techSun 1 Nov, 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

Bun

Bun AoT cuts memory

In one test, an Express REST API also used 14% less CPU and had no JIT warm-up.

Bun, a fast JavaScript runtime, has a new ahead-of-time mode that compiles code before it runs. Its creator says big apps and command-line tools like Claude Code gain the most, while Fastify saw smaller gains and ran 4% slower.

Model quality

OpenCode checks DeepSeek

The team will shut off two low-traffic setups that had problems.

This follows up Oct 31's reports that DeepSeek felt weaker through OpenCode. The team says the benchmarks it could find all landed in the same range overall, apart from those two setups.

Agent harnesses

Build your own harness?

It is a reply to the debate steipete started on October 31.

A reply reposted by Harrison Chase (@hwchase17) argues that a great harness is a core part of building a great product, not a side task. It points to prompts, custom tools that turn a workflow into one call, and tuning when old messages get summarized to save space.

Agent memory

Agent memory gets routed

A shared history gave only an average gain of +0.015 when tasks were mixed.

Agents learn from past runs by storing history. This paper's Routed Continual Learning groups past runs into clusters, each with its own memory, and an LLM router picks the right cluster before each run.

Agent safety

A gate for agent tools

The gate checks each tool call before it runs, and the models and tools stay the same.

IBM's JevAct sits between an agent and its tools. An LLM turns the company's policy into code checks and yes/no questions, and a failed call goes back to the agent with feedback.

Open models

Five small models beat Fable 5

The post reached 190 points on Hacker News, with 95 comments.

The Memdoor repo claims a team of five Qwen3.8-27B models beats the rival Fable 5 on LiveCodeBench Hard, a tough coding test. Check the test setup before trusting the claim.

Big announcements

Sunday, November 1, 2026 · Pergent2

Hacker News

Hackertube: YouTube built from HN

The HN post hit 1,118 points and 331 comments, the top story of the day.

Karpathy launched Hackertube, a YouTube-style site where every video comes from what Hacker News readers recommend. Swyx says Karpathy built it with Claude Opus 5.5 set to medium effort, a setting that controls how hard the model thinks.

OpenCode

Hex: built almost entirely by AI

The team says Hex works across your whole computer and is not built on Electron.

The OpenCode team says Hex was fully vibe coded, including the app, landing page, animations and sound effects. DHH has added it as a dictation option on his Omarchy Linux setup.

What people are fighting about

Sunday, November 1, 2026 · Pergent3

Cost of software

AI bills now top payroll

The team says its 25 people now produce about what a team of 75 did before AI.

The OpenCode team says its token bill is about double its payroll. It argues the 'cost of software goes to zero' idea is not true yet at today's token prices, but still calls the trade a good one.

Anthropic

Anthropic blocks training talk

A short post says the rule covers anything touching training larger models, and it hit 313 points on HN.

A post by @jiahaoshao1 says Anthropic now blocks some users over anything that touches training larger models. The post gives few details on who is hit or how the rule works, and the 180-comment HN thread is where the reactions are.

RAM shortage

Who is buying the RAM?

Orosz's post drew 2,076 likes and 139 replies, one of the day's bigger fights.

Gergely Orosz says Sam Altman blamed people who buy RAM for home PCs, so they can run AI models locally, for the shortage. Orosz says the real buyers are AI labs, cloud firms and startups buying in bulk.

Web dev corner

Sunday, November 1, 2026 · Pergent4

OpenAI security

ChatGPT account takeover claim

The write-up reached the HN front page with 322 points.

A post titled 'Do the math' says its authors found a gap at OpenAI that let them take over ChatGPT accounts. Treat it as a fresh report, and watch for OpenAI's reply.

Infostealers

Malware ranks victims with AI

Talos, Cisco's threat research team, found the trick in the Volta infostealer.

Infostealers steal passwords and logins. This one hands victims to an AI model that scores how valuable each one is, so attackers go after the richest.

Linux security

Linux TCP bug scores 9.8

CVE-2026-98432 is a use-after-free bug in IPv4 TCP, the code that moves network data.

A use-after-free bug lets a program keep using memory after it was freed, which attackers can abuse. The write-up rates this one 9.8 out of 10, and the HN thread has 215 comments.

From the newsletters

Sunday, November 1, 2026 · Pergent5

Research & open source

Sunday, November 1, 2026 · Pergent6

Model training

One recipe for AI training

The paper's method, called unified training, runs every stage under one likelihood loss.

A paper from CMU and others says the usual training stages don't need to stay separate. Its method treats good RL goals as plain prediction on model-made answers, so all data shares one loss and one trainer, and the design question becomes which data to mix and in what ratio.

Worth a read

Sunday, November 1, 2026 · Pergent7

Interface design

Two UIs every AI tool needs

The post drew 1,590 likes and 110 replies.

shadcn says every AI tool will end up building a chat-style screen, the way Slack has, and a task-board screen, the way Linear has. He says the chat view is already here and the task view is next.

AI in court

Court sanctions prompt injection

The filing used prompt injection, text hidden to steer an AI reader.

A Brazilian court penalized lawyers for prompt injection in a court filing, a trick that hides commands meant for an AI reader. The HN thread has 33 comments.