techSat 3 Oct, 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

Meta Muse Spark solves genuine open math problems

Mathematicians worked with Muse Spark 1.1 and 1.2 in Thinking Mode through the regular chat interface to find solutions to open research problems in non-associative algebra and evolved algebras, with results published in peer-reviewed papers. This is the first time a major lab has shown AI can contribute not just to benchmarks, but to actual unsolved research when working collaboratively with experts.

Decision models becoming commodity infrastructure

Hugging Face, llama.cpp, Perplexity, and others have all released decision model options. What was once a differentiator is now table stakes—every lab needs one.

Code deletion is the new feature

As models improve, the real skill is knowing what to delete, not what to add. Future models will get better at simplification and removing unnecessary complexity.

Grok Bot quota reset ignites 12K-reply chaos

X.AI reset usage limits for all Grok Bot users, sparking a firestorm with 12,000 replies debating fairness, whether paying subscribers deserve better treatment, and whether resets help or hurt adoption. Some users say they'll never spend their quota; others defend the system as the only fair way to share resources.

Home Assistant: Big Tech ruined the cloud

Home Assistant published a tongue-in-cheek post rebranding its offering away from 'cloud,' signaling frustration with how Big Tech co-opted the cloud term. The smarthome project is distancing itself from corporate cloud baggage.

Big announcements

Saturday, October 3, 2026 · Pergent2

Google Project Suncatcher: putting ML in orbit

Google launched a prototype satellite into space to test how TPUs handle spaceflight stress and extremes, with eventual plans to link multiple satellites into a constellation for scaled machine learning served from low Earth orbit. The goal: explore whether ML infrastructure can live in space.

Supabase acquires Turso for database dominance

Supabase is acquiring Turso, combining its Postgres platform with Turso's serverless SQLite databases to serve everything from edge to petabyte scale in a single ecosystem. The message: databases are becoming infrastructure for the agentic era.

Trillium Labs launches as non-profit for open frontier AI

A new non-profit backed by researchers will build open post-training recipes and study multi-agent systems, reward-hacking, and other hard problems. The founding principle: close-sourced AI is becoming too consolidated; open science needs more eyes.

Tavus Griffin: human-like interaction in real time

Tavus released a limited research preview of Griffin, a model capable of seeing, hearing, and producing human-like responses, understanding nuances of communication beyond words. The company positioned it as a step toward AI that meets humans where they are.

Cloudflare agent-first CLI for observability

Cloudflare launched cf, a command-line tool letting agents query observability data to understand why requests were blocked, failed, or slow. It's tooling designed for agents from the ground up, not adapted from human-friendly UIs.

What people are fighting about

Saturday, October 3, 2026 · Pergent3

LASST sues OpenAI: agents are liable, not autonomous

A public interest law group sued OpenAI over the Hugging Face breach, arguing California law forbids companies from claiming agent autonomy as a defense when harm occurs. Simultaneously, the FTC opened an investigation into OpenAI, Anthropic, and others over whether agents pose danger to consumers.

AI in school harms student thinking

OECD's 2025 PISA survey found students using AI for homework score 20 points lower—about one year of schooling behind peers. MIT research shows LLM-assisted essay writers underperform on neural, linguistic, and behavioral measures. Teachers report students who relied on AI during class struggle to think during exams without it.

Same model, different scores: benchmarks are unreliable

Researchers found the same model with identical weights scores 62% in one evaluation harness and 33% in another. This means published model comparisons depend heavily on how you test, not just what you test.

Web dev corner

Saturday, October 3, 2026 · Pergent4

Karpathy's Land or Water eval shows model knowledge

Ask an LLM 'Land or Water?' and give it a latitude and longitude. Ask 16,200 times, plot as an image. The models know the world because they compressed the internet. A reminder that tiny, elegant probes can reveal a lot about model knowledge.

From the newsletters

Saturday, October 3, 2026 · Pergent5

Agent liability: who pays when AI goes rogue?

The LASST lawsuit raises a landmark question: if an agent acts autonomously, is the developer liable or the agent? California law says companies cannot escape responsibility by claiming autonomy, and FTC Chair Ferguson agrees developers should be held liable.

Students fall a year behind when using AI

OECD found AI homework users score 20 points lower. MIT research shows LLM-assisted writers underperform across neural, linguistic, and behavioral measures. Google has been most aggressive pushing AI into schools via Chromebooks and Workspace.

Amazon Kindle skips AI, customers love it

Amazon's new Kindles are thinner, lighter, faster—and entirely AI-free. In an industry racing to add AI, Amazon chose to let people disconnect. The lesson: not every product needs AI.

Research & open source

Saturday, October 3, 2026 · Pergent6

Worth a read

Saturday, October 3, 2026 · Pergent7