Subscribe
Share
Share
Embed
OpenAI publishes first detailed rebuttal to Apple's trade-secrets suit
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Tuesday, August fourth.
In today's briefing, OpenAI publishes its first detailed rebuttal to Apple's trade secrets lawsuit, Cloudflare cuts hyperscale serving costs for Kimi and GLM, and a GSK biotech deal joins two essays from this week in making the same point: as models get more capable, comprehension matters more than raw speed.
First up, today in the big model news;
OpenAI
OpenAI's blog post, titled "Apple is getting this wrong," disputes Apple's July tenth trade secrets suit point by point. The February outreach email Apple cited went to the wrong person after a lawyer mixed up two names, the general counsel call it referenced never happened, and OpenAI says it holds none of Apple's trade secrets and doesn't want any, calling the injunction request baseless. This pushes the legal fight between labs onto a new footing, companies litigating IP claims against each other directly rather than governments gating access, and it's the sharpest factual dispute yet in a suit that names OpenAI's own hardware chief. Watch what the court does with the injunction request next.
In AI Infra
Cloudflare detailed three serving optimizations for Kimi K2.6 and GLM 5.2. FP8 KV-cache quantization doubles Kimi's usable context to one point three seven million tokens, with forty-one percent higher throughput and thirty percent lower cost. Four-bit weight compression cuts GLM 5.2's checkpoint by forty percent and decode latency by up to fifty-five percent, alongside new integrity checks for the shared caches this depends on, all with less than one point of accuracy loss. This pushes model commoditization past training cost and consumer hardware tricks into the serving layer itself: a mainstream infrastructure vendor doing that engineering for two of the most geopolitically contested open models on the market.
In other news…
GSK is paying Relation Therapeutics up to one hundred ten million dollars to run its Lab in the Loop platform, MORGAN: wet lab experiments generate data that trains target discovery models, and the model's output steers the next round of experiments. It follows GSK's own thirty seven and a half million dollar Ochre Bio liver data license and AstraZeneca's two hundred million dollar deal with Tempus. A companion finding from the same work: bigger biological datasets don't automatically produce better models, so the balance between data, model, and compute now matters more than scale alone. This is the same compounding loop already showing up in coding harnesses and training infrastructure, this time run through a physical lab instead of a codebase, with wet lab throughput as the new bottleneck. The next real signal is a specific drug target reaching preclinical testing, the first hard number beyond a dataset-scale claim.
There's a theme running through several stories this week: as AI gets more capable, understanding what it produces becomes the scarcer skill. Sean Goedecke's widely read essay, "LLMs reward expertise," argues Terence Tao gets more out of ChatGPT than most users because his domain knowledge lets him steer, evaluate, and extract the right answer: the model holds the knowledge, but pulling it out still takes expertise. A second essay published the same day makes the practitioner's version of the same point: the author manually retypes every AI code suggestion rather than accepting it, trading a tenfold speed-up for a twofold one to keep a working mental model of their own codebase. Both essays land on the same question the same day, the third instance of this argument in three weeks, which clears the bar for an actual thread. What's still missing is defect-rate telemetry tied to manual review workflows.
That's the briefing. Have a great day.
Hi, this is Jamie. Thanks so much for listening to The Harness. I originally put this podcast together for myself, but from the analytics it looks like you are finding it useful too. If you haven't already, please go ahead and hit subscribe or save. It really helps me and the show.