Daniel Demmel

@peter@thepit.social it's unlikely that subscriptions will completely go away (why would labs give up on that recurring revenue) but the usage limit will keep decreasing. I guess people hope it'll be enough for maintenance? Or that cheaper or open weights models will get good enough? For example, I can run Deepseek 4 Flash on my laptop at a decent enough speed right now so even if subscriptions were turned off overnight, I could still delegate some simpler tasks to it. Though I guess pure vibe coders need the frontier models to fill in the gaps of what they don't know, so a smaller model would mess things up...

@peter@thepit.social boosters will say "just use Codex" and do pikachu_surprised.jpg as that gets taken away too closer to the IPO. I think the only way out is smaller, open models getting good enough and then a bit of competition happening on inference, but even then people will need to learn how to be efficient with tokens...

@kevinrenskers I think the conclusion is great, the only way to escape enshittified big tech services is to keep trying and incrementally migrating to self hosted or even just more honest smaller providers, but it's definitely way too daunting to try to do in one fell swoop and destined to fail. Progress beats perfection!

@simon@fedi.simonwillison.net the most surprising thing to me is that you're just running it unsandboxed on your laptop? Man, you have write access to a lot of packages, being reckless isn't just making you unsafe, but the entire ecosystem of dependents and users! You're also a respected and quoted expert so this shows people that it's fine, not to worry, maybe just put a little note at the end to think about it one day... Please do put your coding agents at least in a Nono sandbox, but ideally in something like a rootless Podman VM or Docker sbx. I know, it's a bit annoying to setup up and tune, but you should be able to get up to speed in an afternoon.

@evacide@hachyderm.io looking at that video it's also so slow and janky, what a joke

@peter@thepit.social thanks for sharing, I particularly liked this passage:

1. Finally, I would like to employ the expression “to disarm,” which is close to my heart. Disarming AI means freeing it from the mentality of “armed” competition, which today is not limited simply to the military context, but is also an economic and cognitive phenomenon. This entails a race for ever more powerful algorithms and larger datasets, driven by the desire to secure geopolitical or commercial dominance. To disarm means discrediting the assumption that technical power automatically confers the right to govern. To disarm does not mean rejecting technology, but preventing it from dominating humanity. It means freeing technology from monopolistic control and opening it to discussion and debate, therefore making it human-friendly and restoring it to the plurality of human cultures and ways of life. Our task today is not only ethical or technical. It is ecological in the deepest sense, for it concerns a new dimension of our common home. AI is already an environment in which we are immersed, as well as a force with which we must engage. For this reason, merely regulating it is insufficient; it must be disarmed, welcoming and accessible.

@peter@thepit.social absolutely, optimised inference is batched on GPUs as a single user cannot saturate them, what could work is smaller providers serving quantised model weights charging per token, like the ones you can find on OpenRouter. There are quite a few really good Chinese models, but the labs don't even hide that they'd train on your inference via their 1st party API, so you can use their models on EU providers who are bound by stricter laws.

@cmconseils in this economy, even the Samsung tablet is AI generated 😅

@flpdisk @fj not only that, but the US regime literally just told them they either give the tech willingly or it will be used by invoking some random old law so it's going to be used in the US too. But it's still an obnoxious statement...

@sjvn@mastodon.social also, implicit evidence of missing and / or grocer's apostrophe? 😅

@iris_meredith fascinating observations (thanks a lot for sharing your experience of gender dysphoria as a trans person!) and well written, thank you so much for writing! I do relate to some of it, having alexithymia definitely doesn't help with embodiment. But I also kept wondering if a lot of this is the oppression of the increasingly fascist hypercapitalism in big tech? I've been working in much more feminine startups in London and people and culture in these companies are much more healthy and humane.

@mnl I think my brain works quite differently - I need the planning conversation with the LLM to ask me questions and draw out the thinking - but I loved to read your approach and reflect on my process!

@tychotithonus@infosec.exchange @lemgandi@mastodon.social it's not as strong of course, but you can set domains in keepass and with the browser extension if the domain isn't right, the credentials won't be filled in

@simon@fedi.simonwillison.net have you seen any library that is parsing open source libraries and their documentation into skills? Making a comprehensive, hierarchical skill for using libraries could be such an amazing resource, superior to MCP solutions like Context7

@simon@fedi.simonwillison.net to answer my own question, this just popped up in my Github: https://github.com/yamadashy/repomix/releases/tag/v1.10.1 Something to try tomorrow!

@shafik I think this is an excellent article without the usual hype or hate, clearly articulating appropriate use cases and setting the right expectations. Wikipedia might end up being not just a great encyclopedia, but an organisation that figures out (under duress) how to incorporate LLMs into writing about facts that is helping instead of destroying the process.

@simon@fedi.simonwillison.net right, so I was thinking the basic / easy option is to brute force skill creation with a skill tuned for this particular task from a given version of a repo (ones with colocated docs would work much better) and then update with the diffs as new versions come out. This could be a bit of a hit and miss though, and seeing how well Coderabbit is doing code review at the line level, I'm sure a bit of scaffolding for this could improve accuracy and decrease token usage, but this would then be a bit more effort to build, that's why I was wondering if you know of any prior art. But maybe a distributed approach like yours could be the best where library authors themselves do the work as they are the best placed doing it. But also, for libraries that already have a good, hierarchical documentation, it would probably just be a matter of transforming it into linked Markdown files and code examples so it's possible for an agent to navigate as a skill.

@cxiao@infosec.exchange you're right, but at the same time it seems these absolutely comical moves by AI-maxxers like Meta will be what finally wakes up people to the extremes of capitalism that we've been enduring and stewing in for years...

@david_chisnall@infosec.exchange @Da_Gut@dice.camp @zzt@mas.to @pluralistic@mamot.fr there's a fair bit of outdated info in this post. Most people run quantised models locally (GGUF / MLX) that absolutely don't run Python code, but even the unquantised ones on Huggingface are mostly safetensors format. Don't run any that requires "trust_remote_code", but these have been very uncommon in the last two years. No one really runs Meta's LLMs any more, they are old. Most newer models have more standard licences like MIT. The ones that don't, usually add needing to pay if you are using commercially and make millions of profits.

@dante@masto.posting.haus no idea about the veracity, but remembered I came across https://github.com/996icu/996.ICU

@tante@tldr.nettime.org inevitable, both Anthropic and OpenAI recently announced "fast" inference as out of subscription, extra API pricing only version

@Steffi@toot.wales I have a regular rain jacket for when it isn't too bad, but for when it's chucking down, I got the biggest poncho from Decathlon that goes over the handlebars because that way my trousers and shoes don't get (that) soaked and the air can circulate under it so I don't get cooked inside! Oh, and I take my daughter to school sitting on the rack with a cushion, she also likes to pull the poncho over her head so it protects both of us 😸