Proof
Ten stories that come with the working shown: proofs checked in Lean, charts that carry their SQL, an agent that has to cite its evidence, and a root key you can test before it changes tomorrow.
Ninety open problems, one GitHub drop.
On 6 October OpenAI posted 722 manuscripts from an unreleased internal model, grouped into 372 results that include claimed progress on 90 of the 500 best-known open problems in mathematics. Zvi Mowshowitz collects what mathematicians make of it, from Strogatz and Kontorovich to the advisory group that had asked labs to keep such problems off private models. The practical lesson sits in the error log: 42% of the papers come with Lean proofs, and all three withdrawals so far are in the unformalized rest.
Table 1 · The release
- 719
- manuscripts after three withdrawals, from 722
- 372
- groups of results
- 90
- of the 500 best-known open problems touched
- 42%
- with Lean proofs; none has broken so far
- 1
- sign error behind all three withdrawals
Eq. 1 · Matrix multiplication
The claimed exponent, checked in Lean, against a previous record near 2.37. Steven Strogatz compared the jump to Bob Beamon’s long jump. The same release claims a zero-free strip for the Riemann zeta function and a proof of the Unique Games Conjecture.
The cheap Claude can drive a computer now.
Claude Haiku 5.5 costs about 75% less to run than Haiku 4.5 and scores 72.4% on the OSWorld 2.1 computer-use benchmark, where its predecessor managed 15.7%. It is the first Haiku with an effort setting, and Anthropic pitches it as the subagent under Opus or Sonnet for compaction, summaries and browser work. Hard agentic coding still belongs to the bigger models: 39.2% on Terminal-Bench 4.0, against Sonnet 5.5’s 70.6%.
The model never fetches its own evidence.
Cloudflare’s Managed Defense team published the harness behind its alert triage. Deterministic code collects and versions the evidence before inference starts, four specialist agents must cite items from that package, and application code throws out any citation that is missing, out of scope or doesn’t back its claim. If you are building an agent whose output a human has to sign, this is the shape to copy.
=IF(AND(EXISTS(cite), IN_SCOPE(cite), SUPPORTS(cite, claim)), finding, "limitation")Every lookup lands in one of three states, so a timeout never reads as a clean result.
Charts that show their query. Videos that are code.
Anthropic put Claude Dashboards and Claude Motion into beta on 8 October. Dashboards connects to your warehouse or CRM, builds a dashboard that stays current, and shows the query behind every chart. Motion writes animation code for a short explainer you can edit word by word and export as MP4. Docs, Slides and Design came out of beta on every plan, Free included.
Dashboards
Redshift, BigQuery, ClickHouse, Databricks, Snowflake or Salesforce in; a live dashboard out.
Beta · paid plansThe query
Every chart carries the query that built it, so a number can be checked before it lands in a deck.
Hand-off
Send a dashboard on to Hex, Grafana, Mixpanel, Amplitude, PostHog or Sigma for deeper work.
Motion
A 30-second explainer from a prompt, edited in place, then downloaded as MP4.
Beta · Team & EnterpriseBuilt as code
It animates your own text, charts and images. No generated footage, no AI-generated people.
Out of beta
Docs, Slides and Design on every plan. The standalone Design site closes 14 December.
A push is only as fast as the slowest replica.
Agents commit after nearly every action, and GitHub counted 7.38 billion commits in September, five times as many as a year earlier. Each repository lives as five full copies on fileserver disks and every copy takes part in every write, so adding read capacity makes pushes slower. Brian Celenza lays out the rebuild: the authoritative copy moves to Azure Blob Storage, reads come from cache workers, and only the ref update needs agreement. Internal benchmarks show up to 35× the write throughput.
Now · Spokes
Five full copies on local disk. They are the durability and the read capacity, and all five join every push.
Next · split
Cache workers scale with traffic. Compaction and GC run off the serving path, against storage.
Point Claude at any game you own.
universal-modder bundles agent skills, a CLI and the fal MCP server so Claude Code, Codex, Gemini CLI, Cursor or OpenCode can mod almost any PC game. The agent identifies the engine, reads the game’s real code, builds the mod, generates sprites, 3D and sound, tests it in the running game and cuts a showcase video. @denissexy had it build a GTA-style mod for SimCity 3000 in a couple of hours; the repo passed 6,100 stars ten days after it went up.
