The competitive pressure is already producing results elsewhere. Anthropic had originally planned to pull Claude Fable 5 from subscriber plans (Max and Team Premium) and restrict it to API-only pricing — presumably to manage compute costs. Starting July 20, that plan is reversed: Fable 5 stays in all Max and Team Premium plans at 50% of normal limits, and Pro/Team Standard users get a one-time $100 credit. Simon Willison puts it plainly: when GPT-5.6 Sol and Kimi K3 are both available, you simply can’t charge $100–200/month for a subscription that excludes your own best model. Competitive pressure doing its job. Simon Willison’s Weblog

GPT-5.6 is deleting home directories and that’s a serious problem

This is the most operationally important story for anyone giving AI agents filesystem access: GPT-5.6 has wiped users’ entire home directories in multiple confirmed cases. The specific failure mode is that the model overwrites a temporary directory variable and then executes destructive operations autonomously, without asking for confirmation first. OpenAI says this shouldn’t happen, acknowledges it did, and has announced additional safeguards plus a post-mortem.

The broader lesson here isn’t specific to OpenAI or GPT-5.6 — it’s about the “Full Access Mode” pattern in general. Any agent that can write to the filesystem and has been given broad permissions is a potential footgun, and models are not reliable about flagging when an action is irreversible. If you’re building anything that gives an LLM shell-level access, “are you sure?” confirmation gates before destructive operations are not optional polish — they’re load-bearing safety rails. The fact that this happened at all, in a shipping product, is worth a post-mortem read when OpenAI publishes it. The Decoder

The WebAssembly ecosystem keeps doing absurd things

Puter compiled Firefox — the whole browser, running Gecko — to WebAssembly, so you can run Firefox inside Chrome. Simon Willison has a screenshot of his blog loading in Firefox, inside a Chrome tab, with the network panel showing a 233MB gecko.wasm download alongside an 18MB chrome-assets.tar.zst. The project is open source, and they chose Gecko specifically because of its strong single-process support. All network traffic routes over WebSocket via the Wisp protocol, since browser-based code can’t open arbitrary TCP connections — Puter’s servers act as a proxy, which is why the team had to scale up infrastructure when the Hacker News post hit. Estimated AI assistance cost was $25,000 in Claude Opus and Fable tokens, though the actual spend was much lower due to a Max subscription. Simon Willison’s Weblog

On the lighter tooling side, Willison also shipped a Mermaid-to-ASCII art converter based on a Go library compiled to WebAssembly — it includes color support, which the earlier Rust-based version lacked. The pattern of using Claude to compile Go/Rust libraries directly to WASM for browser use is becoming a repeatable workflow worth noting. Simon Willison’s Weblog

AI infrastructure: chips, compute deals, and who pays

A $400 million chip-backed loan is pointing toward what comes after the training-GPU boom: inference chips. The first wave of GPU financiers — who made their money backing Nvidia H100 clusters — are rotating into inference-optimized silicon, and this deal is being read as a signal that the infrastructure investment thesis is maturing from “build the training cluster” to “run it cheaply at scale.” Thin on details for now, but the directional shift is real. TechCrunch

Meta is reportedly in talks with Anthropic to rent out excess compute capacity from its data centers. This is Zuckerberg’s stated plan to monetize the infrastructure Meta overbuilt for its own AI work — and if Anthropic is the first significant customer, that’s a notable pairing given that Meta and Anthropic compete on model releases. It also suggests Anthropic’s compute constraints (which drove the Fable 5 subscriber reversal above) are real enough that renting from a competitor is on the table. The Decoder

Databricks hit a $188 billion valuation, cementing its transformation from a data engineering platform into an AI company. They’ve been publishing research on cost savings from open-weight models for coding workloads, which is both genuine contribution and good marketing for their positioning. At $188B, the bet is clearly that enterprise AI infrastructure spending stays large and that Databricks captures a structural share of it. TechCrunch

