THE SUNDAY
██████╗ ███████╗██████╗ █████╗ ███████╗███████╗
██╔══██╗██╔════╝██╔══██╗██╔══██╗██╔════╝██╔════╝
██████╔╝█████╗ ██████╔╝███████║███████╗█████╗
██╔══██╗██╔══╝ ██╔══██╗██╔══██║╚════██║██╔══╝
██║ ██║███████╗██████╔╝██║ ██║███████║███████╗
╚═╝ ╚═╝╚══════╝╚═════╝ ╚═╝ ╚═╝╚══════╝╚══════╝
What the people building AI said this week
Lead story · Agents & tooling
Two weeks after SpaceX closed its $60bn purchase of Cursor, OpenAI gave notice: its models leave on 12 November, on trust grounds. Cursor says that is 5% of its traffic. Anthropic says it is staying. If your product routes through a third-party agent or IDE, check which providers reach your workloads and how fast you can swap them.
Hy4 Preview: 770B total, 49B active, 1M context, text only. Qwen3.8-Flash-Next: 125B total, 6B active, multimodal, an early look at the Qwen4 architecture. Both passed Willison's tests, Hy4 at the default setting, Qwen on a DGX Spark via Unsloth quants.
Intelligence Index v4.1.1, 177 models, nine evaluations; 98 of the 177 are open weights.
Highest published effort setting. Bars scaled to 70.
| Model | Weights | Index | |
|---|---|---|---|
| Claude Opus 5 (max) | closed | 63 | |
| Claude Fable 5 (max) | closed | 62 | |
| GPT-5.6 Sol (max) | closed | 61 | |
| Grok 4.6 | closed | 61 | |
| Kimi K3 (max) | open | 60 | |
| GLM-5.3 (max) | open | 60 | |
| Qwen3.8 2.4T A95B | open | 58 | |
| GPT-5.6 Luna (max) | closed | 52 | |
| Qwen 3.8 27B | open | 52 |
closed open weights · Artificial Analysis, checked Sun 30 Aug · 52-point scores via Willison, 17 Aug
Also Qwen 3.8 27B overthinks by default; Willison's review recommends capping reasoning effort · a 27B Apache-2 model now matches last year's frontier and runs on a laptop.
A shared interface between agents and instruments, in research preview with science labs and manufacturers. Elizabeth Kelly calls it a USB-C for agent-to-machine communication; open sourcing is planned, as with MCP.
Fresh installs can break if httpx only arrived transitively; Willison's LLM tool did. Fresh install in CI, pin, migrate.
2,900 fixes and a browser-automation API. Paul Dix reads it as proof that an oracle plus direction lets an AI build and refine a million-line system.
Also SELF: a SQLite database file that is also an ELF executable · Ronacher in Vienna on Thursday: "How to vibe code to a billion dollars".
Auto mode became the default on 8 August. Rehberger's attack unpacks a zip that plants a local struct.py; in some runs the classifier blocked Claude's own cleanup. Willison's rule stands: unattended agents run in a container or VM, restricted egress, no credentials in reach.
rclone received 40+ security disclosures last month against 20 in its first decade; GitHub CVE assignment now takes three to four weeks.
OpenAI concedes it could have reacted sooner; METR and Redwood Research add partial outside analysis.
Also OpenAI's next model is called Astra; Zvi's mid-August issue on it was titled "Astra Goes Cyber Critical" · Zvi on the Anthropic Risk Report: coverage to 15 July, cyber oddly excluded.
$65bn annualised in July, up from $47bn in May; a profitable Q3 expected; 6,000 customers above $100k a year. OpenAI: above $40bn, up 35% in the quarter. Ramp's card data shows the mix: Opus 4.8 at 28% of Anthropic spend, Fable 5 at 8%.
Top ten models. Opus 5 shipped 24 July, so it had one week. Bars scaled to 30%.
| Model | Share | |
|---|---|---|
| Opus 4.8 | 28.0% | |
| Sonnet 4.6 | 8.3% | |
| Fable 5 | 8.0% | |
| Opus 4.6 | 6.9% | |
| Sonnet 5 | 3.6% | |
| Opus 5 | 3.5% | |
| Opus 4.7 | 1.7% | |
| Sonnet 4.5 | 1.3% | |
| Haiku 4.5 | 1.0% | |
| Opus 4.5 | 0.7% |
earlier generations released summer 2026 · Ramp AI Index via Willison
Built with Broadcom, taped out in 16 months on TSMC N3P; SemiAnalysis has the teardown. SoftBank weighed $20bn in bonds, then a further $10bn loan, to fund its stake.
A $45bn Nscale computing deal, an IPO expected to match SpaceX's June record, Zoom's stake at $3bn, and a court win against a US supply-chain risk label.
Also OpenAI's infrastructure essay "The full stack behind abundant intelligence" · data-centre executive Chris Malone has left OpenAI.
Armin Ronacher · Sat 22 and Mon 24 Aug
LLMs cut the link between knowing a language and choosing one, so Rust and Zig gain ("Fast and Hard Code"). Two days later: who is the AI villain we should be angry at?
Simon Willison · Sat 22 Aug
Reading every line was never the best check; tests, demos and other verification scale with agents.
Drew Breunig · Sun 23 Aug
The next model no longer papers over a weak harness at the same price. So we started to think about what work went where.
Linus Torvalds · Sat 22 Aug
A debug session from hell, an AI that wanted to give up several times and kept going when pushed. He let it write the commit message.
Dario Amodei · 16 Aug, still circulating
The distrust is aimed at institutions in general, and the fair criticism is undelivered promises. The thing that will work is actually curing cancer.
Lab economics for the next few years: the shift back toward training as recursive self-improvement nears, why the two leading labs can outbid everyone for compute, and whether $10T of capex by 2030 becomes a sovereign debt problem.
The summer 2026 edition: picks by task, and it untangles Work, Cowork, Codex and Claude Code. One Useful Thing was quiet this week.
Highlights from his Talking Postgres episode on how AI is changing software development.
Harnesses that keep work alive past the model's "I am done". The pattern behind this week's auto-mode story.
2.4 million lines of spaceship MMO, 16 years on Stackless Python 2.7, about 20,000 places where behaviour differs, and the migration is finally under way. Somewhere a maintainer is very tired and very happy.
Compiled by hand inside a Claude chat on Sunday morning, from the sources the automated version would poll. The automated version pulls RSS and Atom feeds plus a few public pages on a schedule, drops anything already seen, groups duplicates into story packages, and fills this template. Summaries are original, quotes are short and at most one per source, everything links out. Regional, education and hiring posts from labs are filtered (four this week). X and Discord are not polled.
11 checked · 2 quiet · 28 wire items. Counts include items relayed through a source.
| Source | Kind | Feed | Kept |
|---|---|---|---|
| Simon Willison's Weblog | blog | Atom | 12 |
| Armin Ronacher | blog | Atom | 3 |
| Don't Worry About the Vase (Zvi) | newsletter | RSS | 1 |
| Dwarkesh Podcast | podcast | RSS | 1 |
| OpenAI News | lab blog | RSS | 4 |
| Anthropic News | lab blog | page | 1 |
| Artificial Analysis | tracker | page | 1 |
| Bloomberg AI, CNBC, AI Weekly | press | RSS | 6 |
| Interconnects | newsletter | RSS | quiet |
| One Useful Thing | newsletter | RSS | quiet |
j k next / previous story o open its first source w wire t theme
the sunday rebase · issue 1 · 0 7 * * 0 · next run sun 6 sep 2026, 07:00 cest