The Signal — August 19, 2026

Three releases this week, and the gap between them is the interesting part. Two hand over something real: a compiler, a 17GB model file. The third hands over a claim about what a model can do in a wet lab, with the lab's own numbers as the only receipt.

Modular open-sourced the Mojo compiler

Modular released the full Mojo compiler and toolchain under Apache 2.0 with LLVM exceptions on August 18, closing a promise the company first made in May 2023. The standard library went open earlier, followed by hundreds of thousands of lines of GPU kernel code, but the compiler itself stayed shut. Mojo hit 1.0 with source stability last week, and now the whole thing sits in the modular/modular repository, buildable from source with a single Bazel command.

Mojo is a bet that you can design a new general-purpose language good enough to program GPUs and AI accelerators directly, rather than layering Python over CUDA forever. That bet is only testable if people outside Modular can read the compiler, and until this week they couldn't.

The timing carries some weight. Qualcomm completed its acquisition of Modular on July 29, which means a chip vendor now owns an open compiler aimed at accelerators it does not make. Apache 2.0 with LLVM exceptions is about as permissive as language licensing gets, so whatever Qualcomm's plans are, the code is out and can't be pulled back. Whether the project keeps its momentum under new ownership is a different question, and one nobody can answer from a license file.

Sources: Modular · Simon Willison · Phoronix

Anthropic says Claude designed protein binders that beat the published record

Anthropic published two experiments on August 18 that it presents as evidence Claude can speed up laboratory science. In the first, Claude (running as Mythos Preview and Opus 4.8) designed protein binders from scratch against 15 targets and got a working binder for 14 of them. Between 22% and 35% of individual designs bound successfully, which Anthropic contrasts with the 10-15% hit rate it describes as typical for protein design campaigns. Some designs bound several times more tightly than the best previously published result for their target.

The second experiment is smaller and, in a way, more legible. Anthropic handed Claude Opus 5 a contract lab's raw NMR and LC-MS files with a two-sentence prompt. Claude returned finished analyses in 23 and 19 minutes that matched the lab's own hydrogen counts and its purity measurement, 96.4% against the lab's 96.33%. That is routine work a chemist does constantly, and shaving it down to twenty minutes with a generally available model is a concrete claim you could check against your own instrument files.

Both results come from Anthropic, about Anthropic's models, with no independent replication attached. Protein binding assays are exactly the kind of experiment where setup choices move the numbers a lot, and "typical hit rate" is a moving target that depends on which campaigns you count. The company says it plans an access program for scientists, which is where this gets falsifiable. Until outside labs run the same targets, the honest reading is that Anthropic has published a strong internal result and invited people to check it.

Sources: Anthropic Research · Techmeme

Qwen 3.8 27B fits in 17GB and spent 21 minutes drawing a bicycle

Alibaba's Qwen lab shipped Qwen 3.8 27B on August 14, an Apache 2 licensed model with vision support, and Simon Willison's hands-on writeup two days later is the most useful thing published about it so far. His summary: the capability is real, the defaults are ridiculous.

Running the 17GB Q4_K_M quantized build on a 128GB M5 Max MacBook Pro and an NVIDIA DGX Spark, he hit LM Studio's default 8,192-token context ceiling almost immediately, because the model ships with reasoning effort set to xhigh and will happily think its way through a mundane request until it runs out of room. Loading the full 262,144-token context fixed that. It also produced the number worth remembering: one SVG drawing took 21 minutes and 22,276 reasoning tokens to generate 3,223 tokens of output. Qwen exposes a reasoning_effort parameter with lower settings, so this is a default choice rather than a hard limit, but defaults are what most people run.

Qwen's self-reported benchmarks claim the model beats both its predecessor and the closed-weight Qwen 3.7-Plus, one of the lab's strongest models of any size as of May. Those are vendor numbers and independent evaluations aren't in yet. The part that does not depend on any benchmark is the file size. A long-context, tool-calling, vision-capable open weights model at 17GB runs on hardware people already own, and that keeps being true regardless of where it lands on a leaderboard.

Sources: Simon Willison · Hugging Face · VentureBeat

On the Editor's Desk

AWS released Dogwood this week, an open source policy language extending Cedar so rules can cover sequences of agent tool calls instead of one request at a time. It's the kind of unglamorous plumbing that decides whether agents are safe to deploy, and we held it only because the source trail is still thin outside AWS's own blog. Expect it back once someone outside Amazon has used it.

We also looked at the FDA's discussion paper on regulating generative AI medical devices and METR's research note asking whether AI has actually accelerated scientific discovery. Both are legitimately interesting. Both overlap heavily with ground we covered in the last week, and rerunning them today would mostly be repetition. The Google purchase of Spirit Airlines data and the Axiom Math proof verification both ran in yesterday's edition.