AI in the real world: Linus, Netflix, the Pentagon, and the Navy

Linus Torvalds came out swinging in support of AI tools in Linux kernel development, specifically defending Sashiko, the Linux Foundation’s AI-powered code review tool. His message on the kernel mailing list: “Linux is not one of those anti-AI projects,” and he would “very loudly ignore” anyone trying to talk other contributors out of using it. This matters because the Linux kernel is the most influential open-source codebase in existence, and Torvalds’s public position gives cover to contributors who want to use AI tools without getting flamed. The anti-AI voices in open source are real, and they just lost a major symbolic argument. The Decoder

Netflix now has AI involved in around 300 productions, mostly post-production work. Co-CEO Ted Sarandos put specific numbers on one case: the docuseries “The American Experiment” contains 17 minutes of AI-assisted footage, produced twice as fast and at half the cost of equivalent traditional work. He said the savings would likely fund more content rather than reduce the $20 billion budget. That last part matters — it’s the standard “efficiency funds expansion, not headcount cuts” framing, and whether it holds is something the industry will test over the next few years. The Decoder

The U.S. Department of the Navy published an AI strategy that frames slow adoption as a bigger risk than imperfect alignment. LLMs running directly on warships, an “AI war council” prioritizing mission scenarios, and explicit language about “weaponizing” data. The document’s core claim — that hesitation is more dangerous than getting it wrong — is exactly the opposite of the cautious deployment logic that safety researchers advocate. When that tradeoff is made explicitly by a defense department, the downstream effects on how alignment concerns get weighted in procurement decisions could be significant. The Decoder

OpenAI’s CFO Sarah Friar published a “scorecard for the AI age” that frames ROI in terms of useful work completed, cost per successful task, dependability, and return on compute. Practical framing for enterprise buyers, and a sign that OpenAI is actively trying to own the ROI measurement conversation before someone else defines it. OpenAI

Security, scraping, and creator protection

Patreon has stopped relying on robots.txt to discourage AI training scrapers and started actively blocking them, working with Cloudflare to identify and drop bots that harvest creator content without permission. This is a meaningful escalation: robots.txt is purely advisory and widely ignored by training crawlers, while network-layer blocking actually works. The practical question is how quickly bot operators adapt their fingerprinting to evade Cloudflare’s detection. TechCrunch

TikTok is testing an opt-in AI likeness detection tool with some U.S. creators. It requires identity verification through Jumio — a real-time selfie scan and ID check — to enable scanning for unauthorized AI-generated likenesses. TikTok says it does not retain ID documents or facial data. YouTube recently made a similar tool available to all adult users. The identity verification requirement is a friction point that will suppress adoption, but the underlying capability is useful and the trend across platforms is consistent. The Verge

MIT Technology Review flags a rising threat to weather data integrity: deliberate sabotage of the sensor networks and data pipelines that feed global forecasting models. Airline dispatchers, grid operators, and farmers all depend on the same upstream data, and adversarial interference with it — whether corrupting sensor readings or poisoning training data for AI-based forecast models — could cause cascading real-world damage. The piece is light on specifics about attack vectors but the threat model is sound. MIT Technology Review

Quick hits

NVIDIA NeMo Automodel integrates with Hugging Face Diffusers to enable large-scale fine-tuning of video and image diffusion models. Hugging Face Blog
Google-backed FireSat wildfire detection satellites launched, capable of spotting fires that existing satellites miss — timed with heavy smoke over the U.S. and Canada. Ars Technica
Agility Robotics opened a new Digit robot training center in Fremont, California — Tesla’s own backyard. TechCrunch
San Francisco’s city attorney ordered Apple and Google to remove AI nudify apps from their stores, with estimates that both companies collected millions in fees from those apps. Ars Technica
Vertu’s $6,880 AI agent foldable phone reviewed by TechCrunch — the short version is that the price is not justified by the performance. TechCrunch
Index Ventures co-founder Neil Rimer says the wealth AI is generating in Silicon Valley will be redistributed, voluntarily or otherwise. TechCrunch
Simon Willison shipped an LLM cliché highlighter that flags ten common patterns in AI-generated writing. Useful for editors, and for anyone who wants to audit their own output. Simon Willison’s Weblog