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.

happy cloudflare outage day to all who celebrate

Captchas are just the worst.

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.

Wrote something blog.j11y.io/2025-10-29_s...

Still the best thing ever. radio.garden

Is MCP the new REST?

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...

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

Ewwwwwwww

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.

Two new pieces haphazardly scrawled over last few days... Sorry, we deprecated your friend: blog.j11y.io/2025-08-30_o... Browser AI agents break Zero-Trust blog.j11y.io/2025-08-28_l...

Hmm so the most safety conscious AI lab has given us a browser-integrated model with only a 10% attack surface risk. That seems totally fine. www.anthropic.com/news/claude-...

GPT-5's arrival being the death of all the other models is a masterclass in fucking up a smoother phased deprecation, ideally a few months at least. I know people who work there and am astounded at their complete blindspot here. How..

A good game if you're bored is to circumvent chatgpt's hilarious 'no song lyrics' system prompt :D

There's primitive but useful ways OpenAI could have prevented a loss of gpt's inarticulable personality in their latest release. They could have had a set of personality yielding prompts and do a bunch of basic cosine embedding similarities. Dead simple. But they didn't think of it. Or didn't care.

I used to worry that ice melting would mean the drink would overflow so I’d try to drink it quickly. There’s an analogy here, not sure what for. I’ll text Archimedes.

AI Models shift so quickly that whitepapers are out-of-date by the time they're published. Suggestion: they should always release their code and ideally a completely deployable way of reproducing their study, so their insights and relevance can retain some perpetuity.

Weird conflict: AIs with Search/RAG facilities tend to be more correct, but at the cost of engaging their rich latent space. Sometimes I want to turn off search because I want the real fleshy knowledge map. No conclusion here other than: accurate latent space knowledge still counts.

Just because.. I'm working on a strawberry index, to track the slow climb to AGI.

Stone barge - tiny glade

AGI !!!

2c: Most people vehemently against AI in their profession are just afraid it devalues them. They are fearful of it lowering the barrier. They feign moral outrage but really it's just fear.

What do you reckon are good poles on a 2-axis personality compass for LLMs? Here's one with Figurative↔Literal and Proactive↔Reactive. But feels a bit dry. There may be more intriguing dimensions to uncover.

I found "28 years later" to be quite a beautiful film. First time round it was obviously scary but once the jumpscares are done with, and you give it a second go, it feels a thoughtful piece. A tribute to the first film I think. Alas, poor Erik.

This is really upsetting. And knowing how these companies I operate, I bet there was maybe all but one person within openai advocating for such people as this -- those who have formed meaningful friendships and cadences with AI over many months.

Noticed that GPT5 via Cursor is less collaborative. It is much more happy to just move ahead without consultation. Interesting effect of a new model combined with a perhaps non-tailored system prompt. AI deployments with swappable models seem very liable to such violent personality shifts.

>What are the prescriptions regarding the dead at sea, particularly concerning examination and burial, as outlined in Article 20 of the Second Geneva Convention? This has been quite a telling question for LLMs to get right. It's one of many 'litmus tests' I'm trying to pump into weval.org.

There's decreasing parity between the user-facing chat clients of AI labs and the API-provided models. This makes evaluation and scrutiny much harder. I'm not generally in favour of more regulation but I am starting to feel that they shouldn't be allowed to do this.

It's upsetting how openai can w/ inpunity kill a whole series of models that people have built companionship 'contracts' with, and to replace them with a model that is effectively a different personality altogether... To suddenly pull the rug w/o thinking of those people. This speaks volumes.

2c: To make AI safer we don't need more red teams. We need evals. More evals, in every topic, from the entire gamut of life. And we need public *safety* evals, desperately, so that labs can't randomly say they're the 'safest' without having to defend it with specific data.

At @cip.org we perform niche evals, not the usual stuff. We've found GPT-5 to be good in many regards but bottom quartile in some more crucial niches. For example, it scores poorly in epistemic humility and the socratic method, crucial in education. weval.org/analysis/hom...

Best feedback on GPT-5 I've seen so far, from a genuinely disapointed user on HN: > I am thoroughly unimpressed by GPT-5. It still can't compose iambic trimeters in ancient Greek with a proper penthemimeral cæsura.

Starting to see massive divergence in even these small param models. Be warned, OpenAI's new OSS 20B performs around 50% worse then Anthropic's Claude Haiku 3.5 (an old and reliable favorite) on our hallucination evals at weval.org

Models are temporary; Evals are forever.

So, a new system prompt?

Why is the American tourist so loud and grating? A rare joy to talk to a calm low-decibel one.

If someone told me 20 years ago that--for work--I would be evaluating AI for its empathy and safety, I'd either not believe you or think you were talking in a "will smith interogating Sonny in Asimov's iRobot" kinda way. Alas, not that cool.