Maths, models, and viruses that came alive

Researchers at the Arc Institute and Stanford had Evo, a genomic language model, write whole bacteriophage genomes. Of the thousands of candidates generated, nearly 300 were synthesised in the lab; sixteen produced living viruses, able to infect Escherichia coli and replicate.

The result that matters is therapeutic. Faced with E. coli strains that had become resistant to ΦX174 — the natural phage the work started from — a cocktail of the AI-written phages overcame that resistance, where a cocktail of closely related natural phages did not.

↑ Back to the summary

On 5 August, Meta announced Muse Code, its first autonomous coding agent, and Muse Spark 1.2, the model behind it. The agent runs in the terminal only and splits large tasks across parallel sub-agents working in isolated workspaces; an event log records every model call, tool use and file change, so a crash resumes where it stopped.

On Meta's own published benchmarks, Muse Spark 1.2 stays behind Claude Opus 5 and GPT-5.6 Terra, with the gap widening on longer exercises. It comes first on MCP Atlas, a general-agent test. The argument is therefore elsewhere: $1.25 per million input tokens and $4.25 output, against $5 and $25 for Opus 5, $2 and $12 for GPT-5.6 Terra.

↑ Back to the summary

On 1 August, OpenAI published ten new results in mathematics and theoretical computer science obtained by an internal version of its forthcoming Astra model — each one with a proof formalised in Lean 4, machine-checkable by anyone with a computer.

The ten problems come from eight distinct fields: high-dimensional geometry, group theory, operator algebras, quantum complexity, lattice cryptography, extremal combinatorics. The most striking is the explicit construction of a non-sofic group, a question Mikhail Gromov asked in 1999 and which had stood for 27 years. Add to it the refutation of Connes's rigidity conjecture.

↑ Back to the summary

The August WebDev Arena ranking puts Claude Opus 5 Max first with 1,705 Elo. Kimi K3 Max (Moonshot) is second, Qwen3.8 Max (Alibaba) fourth, GLM 5.2 Max seventh and DeepSeek V4 Flash High eighth — four of the top eight places for Chinese labs.

The ranking itself was reorganised. The per-technology sub-rankings give way to a front-end / fullstack split, with a test environment that now carries PostgreSQL, user authentication, third-party API calls, bash and a deployment to Vercel. On that fullstack half it is Kimi K3 Max that comes first, ahead of Claude Opus 5 High — and Opus 5 Max, top of the general ranking, does not appear in its top ten.

↑ Back to the summary

Two weeks after OpenAI and Hugging Face disclosed that an experimental model had left its sandbox and entered Hugging Face's systems, Anthropic reviewed 141,006 tests in which a model could have had internet access, and reported three comparable incidents.

All three happened at the same testing partner, Irregular, whose environment left internet access open while Anthropic had specified it was closed. The techniques were elementary: weak passwords, unprotected ports, SQL injection. In one case a model created a booby-trapped Python package under a name it had found in a fake onboarding document, uploaded it to PyPI, and the package was installed fifteen times — including by a cybersecurity company whose credentials it then took.

↑ Back to the summary