hjhornbeck

Are people still debating this? Here's my contribution: even if you believe substrate independence, you can still argue LLMs cannot be conscious. How?

1. Seizures change the discrete firing timings of neurons in the short term.
2. Seizures can effect consciousness.
3. Therefore, short-term discrete firing timings are important for consciousness.
4. LLMs cannot model short-term discrete firing timings.
5. Ergo, LLMs are missing a key ingredient of consciousness.

That's just a summary, the full argument is at the end of this blog post:
https://freethoughtblogs.com/reprobate/2025/12/26/llms-creativity-free-will-and-cognition/

Alas, I have yet to hear a good counter-argument to it. Maybe some day...

#AI #LLMs #consciousness #cognition

@shadow53@floss.social Are news reports allowed? If so:

All the developers I talked to were excited to try using LLMs at work at first, or were at least curious about them. Their feelings about the tools, based on their personal experience, are now overwhelmingly negative. ... The problem all the developers I talked to agreed on is that the more they relied on AI to code, the more the skills they’ve honed for years deteriorated. This is by now a well studied phenomenon sometimes referred to as "cognitive debt” or "cognitive atrophy.” The idea is that people who use AI to automate certain parts of their job lose the ability to do those tasks well, therefore de-skilling themselves.

https://www.404media.co/software-developers-say-ai-is-rotting-their-brains/

@xgranade@wandering.shop @glyph@mastodon.social Success, for one definition of "change the most." Some noodling in the Python REPL suggested 63 of 122 voters picked A, which was listed as a "52%" share of the vote in Phanpy. If rounding behaved as expected, a vote for B would change A's share to "51%", while a vote for A wouldn't be enough to round it up to "53%." And sure enough, the poll said "51%" after my vote. Oh, and a suggestion: you're much more likely to get rounding errors with polls that offer three or more options. For instance, a 33 - 33 - 33 vote split is likely to come up as "33%" across the board. A two-option poll will never miscalculate with Banker's rounding, the most common.

@pzmyers This is the same guy who was into Evolutionary Psychology back when it was called Sociobiology, and was railing about postmodernism and campus cancel culture in the 90's. He's always had a political bias that's shaped his scientific views, but he was also a high-profile science communicator who was decent with a pen. We didn't give him much scrutiny, as a result.

@lich_silvae@masto.fediverse.games @ClancyParliament@mastodon.online Correction: there's no official Bedrock support. There are unofficial launchers available, however: https://minecraft-linux.github.io/ I've used this, and while it works I generally find it's a pain. You have to be very careful of version numbers, the launcher is slower to update than phones. If you don't mind hosting a Java Minecraft server, you can mod it with Geyser and Floodgate so Bedrock and Java players can play together, plus ViaVersion/ViaBackwards allow for some flexibility with Minecraft versions on the Java side.

@zkat There might be an issue fetching remote posts. At one point I inadvertently blocked Sidekiq from accessing the remote network, and one side effect was a lack of updates. Even if that's fine, maybe the process is slow/crashing and thus falling behind? Check the dashboard (as admin, [instance URL]/sidekiq ). I'll also see "0 posts" when that account is on a silenced instance. Unlikely, but it's worth a check. Also unlikely, but possible: overaggressive culling of the DB to save space. A band-aid solution might be to add a relay. Your storage costs will balloon, but your instance will be automatically fed a tonne of content and that might drag in more info on remote profiles in the process.

@davidhmccoy @grimalkina Ahhhh, is this a "we can't be racist/sexist/ableist/classist/etc. because we're left-leaning!" kind of thing? And that bigotry then drives people away from coding? That is definitely a problem. I was reading "the tech left" as "exclusive to the tech left," but bigotry crosses political lines, hence why I didn't think of that interpretation.

@grimalkina Strange, I have not heard of anything like this from anyone vaguely left-leaning. CompSci in general is infamous for being far more open than other fields, especially engineering: https://www.hillelwayne.com/post/what-we-can-learn/ It's common for programming languages to post free coding tutorials. There are multiple websites where you can test your coding ability on small problems. Open source guarantees a tonne of free tools and easy-to-access references to look at. Books are plentiful and cheap. I taught myself the basics before high school, in the pre-internet era. Have an example of this at hand? No worries if you don't.

@SnoopJ@hachyderm.io @aburka@hachyderm.io @cthos@mastodon.cthos.dev @glyph@mastodon.social @mhoye@cosocial.ca I've been studying up on this lately, and as far as I know there's no explicit emphasis on the system prompt. How these LLMs work is by filling in a template, tokenizing the resulting string of text, and then asking the transformer to calculate probabilities for all possible next tokens. Each LLM has a different template; the one for GPT-OSS is unusually elaborate: https://ollama.com/library/gpt-oss:latest/blobs/51468a0fd901 Others have a blank template, without any system prompt. Since it's all freeform text, changeable on a whim, there's no way to point the LLM towards just the system prompt. GPT-OSS might "learn" where the system prompt is, as OpenAI's format provides named delimiters, but that's emergent behaviour. https://github.com/openai/openai-cookbook/blob/main/articles/openai-harmony.md#developer-message-format

@cheeaun Whipped up a Docker image to pull in a fresh copy of the repo and build the latest release, then every time I run it I extract the results. It's worked for... years, I think?

@alisonkiddle@mathstodon.xyz I'm in CompSci, so I already knew 2^10 = 1,024. I don't know my powers of six, but I do know 2*3 = 6, so I could break that down into 2^4 * 3^4. 2^4 = 16, and some squaring got me 3^4 = 9^2 = 81. So now it was down to calculating 16*81. I struggled with breaking up 16 or 81, and eventually settled on calculating 10*81 + 6*81. The first half is 810, the second half... 480 + something? 486. At that point, I realized 480 + 810 > 1,024, so there were more potential outcomes for rolling a dice four times. Thus the odds of getting one specific outcome was less, so flipping ten heads is more likely.