Good morning. The Apple-versus-OpenAI story escalated overnight from courtroom drama to legal siege, with Cupertino sending document retention letters to dozens of OpenAI employees and a trade secrets lawsuit landing at the worst possible moment for OpenAI’s IPO. Elsewhere, Mozilla put out an open-source AI report that a lot of readers think an AI wrote, and Isomorphic Labs is making some very big claims about the successor to AlphaFold 3.
Apple turns up the pressure on OpenAI. The Financial Times reports that Apple has sent document retention letters to dozens of OpenAI employees — many of them former Apple hardware people — as part of its widening trade secrets case. The HN thread is split: some ex-Apple lawyers call hold letters routine, while others read Apple’s willingness to escalate as a sign it has hard evidence. The nuclear option quietly hanging over things is the App Store; Apple pulled Fortnite once, and ChatGPT is not immune.
The timing is what makes this ugly for OpenAI. TechCrunch’s Equity podcast notes the suit names OpenAI’s chief hardware officer and claims over 400 former Apple employees now work there — right as OpenAI is reportedly targeting a late-2026 IPO. The Vergecast is more skeptical of Apple’s specific allegations, with legal experts saying much of what’s cited looks like standard industry poaching. Whether Apple wins in court or just wins by dragging things out through discovery, the hardware roadmap gets harder either way.
Kimi K3, tested. Simon Willison ran K3 through his pelican-on-a-bicycle test — the whole run cost 25 cents thanks to 13,241 reasoning tokens on a single SVG — and the HN discussion is a good read on the benchmark’s limits. One commenter noted the obvious problem that pelican-bike prompts are almost certainly in training data by now, and someone else pointed out the benchmark says nothing about agentic tool use, which is what actually matters for how these models get used. Consensus estimate: K3 is roughly three months behind GPT-5.5.
Mozilla’s state of open source AI, apparently written by AI. Mozilla’s report argues open-weight models have closed the practical capability gap, inference costs have dropped ~50x in three years, and open models now handle most production tokens — with OpenRouter data showing open-model volume grew nearly 5x in four months to overtake closed models 63/37. Good data. The problem, as HN commenters quickly noticed, is that the prose reads like “an LLM’s idea of a CTO presentation,” and a Pangram check flagged it as machine-generated. Advocating for thoughtful open AI development in copy your executives clearly didn’t write is a choice.
Isomorphic Labs claims a big leap past AlphaFold 3. DeepMind’s drug discovery spinout announced IsoDDE, a drug design engine they say more than doubles AlphaFold 3’s accuracy on protein-ligand structure prediction, predicts binding affinities faster than physics-based methods, and can identify novel binding pockets from sequence alone. The HN response is a wall of skepticism: no peer review, no public access, no technical detail on what actually drives the gains, and pointed reminders that AlphaFold 3 didn’t meaningfully outperform older methods on CASP16’s protein-ligand benchmarks. Interesting if true, and we might not get to find out either way.
Kaiser nurses push back on call-center AI. Local News Matters reports that Kaiser Permanente call-center nurses say AI-driven surveillance — including tone and empathy scoring, plus handle-time tracking — is pressuring them to keep patient calls under 15 minutes even during potential suicide crises. Kaiser disputes the metrics claim and says the empathy-scoring pilot was scrapped in 2024. Several HN commenters argued the article conflates old-school call-center metric abuse with anything specifically AI, though others confirmed similar tooling is spreading at UHC.
A Kaggle/DeepMind AGI hackathon judged by AI, apparently. A Kaggle discussion post lays out inconsistencies in the judging of a Kaggle/Google DeepMind AGI hackathon, and the HN thread turned into a broader indictment: AI-generated submissions being scored by AI judges, with at least one allegation of prompt injection inside a winning project. A Kaggle PM showed up to acknowledge the concerns. It’s a small story with an unflattering shape — the hackathon nominally about measuring AGI couldn’t reliably measure its own contestants.
That’s the wrap. Watch to see whether Apple’s retention letters produce any actual named defendants in the coming weeks — that’s the point at which the IPO conversation gets genuinely awkward.