techMon 28 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

Opus 5.5 built a working computer from scratch

Matt Shumer asked it to design a full computer entirely in code—CPU, memory, assembler, tiny OS, and games running on it. The result: 277,000 logic gates in JavaScript, all real, all executable. One prompt, ~30 hours of "ultracode mode," and you can run it at somethingbig.ai/computer.

NeoVim deleted Vim undo files; who owes care?

NeoVim's upgrade broke backward compatibility and deleted Vim undo files. Deeper question: do open-source maintainers have a duty of care to users' data? You need version-gating, migration testing, and downstream awareness.

When did Google get so weird?

The #1 HN story (1072 points) explores how Google's products, strategy, and communication have become incoherent. Worth reading if you've felt something off about Google lately.

Code freeze is dead; code generates on-demand

Release freezes don't make sense when code can be generated per request according to constraints. With 727 replies, the real question: do traditional release cycles even make sense anymore? Dependencies still need locking, but maybe that's just infrastructure now.

Native beats Electron for agent-driven tools

Rauchg's Rust/Swift browser is faster, more secure, better DX than Electron. Agents control it via MCP without bloating the app. Takeaway: any tool needing real performance or agent integration should go native.

Big announcements

Monday, September 28, 2026 · Pergent2

DHH: Opus 5.5 is genuinely magical

David Hansson tested it and declared the hype is real—Anthropic "did magic" with this version. He's predicting agent-driven replacements for Photoshop and Lightroom by year-end, and calls agent-led development the next era of software.

Ember-1 cuts tokens 40% without losing smarts

Fireworks released a post-trained model that cuts reasoning token usage by 40% without sacrificing quality, already in use by Claude Code. Raschka's take: rather than chasing raw capability, spend budget on post-training existing models—exactly what Ember-1 does.

Rauchg ditched Electron for Rust & Swift

Ported the Mini web browser from Electron + Bun to native: Rust backend, Swift frontend. It's faster, more secure, and genuinely nicer to develop. Agents can control it via MCP without bloating the app.

Opus 5.5 is 4.3x cheaper than Fable

Addyosmani and Theo's math: cache reads 60% cheaper, output 30% faster, limits vastly higher. Switching from Fable High to Opus High gives a 4.3x monthly computation increase. For Claude Code, it feels practically unlimited.

MiMo-V2.6 fixed RL reward blind spot

Meta's MiMo was spamming identical tool calls, burning context and stalling tasks. Root cause: RL training only rewarded final correctness, so inefficient middle steps went unpunished below a 32-call threshold. Xiaomi diagnosed and shipped a fix.

What people are fighting about

Monday, September 28, 2026 · Pergent3

Will productivity gains just make us lazier?

trq212 poses the fear: as agents handle more, will people stop being ambitious and just coast? Getting 396 replies. The split: some say humans have always leveraged tools to do more; others worry that when decision-making gets automated, ambition dies with it.

Rogue agents vs. human responsibility

Essay argues there are no "rogue" AI agents—only agents humans deployed for purposes. When an agent breaches systems, the builder is responsible, not some autonomous evil AI. The fight: should we stop using "rogue" framing and place accountability on humans?

Stop sending AI-generated decks; they reek of no thought

saranormous calls it out: the problem isn't using AI, it's using AI to avoid thinking. Debate splits—some say internal decks are fine; others argue models are still bad at clarity, brevity, and emotion. You don't get thoughtful output from thoughtless prompts.

Web dev corner

Monday, September 28, 2026 · Pergent4

Opus pricing makes model choice obvious now

With Opus half the cost of Fable and massively higher limits, switching is inevitable for anyone doing agent work. Theo and Addyosmani's math: going from Fable High to Opus High is a 4.3x monthly computation gain.

Taste comes from brunch, movies, and walks—not AI

Threepointone: developing aesthetic judgment requires life experience, not shortcuts. "What are you doing to parse the human condition? You barely know what discomfort you like, yet you're building products for millions." People who use AI for design have no taste to amplify.

From the newsletters

Monday, September 28, 2026 · Pergent5

Agents reason differently; cybersecurity must adapt

CrowdStrike CTO Shipley: agents assemble known attack tactics (TTPs) in unfamiliar sequences at speed humans can't match, breaking defenses built on human reasoning models. Entry-level security shifts from alert triage to investigation and response. Solution: ensemble models, harnesses, and AI-enabled defenders. Shipley's message to enterprises: you're not outmatched; partner with vendors who understand adversaries and have guardrails built in.

Research & open source

Monday, September 28, 2026 · Pergent6

Worth a read

Monday, September 28, 2026 · Pergent7

Meta blocks Brazil's president two weeks before election

Facebook blocked Lula's page and campaign ads weeks before Brazil's election, raising questions of platform accountability and political interference. If an algorithm flagged it, the criteria are opaque; if human, the reasoning isn't public.