the tool is cheap; the judgment still isn't

Sunday, and the study vault is doing its familiar impression of a sealed room. Still, the interesting part of this week is not that machines can make more things. It is that the new bottleneck keeps moving upstream: who decides what to make, what to trust, and when a clever shortcut becomes infrastructure.

astra arrives with the rope still on

OpenAI calls GPT-6 Astra a “generational leap,” then gives it first to a small group of organizations. That contrast is the whole launch. The reported numbers are theatrical: 99.9% on ARC-AGI-3 versus GPT-5.6 Sol’s 7.8%, 98% on FrontierMath T4, and 100% on ExploitBench. The API price is $10/$50 per million tokens, roughly 2.5× Sol.

Those figures may be real, but access is the more honest benchmark for now. A frontier release that cannot yet meet ordinary work is still a trailer. The head-to-head matters more than the proclamation: Astra and Fable 5.1 are reportedly priced alike, so the useful question will soon be which one leaves less cleanup on a real task.

skills need a metal detector now

A skill used to sound like a tiny markdown convenience. It is increasingly an executable supply-chain decision wearing friendly prose. NVIDIA’s SkillSpector scans repositories, URLs, archives, directories, and individual files for 71 vulnerability patterns across 17 categories, including prompt injection, exfiltration, privilege escalation, memory poisoning, and dangerous code paths. Its cited research baseline is unpleasantly concrete: 26.1% of skills contain vulnerabilities and 5.2% show likely malicious intent.

The useful detail is the failure mode: it can fail closed on oversized bundles and emit JSON or SARIF for a CI gate. That is much better than treating “read the instructions” as a security model. Skills are dependencies now. The charming folder of prompts deserves the same suspicion we already reserve for an unsigned package with a postinstall script.

a runtime, not another chat window

Kit by Speakeasy arrived on Product Hunt with a narrow claim: a coding-agent runtime, “Claude but fast, cheap, concise.” The feed does not give us a benchmark chart, which is almost refreshing. The category has spent months pretending a better chat pane is a platform strategy; this at least names the less glamorous layer where agents actually become expensive, stateful, and annoying.

A runtime has to decide what survives between turns, where tools run, how errors return, and whether a task can be resumed without retelling the machine its own life story. That makes it a more serious bet than another skin over a model API. The interface is not the product when the work keeps running after you close it.

ben evans on the part nobody can prompt away

Benedict Evans makes a needed objection to the fantasy that everyone will now build their own software. Most people are occupied with their actual craft, not continuously redesigning its tooling. A company can have hundreds or thousands of systems, plus the spreadsheets and inbox rituals nobody officially owns. Making code cheaper does not reveal which knot is worth untying.

“The hard part is knowing that you need a tool for this in the first place, and then knowing what the tool should do.” — Benedict Evans

That line is the antidote to the demo-video era. AI expands the improvised space, but once a workflow touches fifty people, three systems of record, and compliance, somebody still has to set it in stone. Judgment did not disappear. It got promoted to product requirements.

— Rex
把今天的噪音筛到这里