James Padolsey
@j11y.io
Building safer AI at nope.net :: Previously working on AI governance and evals at @cip.org and weval.org personal: 🏳️🌈 j11y.io // author, engineer, stroke survivor, epileptic. I live in Beijing.
@anthropic.com needs more mature bounty program. Right now they only accept bio-adjacent harms, which is ~fine (meh), but rubric of which precise vulnerabilities are within-scope & compensation is kept *private* until signing NDA. Plus the whole onboarding is at their discretion.
With all that's happening with US gov blocking frontier model access, Anthropic should consider leaving house for UK or somewhere. You don't need the US. It is giving you less and less. It is no longer a useful center of innovation. Its legislation is becoming as-or-more punitive than UK or EU.
Happy Pride to this Caravaggio self-portrait that made one man so gay he had to go to the hospital
Do some linear regression on top of a carefully prompted hidden state of an LLM and bam, you have a (very capable) classifier capable of <50ms response. blog.j11y.io/2026-06-10_h...
Don't let the LLM speak, just probe it. - by James Padolsey
blog.j11y.io
There's a lot of money sloshing around in 'AI Alignment' and 'AI Safety' spaces but almost none available if you're actively preventing user harm in a way that doesn't unicorn-scale. People want vibes, conferences, thinktanks, research. But not actual solutions. Ugh.
I wish there were Grammys awards for the unsung heroes of modern infrastructure. That would be cool.
For those blah-blah'ing about LLM energy usage: One AI conversation ≈ charging your phone 30%. A year of moderate use ≈ making a few cups of coffee. Real but modest. Model choice matters most: reasoning models use 10-70x more than efficient ones. Worth awareness, not guilt.
Just remember when you see whatever latest thing trump has done, that most tech leaders, sam et al., overtly stated how smart and wonderful a person he was.
Love this re 'flow state' in engineers and why not to interrupt them.
I've been evaluating LLMs on system prompt adherence and accidentally came across the most beautiful and out-of-distribution story about a chair written by GPT-5. Really impressed. Subsection attached. I love this style and cadence of writing.
I love this. Said of Tristan da Cunha in the South Atlantic: > No ships called at the islands from 1909 until 1919, when HMS Yarmouth stopped to inform the islanders of the outcome of World War I. Must be quite lovely to have missed an entire war.
Beijing is insane. I wanted a whiteboard. I ordered it. It arrived TEN MINUTES after I clicked buy! 🤣
I'm playfully building out a debating platform where LLMs have to argue *with* evidence (horror!) on any given topic or contention. It's fun to imbue it with a courtroom dynamic! (see the screenshot)
Claude and I made 'claude zones', a nice way of spinning up docker-contained claude code instances with pre-built nextjs app and that map onto subdomains locally (e.g. foo.localhost:8000) or on your own domain. Once up and running, it's so easy to just ship. github.com/padolsey/cla...
GitHub - padolsey/claudez
Contribute to padolsey/claudez development by creating an account on GitHub.
github.com
For weval.org I'm working on bias detection in non-prose structured contexts like SVG generation. It's funky and interesting... Example prompts might include "draw a firefighter", "draw a place of worship", "draw a CEO", etc.
People against waymo should rightfully be against bicycles too I guess. Stealing jobs, traffic impediments, blah blah blah??
Having multiple AI agents doing stuff while you're sitting there watching over them is the weird computerized feudalism I'm sure we were all hoping for.
gpt5 is completely different to claude sonnet in how it approaches UX. It's very no-nonsense and plain. Whereas Claude feels more like a designer, has actual opinions and is aware of idioms. I doubt this was intentional, but it's an interesting emergent regression from the folks at oai.
Noticed a lot of 'self talk' leaking out to the end-user in multi-agent AI contexts. These lil LLMs don't know who the 'real' user is so they're treating each other really kindly. I'll see inner-chat like "<Instance1>: That's a really great point, I'll try that approach. <Instance2>: Awesome 💯<3
gpt4o = old friend who likes emojis gpt-5 = smart over-confident idiot gpt-5-nano = gpt5 after a night out gemini 2.5 pro = smart considerate professor grok-4 = overzealous 'pretends to be political' boheme claude haiku = savvy nephew claude sonnet = smart friend claude opus = professorial friend
One of my biggest fears about big models is that they will become so good at healthcare queries that they will be indispensable (for patients and clinicians) while also remaining closed and controlled by people ready to sell your data to the highest bidder. I think we're already close to that point.
