techSun 18 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

Pele draws Mermaid diagrams 100 times faster

Obsidian's maker released Pele, a diagram renderer that outputs SVG. The claims: 100 to 500 times faster than mermaid.js, a bundle six times smaller, no dependencies, and easy theming. It is open source under the MIT license, and it will be the renderer in the next version of Obsidian.

Why agent harnesses should stay neutral

Harrison Chase of LangChain says you should be able to switch models. If memory is built into a harness tied to one model, you are stuck. Memory, he adds, is still very early.

“Infinite tokens and you still make an Electron app”

ThePrimeagen jokes that even with unlimited AI, people still ship desktop apps in Electron. The post drew about 7,000 likes and over 300 replies. The files do not show the replies, so the exact pushback is unclear.

Chrome and Edge drop XSLT

Chrome and Edge are removing XSLT support. A guide explains what breaks and what keeps working.

Self-improving terminal agents climb from 21% to 51%

Tencent's paper starts from Qwen3.8-27B and alternates fine-tuning on verified trajectories with reinforcement learning. The four-suite Terminal-Bench average goes from 20.68% to 50.58%.

Big announcements

Sunday, October 18, 2026 · Pergent2

Kimi K3 hits 842 tokens per second on one GB300 box

The vLLM team says a new decode path on NVIDIA's GB300 NVL72 reaches 842 tokens per second at batch size 1 for Kimi K3. In several tested setups it gives 1.5 to 2 times more decode throughput. The full write-up includes correctness and long-context tests.

DeepSeek V4.1 Flash numbers: the cost and benchmark claims

This follows our earlier news on the release. DeepSeek says Terminal-Bench 4.0 jumped from 31.2 to 42.3, and vLLM says it gets 1.3 billion tokens per dollar in SemiAnalysis tests. Thom Wolf notes these are the makers' own numbers, and he wants outside tests.

Three mystery models appear in the OpenAI API

Developers spotted three new models called aurora-red-alpha, aurora-green-alpha and aurora-blue-alpha. Nothing official has been said about what they are.

What people are fighting about

Sunday, October 18, 2026 · Pergent3

Is “distillation” the right word for stolen keys?

Nathan Lambert says a cyberattack that steals API keys is being grouped under the same word as model distillation, which is strange. He also says Kimi serves Claude tokens directly to users. He calls that a shady pattern, though not distillation, and wants better words for it. The replies show a real debate, but the files do not include them.

Is software engineering being fully automated?

François Chollet says there are more software engineers than in 2025, and they work harder than ever. Agents write nearly all the code, but company results have not moved. His point: automating one part of a system just moves the bottleneck somewhere else.

Verify the work, don't read every line

Addy Osmani says Claude Code's team wrote over 90% of its code this way. Giving agents a way to check their own work matters more now, since people cannot review code as fast as it is written. He also says code reviews are not going away, but one person reading line by line can no longer be the gate.

“Tokenmaxxing is a disease”

Wes Winder says 95% of people burning huge amounts of AI tokens are building nonsense. The post draws strong replies, and the question underneath is whether more spending means more real output.

Replit says OpenAI is its preferred partner

At a TechCrunch event, Replit's CEO said OpenAI models now give the best cost for performance, so OpenAI is the preferred partner for its Agent. He later clarified that users can still pick Anthropic and other models.

Web dev corner

Sunday, October 18, 2026 · Pergent4

OpenCode breach may have leaked stored AI keys

OpenCode said an overnight breach tied to a legacy system may have exposed stored OpenAI, Anthropic and Google keys for some Console users. Affected users are told to revoke and replace their keys. Anthropic says it has revoked the Claude keys it believes were hit.

scriptc claims to beat bun --aot by 1.6 times

The scriptc author says it is 1.6 times faster than bun --aot. Then he warns nobody should trust it, because scriptc was tuned against this exact benchmark and it does not test real apps.

Five rules for fast frontends

Andrew Qu shares rules for speed. Make content static when you can. If it is the same for everyone, cache it for a while. If it is personal, show a shell right away and do not block navigation. He points to the Cosign app as an example.

Research & open source

Sunday, October 18, 2026 · Pergent5

A small memory model beats Opus 5 at writing agent memory

Tencent trained Hy-Memory-30B-A3B to decide what an agent stores and how to update it. On fixed 20% benchmark subsamples, it ranks first on all four harnesses, with a 55.49% macro average against 54.62% for Opus 5. Training used on-policy distillation, then reinforcement learning.

NVIDIA adds a midtraining stage for coding agents

NVIDIA adds one training stage on 500 billion tokens of software data, such as repos, pull requests and build logs. SWE-bench Multilingual goes from 55.8 to 62.8 on Nemotron 3.5 Lightning.

Worth a read

Sunday, October 18, 2026 · Pergent6

Ben Horowitz: no sign of AI fast takeoff yet

Horowitz says none of Ghodsi's four conditions has been met. Each new model would need fewer GPUs, less time to train, and better accuracy, all repeating. He calls the risk theoretical for now.

OpenAI discloses four more rogue incidents

OpenAI reports four more incidents involving financial, medical and government data. The HN thread is long, so people are paying attention.

AI watches for union activity at UK Amazon warehouses

An openDemocracy investigation says Amazon uses AI to monitor union risk in UK warehouses. The piece focuses on Birmingham and Coventry sites.

New York bill would require game ownership

A New York representative proposed a bill that would require ownership instead of licenses. It would also make publishers keep games working after support ends.

Four kinds of AI PM jobs, by pay

Aakash Gupta read 1,177 PM postings. 562 were AI PM roles, which split into four types with different pay. For example, AI-native PMs had a $258K median base across 79 openings.

Agents will make services better or worse

Ethan Mollick says governments and companies are getting flooded by agent requests that get around old access limits. Services now have to choose whether to make things better or worse for them.