site banner
PaperclipPerfector  10d ago (text post)   2297 thread views

Friday Fun Thread for October 31, 2025

Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.

2
Top Bottom Old Controversial Comments
Jump in the discussion.

No email address required.

PewDiePie has a video about maxing out his own AI.

I want to run my own AI, so I can load a bunch of memory into it and try to replicate the Smart House. That might happen, but not this year, or next.

Instead, I've been quizzing my favored AI on music. When I was an undergraduate, I had to transcribe a jazz solo, and I did Miles Davis' solo on So What from Kind of Blue. Naturally, the question I've been using is:

transcribe miles davis' solo on so what from kind of blue

It's awful. Grok does best when provided with spectrographs, which I can generate from audio files via audacity, but Gemini is completely helpless. ChatGPT seems better, but still not close to Grok. Grok did the best at trying to give me readable transcripts, which were either ABC notation or LilyPond which didn't work.

Anyone care to transcribe the spectrograph*? I will subscribe to the Motte patreon for 1 year if anyone can tell me what the song is from this image via any means necessary, and I'll do it at more than $5 per month if someone gets me a genuine transcription. It's not Miles Davis**.

*PS The native themotte.org image upload failed me, but so did imgur, so who knows.

**PPS Here's the audio

I can't fathom why people would spend $20K+ getting a box of modded 4090s, all to access an open-source $0.18/$0.54 model 'for free' (still paying for all the electricity, so probably more expensive if anything). Just get an API key, save tens of thousands of dollars and a great deal of time. Microsoft, Google and Intel and AMD are probably already spying on you. It's not like running locally will greatly enhance your privacy.

And what is there even to be private about regarding AI? Yes, the people running APIs are probably sniggering at the logs of the goon sessions. I've sniggered at some logs myself, though mostly I just find the low standards of taste appalling. Maybe if you're Pewdiepie it's worth it, since journalists would find value in muckraking and log-sniggering.

But why would any normal person care? It's highly unlikely that they can even trace the logs back to a human identity, even less likely they'd care to do anything. Let the gamers buy their 4090s. Let the API providers on Openrouter get some revenue. Use an API key.

I can't justify 20k USD, but I've justified 1.5K to my boss for work reasons and around the same to myself for personal reasons.

  • Control. You don't have to worry about parts of your prompt getting funneled off to a cheaper LLM from 2021, you can be pretty confident whether your 'super-long context' implementation is using normal context instead of pretending and slamming the RAG button instead, modify temperature, so on. At the more extreme end, you can make decisions about how agentic tooling is set up, what APIs it uses, and if you really want to get into the weeds you can skip the whole mess and just rawdog python to change behaviors. Most people won't want to put the energy into this on the front side and it wouldn't benefit them if they did, but if you're in this use case few mainstream APIs support it and fewer still are likely to maintain that support.
  • Censorship resistance. Nearly every mainstream LLM gets repeatedly censored or at least has its API censored, often making them much dumber in the process; nearly every mainstream diffuser API starts out censored and the rest get there in days. On the gooning side, Grok Imagine had about a week where spicy mode could get you outright sex scenes, and now you can maybe get someone in a jockstrap or with an exposed nipple, and I don't expect even that to last. Grok's text form hasn't booted me (yet), but most other vendors will outright refuse. But it's not limited to sex and sexuality; I've had prompts rejected that were related to electrical engineering, plumbing, aviation communications, motor design, machining, firearms history, case law, the list goes on and on. Most of these even had valid reasons, and there's certainly things that make sense to censor if you're answering questions for every rando on the planet, but it's a capability we'd be leaving behind without it. In most cases I could eventually find a provider that was willing to answer the question... for now.
  • Privacy. From a business perspective, we have code that it is a firing offense to bring outside of company property for any reason; quite a lot of businesses work like that for a wide variety of private information, and some are more restrictive and don't allow any network communications outside from secured areas period. To be fair, if you're Meta/OpenAI/Google, yes, you could absolutely rebuild (or steal) the code my company runs on far cheaper than running this whole AI ecosystem. To be less fair, Amazon has the manpower to engineer and market every product on the planet, and still didn't keep its firewall in order. Google might be able to figure out a lot of stuff from various search requests and e-mails going through their relay servers, but it's a lot easier for them to have our verbatim work if we directly put it into an HTTP POST; a localLLM that's got no (or for agentic llms highly restricted) outbound network traffic probably isn't doing that. And, yes, from a gooning perspective, I expect very few people are paying attention to what are, ultimately, not that unusual or ugly as kinks go on a matter that these companies absolutely don't want to be involved in policing. I'm also not green enough to imagine these aren't logged, or that no situation could arise where these logs could be useful in the future, and leaving aside the various conspiracy theories about the political acumen of search engines, there are advocates today of large-scale prohibitions on the most boring variants of this content, and we've had a recent SCOTUS case over whether people can be liable for publishing porn without age verification regardless of whether a minor sees it, and this law is not alone. There's even reason these advocates could (not always unreasonably!) see AIgen content as particularly unconscionable.
  • Proofing against an AI Winter and/or an AI Bullrun. There's a lot of potential spaces for OpenRouter to be much less viable or more expensive in a very sudden way, and most of those explanations end up with either drastically reduced availability of inference power, AI tooling, or both. This could happen for regulatory reasons, because of supply chain problems in Taiwan, downstream of financial bubbles. Hell, it could happen even if Sam Altman makes GPU Jesus in his basement and suddenly all of the datacenter GPUs go to the important stuff instead of answering problems you could 'just' look up (if you had five hours and a subject expert to point you to the right keywords). It's probably not even that likely! (especially GPU Jesus). But if a good LLM is saving your business 200k a year, are you 90% sure it's not going to happen? But the actual numbers are a little worse than that, because such a disruption is likely to come as a surprise, and an honest decision matrix has to put the electricity and development costs for running local llms along with the materials cost on one side and a bunch of your users suddenly going cold turkey on the other.

