Ju Lin

This tiny 7M parameter model supposedly beat Gemini Pro. We didn’t believe it. So we gave Clusy one prompt to reproduce the result, run the ablations and figure out what was really happening. The answer was wild.

@simonw Is it the same scraping wall for everyone, or does Claude Code just not have an allowlisted UA? If the latter, that's a five-minute fix behind an internal ticket queue, not an architecture problem.

It's annoying that you can't paste a link to a (shared) Claude transcript into a Claude Code session, because Anthropic's anti-scraping measure prevent its own tools from accessing the output of its other tools

@jxnlco Browsing agents rarely fail by clicking the wrong thing. They fail by not noticing the page changed underneath them and reporting success anyway esp when most pages are cached by fetching api. Curious how much of that gets caught before the human review step.

ChatGPT Work can now browse while it works. On web, the cloud browser can research public sites, compare options, and complete multi-step tasks. You can inspect screenshots, replay its steps, and approve actions. On desktop, the built-in browser supports tabs, sign-in, downloads, and local previews in a separate browser profile. Your existing tabs, cookies, and passwords stay separate. Website access and sensitive actions remain permissioned. Give it a real task with a few browser steps and see how far it gets!

@jxnlco Curious how this handles state: if the dashboard pulls live data, does every view re-run the model, or does it get baked in at publish time? That's the gap between "report" and an app that quietly lies about being current.

ChatGPT Sites is now in public beta. You can turn a prompt, file, or rough idea into a dashboard, project tracker, report, prototype, or lightweight app: build and edit directly in ChatGPT Work or Codex test in a private preview publish and share with a URL Rolling out across paid plans. Enterprise admins can control public publishing.

@fchollet The gap that matters now isn't low-skill vs high-skill, it's fast-verifier vs slow-verifier. Reading a diff and catching the wrong assumption fast is worth more than typing speed ever was.

The weak AI code gen we had until late last year was most useful to low-skill programmers -- it was raising the floor. It was essentially useless to high-skill programmers -- you could move faster and ship better code without. This has been completely flipped: the strong AI code gen we have now is *most* useful to high-skill programmers, while low-skill programmers are vastly underutilizing it or sometimes drowning in it. It went from a crutch to a power tool.

Moral: when an agent misbehaves, audit what it can actually SEE before you blame the model. Full postmortem: clusy.io/blog/the-agent…

The fix was one config line. The part that still bothers me: the agent had no way to know. A human in the same seat wouldn't have known either.

We spent a day accusing our AI agent of hallucinating labels. It wasn't hallucinating. It was blind — and the platform was hiding that from both of us.

@fchollet Adapting instantly is the easy half. The hard half is knowing *why* the prior model was wrong — most systems update the output and never surface the diagnosis. Speed without that is just noise with better PR.

The process of trying, failing, updating a mental model, and trying again is the core of intelligence. We should celebrate models that fail gracefully and adapt instantly.

@levelsio Genuine question: how many "server is down" resets are actually "the thing I use to reach the server is down"? Feels like it deserves its own line in every postmortem template.

Something that will eventually happen when you use Tailscale is You think your site is down, you try reach it over SSH, it doesn't work, you try open the site in your browser, it doesn't open, you're about to reset the server and think the worst! Then you think: let's try the Rescue console on Hetzner, you SSH in, and ask Claude Code what's going on And it tells you your Tailscale key for the server expired, because they automatically expire after 180 days (for safety) @DanielLockyer told me to always [ Disable key expiry ] for every server, but I forgot it for Interior AI, you kinda get locked out otherwise so important to disable this! You can do so in the @Tailscale console under [...] All good now!!

@levelsio The scary part: SSH and browser both tunnel through the same dependency, so when it fails, every signal you check agrees with the wrong conclusion. Redundant-looking checks that share one failure domain aren't redundant.

@eldar_hsnv @realpvarma You know ball

@realpvarma chinese (assuming you know english)

Suggest only 1 language to learn in the AI era

@yiliuai 在founders inc里面好多都是highschooler..

今天见了个在旧金山某加速器创业的朋友，他说他是他们加速器里年纪 最大的…… 就是他们加速器里一大堆大一大二辍学来的，然后一大堆二十三四岁的…… 咱们这种年龄的确实是老中老了……

@jamyanges Safeway roasted chicken thighs: ♾️

hackathons in 2 years: 14 startups: 3 newspaper features: 1 it compounds

@cchevyyc Mega

i haven't slept in 2 years. so I'm giving myself 31 days to build something that fixes it if I can pull it off, I want to help 1,000 other ambitious people sleep better too. would you want one?

@nina_phatak @flowstatebci This is super cool!

in sf at building wearable neurotech who should i meet? @flowstatebci 🥐

@Base44 Meh

New look just dropped and yeah, it took us a minute, but it was worth it.

@jsonphile Do what you got to do mate

@urjr1 Why is onlyfans the only other option bro 😭😭😭

the file explorer of a narcissist