Sun 09 Aug 26
Five picks, and four of them hand you something that runs: a smaller model that outscores its own flagship, an API that returns a live preview URL, and a disclosure curve nobody's patch cadence was sized for.
Same architecture. The small one wins.
DeepSeek changed the post-training and left the architecture alone, and the result is a Flash model that outscores the Pro tier sitting above it. V4-Flash-0731 is a 304-billion-parameter mixture of experts with a million-token context, a speculative-decoding module called DSpark, and three reasoning effort levels you pick per request. Weights are on Hugging Face under MIT, in BF16 with FP8 quantisation supported.
- Parameters
- 304B MoE
- Context
- 1M tokens
- Cybergym
- 76.7
- DSBench full-stack
- 68.7
- Licence
- MIT
Two bugs in a week, from footage nobody was going to watch.
Replay Vision puts scanners on top of PostHog session recordings. Each scanner does one job: monitor answers a yes-or-no question about behaviour, score ranks sessions by frustration, classify tags what the user did, summarise writes it up in your product's own vocabulary. PostHog pointed it at their own app and it found a button that looked clickable and wasn't, plus a search returning false "no results" from a caching bug. Both came back as pull requests the team reviewed and merged. Reachable from Slack, the desktop app, the web app, or the PostHog MCP server.
Send a prompt. Get back a running dev server.
The rebuilt v0 API takes a prompt, generates an app, boots a dev server inside a Vercel Sandbox and hands you a preview URL you can drop into an iframe. Chats hold the app's state, so a follow-up message continues from where the app is now rather than from the original prompt. The agent's work streams back as parts you can render: text, thinking, file edits, commands. Call it synchronously, queue it and poll, or stream. There's a TypeScript SDK, an MCP server, AI SDK tools, and a single call that deploys the result to Vercel.
Get started
pnpm add v0@latest
const chat = await v0.chats.create({ message })
chat.demo // → live preview URL
2,500 critical CVEs, in July alone.
Epoch AI tracks the high- and critical-severity CVEs that 21 major vendors — Microsoft, Google, Apple, Adobe and the rest — disclose each month. July 2026 landed near 2,500. That is roughly five times the monthly record set before Anthropic announced that Claude Mythos Preview could find and exploit vulnerabilities on its own, and 60% above June. The old baseline was about 490 a month. It is not going back to 490, and most patch calendars were sized for that number.
Source: Epoch AI, 31 Jul 2026 · 21 tracked disclosing organisations
Keep your OTel spans. Point them at the error tracker.
Sentry's new OtlpIntegration does three things: it configures an OTLP exporter aimed at Sentry's ingest endpoint, it reads the active OpenTelemetry trace context, and it stamps trace and span IDs onto every Sentry event — so an exception lands inside the trace that produced it instead of beside it. Setup is one integration line and two environment variables. The limits are stated plainly: errors still need a Sentry SDK, because OTLP carries traces and logs only, and you can't interleave Sentry and OTel spans in the same process.
That's today.
Date 09 August 2026 · Zürich
Issue 108 · five stories
Re Sources that fed today's picks
Hugging Face · The Batch · PostHog · Vercel · Epoch AI · Sentry Engineering ·
@seeallochnaya
Also read and passed over: Cloudflare, GitHub, OpenAI, Anthropic, Next.js,
Fly.io, Import AI, Don't Worry About the Vase, Jeff Bullas, @TochkiNadAI,
@denissexy, @rvnikita_blog, @ProductsAndStartups. Every source responded.
The rubric, in order: tools you could adopt this week, creative software, dev tools and agentic coding, privacy and security when the signal is high, research with a practical kernel, and anything a senior engineer or founder can act on tomorrow. Nothing published more than thirty days ago. No story that has run here before.