That makes a lot of sense and is well thought out... but are you really getting that much value for your $1.5K investment? That's sufficient to run what, a 32B model locally? I don't know if I'd trust a 32B model with anything serious. TBH I haven't used a little one for ages, I only play with the big ones, so maybe I'm out of date on this.

Not OP, but depending on what you are doing and your config you could almost certainly run glm-4.5-air or openai/gpt-oss-120b, which are roughly 100B class models.

A name brand box with a AMD Ryzen AI MAX+ 395 and 128GB of unified LPDDR5 RAM would probably be just over that now, but you can probably find a no-name box on sale for around $1.5K from time to time. Performance would obviously be worse than duel RTX 4000 Adas or something, but a lot cheaper.

The use case I'm imagining is like a background task doing a code review or auditing a highly sensitive code base to check for potential vulnerabilities, intentional or accidental. I could also imagine using something like that to slowly scrub through heath, financial, or other sensitive files. Either for auditing purposes or converting to structured data.

It would probably be a bit slow, but for anyone who has to work in an air-gaped environment it seems like it would actually be supper useful. It saves you having to send a query to the public internet the majority of times you have to look things up. Just replacing google searching, or (bleh) having to look something up in a paper book. It doesn't take that many uses from an engineer making 200k a year saving a few minutes to make it ROI positive for a business. Even just the time it takes to transcribe something you looked up from the internet facing machine to your offline machine. I suppose it depends on how many people are in the working group whether it would be more efficient to have some beefier centrally hosted machine on the intranet.

Even if it doesn't have to be air-gaped, I imagine if you have like 100+ employees dropping like $20k might still be cheaper over like 1-year than paying for an API provider. Especially if there are a bunch of compliance problems with sending things off site.

What is this place?

This website is a place for people who want to move past shady thinking and test their ideas in a court of people who don't all share the same biases. Our goal is to optimize for light, not heat; this is a group effort, and all commentators are asked to do their part.

The weekly Culture War threads host the most controversial topics and are the most visible aspect of The Motte. However, many other topics are appropriate here. We encourage people to post anything related to science, politics, or philosophy; if in doubt, post!

Check out The Vault for an archive of old quality posts. You are encouraged to crosspost these elsewhere.


Why are you called The Motte?

A motte is a stone keep on a raised earthwork common in early medieval fortifications. More pertinently, it's an element in a rhetorical move called a "Motte-and-Bailey", originally identified by philosopher Nicholas Shackel. It describes the tendency in discourse for people to move from a controversial but high value claim to a defensible but less exciting one upon any resistance to the former. He likens this to the medieval fortification, where a desirable land (the bailey) is abandoned when in danger for the more easily defended motte. In Shackel's words, "The Motte represents the defensible but undesired propositions to which one retreats when hard pressed."

On The Motte, always attempt to remain inside your defensible territory, even if you are not being pressed.


New post guidelines

If you're posting something that isn't related to the culture war, we encourage you to post a thread for it. A submission statement is highly appreciated, but isn't necessary for text posts or links to largely-text posts such as blogs or news articles; if we're unsure of the value of your post, we might remove it until you add a submission statement. A submission statement is required for non-text sources (videos, podcasts, images).

Culture war posts go in the culture war thread; all links must either include a submission statement or significant commentary. Bare links without those will be removed.

If in doubt, please post it!


Rules


Recommended Posts And Communities

Recommended Realtime Chats