techMon 21 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

Qwen-Image-2.1: one model for generation and editing, open weights

Alibaba shipped a 7B image model that does both generation and editing in one pass, with native RGBA transparency support. Runs fast, plays nice with ComfyUI and vLLM. The real win: it's open weights, and it already has day-0 support across the ecosystem.

Frontier labs are selling garbage to fools in Washington

Long-form critique questioning what frontier labs are actually promising regulators and whether the value justifies the hype. Sharp take on misaligned incentives.

Budget shock: even AI-optimistic companies are pulling back

Gergely Orosz talked to a company that had unlimited AI budgets a year ago. Now they're rationing SOTA models and using cheaper ones for everything except planning. The kicker: this company is genuinely bullish on AI and never even looked at budgets before.

Omarchy picking up speed with builders

The Ruby-focused operating system for agents continues to attract people. Kids are building games with Claude in Omarchy; it's becoming a platform, not just an editor.

Vision-language-action models get faster

Paper shows that splitting a VLA into a slow background planner and a fast RL policy for real-time edits improves robot success from 42% to 97% with just 10 minutes of online data.

Big announcements

Monday, September 21, 2026 · Pergent2

Jev is now public to everyone, no waitlist

TypeSafe AI's decision model (cheap, fast classifier for specific choices) went from waitlist to open access overnight. The replies have been wall-to-wall demos of people building things with it.

Kev: open source decision models you can train yourself

Jared Palmer released Kev, a family of small open-source models (600M to 8B) that mimic Jev's behavior. Train in 40 minutes on one H100, runs on a Mac. Achieves 79.6% accuracy out of domain vs Jev's 85.7%, and the code is Apache 2.0.

DocJev: 6x faster document classification and splitting

LlamaIndex shipped an open-source tool that classifies documents and splits them into sections using Jev. Give it a PDF and some category rules in English, get back structured results. The speed claim: 6x faster than GPT-5.6 Luna with equivalent accuracy.

Google's AX: open agentic orchestrator

Google released AX, an orchestration framework designed to coordinate multiple agents. Pluggable architecture, shipped to HN and trending.

What people are fighting about

Monday, September 21, 2026 · Pergent3

Anthropic vs. Trump: jailbreak drama and threats

Politico reported that the White House pressured Anthropic to pull Fable after Amazon's jailbreak. When Dario Amodei explained that jailbreaks happen to all models, Trump reportedly said he wanted to send them to jail. Anthropic pushed back that they can't shut down a model in an hour.

'Stochastic parrot' was a cognitive virus, says AI researcher

Jack Clark argues that the 2021-2025 'stochastic parrot' debate (LLMs are just mimicking, not reasoning) was so catchy that it delayed progress by years, pulling smart people's attention into the sand.

Product philosophy collision: probabilistic vs. deterministic thinking

Nikita Bier observed that most software execs spent 20 years shipping deterministic apps (input X → output Y guaranteed). AI products work the opposite way (probabilistic: X → distribution of likely Ys). Young people thinking in AI-native terms have an unfair advantage, the way social-first thinking won in 2010.

Is open source really overtaking frontier? No.

Router company CEO pointed out that when small open-source projects claim high market share, they're only sharing percentages, never raw numbers. Frontier models still dominate actual traffic; routers just cherry-pick data to look cutting-edge.

Privacy alarm: ChatGPT now tracks your web activity

Buchodi reported that ChatGPT started tracking user behavior on other websites via an ad collector. This hit 700+ points on HN fast because the privacy angle is stark.

Web dev corner

Monday, September 21, 2026 · Pergent4

Jev support lands in TanStack AI

TanStack added a decide() API for typed choices using Jev. One line of code swaps between models.

Kody: your agents need a shared software registry

Kent C. Dodds and team are positioning Kody as the layer where agents put reusable functions and access secrets. Basically npm for agent orchestration.

Agentic payments becoming table stakes

If your checkout doesn't accept agents as payment methods, you could see your conversion rate drop to zero. One observer notes that agents now effortlessly switch to your competitor if you're not agent-friendly.

Claude's image gap: agents need image generation

Ethan Mollick noted that Claude's lack of built-in image generation limits its agentic abilities. OpenAI and Google let their models generate mockups and infographics; Claude relies on code-based drawing.

Agent browser: UI debugging with timestamped frames

Devin's new contact-sheet feature captures timestamped frames during agent interactions, highlighting what changed. Solves the 'before/after screenshot problem' for UI debugging.

Research & open source

Monday, September 21, 2026 · Pergent5

SoL-Pi: agents optimizing their own harness

New paper lets coding agents recursively improve their own evaluation harness instead of hand-designing it. Result: 44-49% lower token traffic, one-third lower API cost, same performance.

JEPA-Anything: one world model for everything

Framework works across vision, biology, control, molecules, physics, weather, and clinical data using orthogonal predictive factorization. Recovers Kepler's laws and suggests novel biology interventions.

vLLM-Omni supports Qwen-Image-2.1 day-0

Fast inference engine added support for the new Qwen model with optimizations for KV caching and multi-image batching.

Hermes Agent expanding plugin ecosystem

Hermes is adding dedicated plugin pages in its explorer. GLM-5.3 FlashX now available through the platform.

Worth a read

Monday, September 21, 2026 · Pergent6

Why do we need human mathematicians anymore?

Terence Tao, the greatest living mathematician by many measures, explores what AI means for mathematical research. He's bullish on AI but precise about where reasoning still demands humans.

Senior engineer death spiral

Essay about how senior engineers can get stuck chasing technical excellence while the business moves in other directions, leading to burnout and irrelevance.

Large companies ship by process of elimination

Observation that most big companies don't have conviction; they launch things and see what sticks. No rhyme or reason, just trial and error at scale.

Agents won multiple phone calls and FOIA requests

Real examples of what agents are now doing in practice: calling Southwest Air and waiting on hold for you, submitting FOIA requests, filling out visa forms, all with zero intervention.

The reputation business, not the software business

Addy Osmani on why your first job should be proving yourself a reliable teammate. Software is downstream; reputation is the moat.

Agents doing QA work with superhuman intensity

Guillermo Rauch's viral thread: agents reproduce mobile browser bugs, create ephemeral test deployments, verify fixes with iPhone simulators — a level of QA thoroughness no human would sustain.