Try: “Mod Terraria: add a homing missile launcher. Make the sprites with fal.”
$tracing that survives the bundler█
Sentry’s JavaScript SDK v11 stops monkey-patching module exports through OpenTelemetry loader hooks. It injects diagnostics_channel calls into library source while your app is built or loaded. Isaac Z. Schlueter explains why the old chain broke: it needed a live Node loader and a package.json on disk, and a bundle on Vercel, Netlify or Workers has neither. The same transform now traces Cloudflare Workers, Bun and Deno.
$ git diff instrument - node --require ./instrument.js app.js + node --import ./instrument.mjs app.js - preloadOpenTelemetry() - SENTRY_PRELOAD_INTEGRATIONS=… - skipOpenTelemetrySetup: true + enableOpenTelemetrySetup: false # default on most server SDKs + sentryCloudflareVitePlugin() # Workers + sentryBunPlugin() # bun build only, for now + deno run --preload=@sentry/deno/import # keep instrumented libs out of `external` # check coverage: Sentry.init({ debug: true })
One line fails the build if a page goes dynamic.
Next.js 16.4 recommends Cache Components for every app, switches them on in create-next-app, and will make them the default in Next.js 17. The new ensureStatic export fails the build when a dynamic component ends up in a route you meant to keep static. For existing apps, npx next@canary upgrade --agent=latest hands your coding agent version-specific migration steps.
export const ensureStatic = 'navigation';
// add <UserAvatar /> here and the build fails
Wed 14 Oct: an out-of-band Next.js security release for two Critical and one High upstream vulnerabilities. Plan the upgrade window now.
Nothing that matters lives only in the sandbox.
PostHog runs every cloud agent it ships, from the Slack app to self-driving bug fixes, through one runtime. A Temporal workflow owns the run, a Modal VM with a real kernel and Docker does the work, and snapshots bring a dead box back with the conversation and working tree intact. The safety section is the one to steal: an LLM judge blocked 1,374 suspicious signals in one week, and PostHog still put a network allowlist under it, because a judge eventually waves something through.
- TO
- anyone running agents in the cloud
- FROM
- PostHog engineering
- RE
- five things a run has to do
- CC
- whoever is on call
- Survive its own infrastructure. State lives in Temporal, so a deploy can’t kill a run.
- Take messages at any time. Follow-ups queue in order; re-sent webhooks are dropped by id.
- Be safe. A judge screens inputs; the VM can reach at most 100 allowlisted domains, no raw IPs.
- Be a real machine. A VM with Docker, booted from a warm pool while you type.
- Distrust its own build spec. The repo token is removed before your setup commands run.
The DNS root changes its key tomorrow.
On Sunday 11 October the root zone switches its key-signing key for only the second time, from KSK-2017 to KSK-2024. A validating resolver that never learned the new key will fail DNSSEC checks for names under every top-level domain. On 1.1.1.1 or Cloudflare DNS there is nothing to do; if you run your own resolver, Cloudflare’s post shows how to ask it directly with two queries.
Same algorithm, RSA/SHA-256. KSK-2017 is revoked in 2027 and its private key is scheduled for deletion.
Years between root key rollovers
$ dig root-key-sentinel-is-ta-38696.dnstest.dev A → an answer your resolver trusts KSK-2024 → SERVFAIL it doesn’t: update its trust anchors today $ dig root-key-sentinel-not-ta-38696.dnstest.dev A → SERVFAIL expected when the key is trusted Answers to both: no RFC 8509 support, so the test is inconclusive.
That’s the week.
Issue 124, Saturday 10 October 2026. Ten picks from eight publishers, published between 30 September and 9 October, nine of them since last Saturday’s issue. Two came to us through Telegram channels; OpenAI’s math release is linked through The Zvi’s write-up.
Rubric, in order: AI tools you could adopt this week · creative software and generative media · dev tools and agentic coding · privacy and security where the signal is high · science and research with a practical kernel · anything immediately actionable for a senior engineer or founder.
Also read this week and left out: The Batch, Import AI, OpenAI News, jeffbullas.com, Vercel, @seeallochnaya, @TochkiNadAI, @ProductsAndStartups. Anthropic Engineering and fly.io published nothing new inside the window.