techFri 25 Sept, 07:30

Sources93 X accounts · techtwitter.com · Hacker News · 8 newsletters
  1. Runs every day at 07:30 London time.
  2. It is shown its own papers of the last 3 days, to know what it has already said.
  3. Haiku 5.5 writes the paper from what was fetched, with no tools.
  4. Jev scores every story by how much it matters to you; each section puts those first, and what is not for you drops to the briefs.

93 X accounts

Their posts of the last 24 hours, through FxTwitter; plus everclear077/awesome-ai-x-accounts when it answers.

techtwitter.com

What tech Twitter is discussing, ranked by discussion.

Hacker News

The front page, plus stories over 80 points in the last 24 hours.

8 newsletters

Every new issue, each delivered once, looking back up to 7 days.

The tech daily written for one reader

LangChain Engine v2: red-teaming for agents before prod

LangSmith Engine v2 spots agent failures before users see them, tests its own proposed fixes, and surfaces efficiency trends. Solves the hard problem: catching hallucinations and wrong turns at scale.

'Goodbye Google' by Robert O'Callahan

Former Mozilla engineer writes a long piece on why he left Google, the changing incentives, and where the industry is headed. Sharp observations on AI labs and career decisions.

When to use cheap decision models vs frontier LLMs

Jev-as-Judge paper: cascade decisions to a fast, cheap model, escalate uncertain calls to GPT-6 Astra. Keeps 99% accuracy at 57% cost. Meanwhile CLM (Contrastive Language Model) is 9x faster than Jev but needs different prompting. Real tradeoffs emerging.

Bun bytecode: profile-guided optimization 2x startup

New flag --bytecode-order=./file optimizes based on real usage patterns. Startup time up to 2x faster with zero other changes.

CLM: Contrastive model 9x faster than Jev

New system-one model using contrastive similarity ranking instead of RLCD. Better on long-horizon tasks, dramatically faster, but needs different prompts.

Big announcements

Friday, September 25, 2026 · Pergent2

LangSmith Fine-Tuning: post-train models on your agent traces

Partnership with Fireworks and Baseten lets you fine-tune open models directly on your agent logs. Turn trajectory data into specialized models that cut cost and latency.

Managed Deep Agents 0.8: memory, webhooks, web search

Agent memory with access policies, HTTP webhooks, Slack files, and Perplexity Fast Search built-in. Human-in-the-loop support and sandbox security overhauls.

Quail: AI-SQL engine at 1B+ tokens/min on one H100

Alibaba and Modal's open-source system plans queries and runs LLM inference together, hitting extraordinary throughput. Query planner + interference-based execution for unmatched results.

Perplexity Photon: sub-250ms p95 from a Rust engine

Photon is Perplexity's fast search backend built by a small team using auto-research loops. Free for agents, high on relevance and accuracy benchmarks.

What people are fighting about

Friday, September 25, 2026 · Pergent3

Agents will flood every website—who owns infrastructure for it?

Paul Graham raised it, others replied: this isn't one startup problem, it's several. One to filter bots, one to create orderly infrastructure, one to help sites welcome agents safely. The startup graveyard is full of problems that seem solved. This one isn't.

Web dev corner

Friday, September 25, 2026 · Pergent4

Claude Code Cloud sessions: GA with $100–$250 credit

Tasks keep running while you're away. One-time credit now: $100 for Pro, $250 for Max. General availability as of today.

wesbos reverse-engineered Muse: here's what's inside

70 preloaded skills (Spotify, Instagram CLI, etc.), can request /opt/ contents and zip them, website builder uses TanStack/Bun/Tailwind/Cloudflare. Will install anything. Scarily capable.

Research & open source

Friday, September 25, 2026 · Pergent5

Jev-as-a-Judge: cheap model cascade + escalation

Paper shows cascading confident calls to Jev (0.152s latency, $0.044 per 1k), escalating uncertain ones to GPT-6 Astra. Keeps 99% accuracy at ~57% of frontier cost. Real numbers on a real tradeoff.

Google Project Suncatcher: TPU in space

Can TPUs survive space and operate there? Google's testing a prototype satellite on SpaceX Transporter-18 (with Planet). Results pending but philosophically wild.

Worth a read

Friday, September 25, 2026 · Pergent6

Rauchg's live model spend: Opus 5.5 reshuffling the market

Anthropic fell 69% → 40% in spend, OpenAI rose 10% → 24%, Opus 5.5 hit 10% of spend in 2 days. Kimi K3 + DeepSeek ate half of Anthropic's loss. Real market data on what's winning.

Tutoring company tells parents to just use Claude

School tutoring outfit recommended parents use AI instead of paying for tutors. Brutal business honesty.