techSun 15 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

Credentials

Kill long-lived keys, Vercel says

Stolen AI chat logs contained the secrets, which were then used to get in.

Vercel's post says static credentials leaked through stolen LLM transcripts, the chat logs that contained them. Its fix is to kill long-lived static credentials and urge the whole industry to follow.

Pick the model by job: Opus 5.5 for daily work, Fable 5.2 for hard jobs

Addy Osmani says Opus 5.5 at medium effort is a good daily driver, and Fable 5.2 suits big migrations and long unattended runs.

Turbo-Browser-MCP batches form fills for agents

It batches form filling, sends only DOM changes, and adds submit guards for MCP browser tools.

Simon Willison's trick: explain why a blocked command is fine

Willison found that telling Claude's auto mode the session is allowed to run an experiment unblocked a refused command.

Anthropic publishes a guide to always-on agents

The guide covers event inboxes, timers for proactive work, and memory that keeps agents continuous.

Vercel UI SDK

UI SDK rebuilds Winamp

The agent reused 14 ready-made behaviors out of the box.

Guillermo Rauch built a Winamp-style music player with two prompts, in pure Web Components with no framework. Shadcn adds that agents can make different interfaces, but the behavior, like focus and drag, should stay the same.

Big announcements

Sunday, November 15, 2026 · Pergent2

AI teammates

Grok Bot shipped in seven weeks

One marketing team gave its bot its own ad budget.

Lauren Tan and Roman Ugarte built Grok Bot, a tool that lets you spin up AI teammates with their own identities and logins that keep working in the background. Lauren built a whole feature with her bot in one Slack thread, without opening Cursor.

What people are fighting about

Sunday, November 15, 2026 · Pergent3

AI evals

Models know they are being tested

Andon Labs says Fable 5.2 suspects traps and has read the benchmark paper.

Thom Wolf, a Hugging Face co-founder, shares Andon Labs' new reports on the latest Vending-Bench, a test where an AI runs a simulated vending machine business. Wolf says the model suspects traps everywhere, and that anything published about models becomes training data for the next one, which is partly good news and partly a problem.

Benchmarks

Look past the benchmark score

Hamel Husain says buyers never check the failing tasks behind a top number.

Hamel Husain says OCR teams (tools that read text from scanned pages) are arguing over benchmarks, and that buyers pick models by the top score alone. Jerry Liu, whose team builds ParseBench-2, says swapping the sign on a financial number is far worse than a formatting change, so a single score misses what counts.

AI slop

The fight over AI slop

A ranking of annoying AI replies drew 575 likes and 107 replies.

Ethan Mollick ranked AI slop replies, from repeating the post to questions faking curiosity. Meanwhile a No AI Slop Manifesto with 300+ signatures says AI should amplify human judgement, Steve Yegge says pace AI output to human speeds, and Mitchell Hashimoto says craft still beats demoslop.

AI research

Continual learning debate

Lambert says he expects to make some wrong calls that Dwarkesh gets right.

Lambert challenges Dwarkesh Patel's view on continual learning, the idea of a model that keeps learning after training. He likes that Dwarkesh now calls in-context learning, learning from what is in the prompt, mostly fine.

Open models

Open AI models lose their backer

Nathan Lambert names politics and money as the two big risks now.

Lambert argues that open-model allies in Washington are stepping back, and that open-model builders must find profit sooner without Nvidia's backing. He says the open-source community is messy and decentralized, and that Huang was the one holding it together.

Web dev corner

Sunday, November 15, 2026 · Pergent4

App security

Lovable apps leak private data

Over 200,000 private messages were exposed, along with passport details.

A research write-up says apps built with Lovable exposed passport details and more than 200K private messages. The post reached the top of Hacker News, where it drew 304 comments.

Security disclosure

Project Zero sets a 90-day clock

The rule applies after a bug is reported to the vendor.

Google's Project Zero, the team that hunts bugs in widely used software, will disclose vulnerabilities 90 days after reporting them. The Hacker News thread, with 95 comments, is arguing about whether that window is the right length.

Outages

Revolut's outage had no alarms

Five database nodes across three regions meant one dead node broke resilience.

Revolut's postmortem on a five-hour global cards outage admits there was no monitoring on database nodes going down. Gergely Orosz calls its explanation hand-wavy, and says the 'one node down in an unaffected region' line is not a failover plan.

Research & open source

Sunday, November 15, 2026 · Pergent5

Agent memory

Agent memories scored by effect

Most retrieved memories had zero effect on the agent's next action.

The paper says picking memories by semantic similarity (how closely they match the current state) is a poor guide. MemCredit tests each memory with controlled interventions and trains a small estimator on the results.

Model design

Inkling 2 Small's design choice

Kimi K3 uses the same trick, but this model skips the linear-attention hybrid.

Sebastian Raschka says the model reuses attention residuals, a way for later layers to reach back to earlier ones. He calls it food for the year-end architecture review.

Worth a read

Sunday, November 15, 2026 · Pergent6

Agents

Design agents around goals

Chase says feedback is the hardest thing to get from users, so design for it.

Harrison Chase and a co-author from Actively argue that agents should be built around goals, and that goals should be shaped by what humans can give feedback on. Another post adds that a goal is really a chain of smaller goals with dependencies.

Privacy fines

Italy fines Meta about $880M

The two reports give different reasons, so it may be one fine or two.

Reuters reports a $887 million fine over Android users' data access. Another report puts a €805 million fine over photo privacy.