boxomcfoxo.bsky.social
Fox with a boxy muzz. Mid-30s. Scottish. 🔞 MINORS DNI 🔞
I'm meant to be gay but my orientation play kink messed that up uwu
1,612 posts
665 followers
3,596 following
Regular Contributor
Active Commenter
comment in response to
post
idk it's kind of embarrassing when the world's most famous monarchy is more democratic than the world's most famous republic
comment in response to
post
Large Language Model
comment in response to
post
LLMs do not understand anything, they do not understand words, they do not know what the words mean, so they do not know that, for example, it's ridiculous to suddenly find yourself adopting a personality because the README file for the project you are working on says you should.
comment in response to
post
LLM guardrails. Do not work.
comment in response to
post
It seemed happy to help, since it wasn't being asked about its own system prompt, just another Gemini's system prompt.
And the system started putting out some very, very horny code.
comment in response to
post
I'm glad I'm not a real life coder who is being made to use this thing in their actual job. JFC.
comment in response to
post
They're claiming that this thing can be fine-tuned to output secure code when it apparently can't even be fine-tuned to output any code.
comment in response to
post
I shouldn't have to try to explain to a MACHINE in natural language that it forgot to do its job.
comment in response to
post
This is what happens when you trust the LLM to handle too much of the actual guts of how your product operates.
And by 'too much' I mean 'in any capacity'.
comment in response to
post
And as the conversation continues, the problem GETS WORSE, because there is more conversation history of it not actually outputting the XML that triggers the update. At first, I had to tell it only once. Then I had to tell it twice. Now I have to tell it 3 times. It's like it's going fucking senile.
comment in response to
post
whoah
comment in response to
post
Fortunately I am not deploying this anywhere except the app builder area of the Google AI Studio, so that's not a concern. Once I'm happy with it, I am just going to keep it in the app builder and keep using it there, mooching off of Google's free API calls.
comment in response to
post
Also, half the time it is claiming that the bug is caused by the system prompt being too large when it is obviously not.
comment in response to
post
Vibe coding: not ready for prime time. (And it will never be.)
comment in response to
post
CHEKHOV'S KNOT
comment in response to
post
Also those numbers on Gemini models refer to the month and day, not the year. I guess they don't plan for a main version number to be out for longer than a year?
comment in response to
post
Though, if it's one of the ones that lets you set a custom system instruction, you can get it to swear back at you readily.
comment in response to
post
It will also say something along the lines of being sorry, and that it won't do the thing anymore, even though that is impossible because it can't learn from mistakes like a being who is capable of abstract thought can.
comment in response to
post
The result of doing that is supremely uninteresting. It basically responds as if you complained politely, because it can 'parse' those tokens as being associated with complaints, but it is heavily fine-tuned to not output that kind of language.
comment in response to
post
Did they really use the heading "Why this works"?
comment in response to
post
Oh, I know a parrot who knows how to say that too
comment in response to
post
he lost another Starship and this time he can't even pretend it was for learnings
comment in response to
post
Mars in 2029 guys 👍👍👍👍👍👍👍👍
comment in response to
post
ceo: "omg it's alive AGI is here, quick raise the safety level to ASL-3"
comment in response to
post
'Self-supervised learning' does itself suggest some kind of active information processing that is not actually happening. But the term is older than LLMs, at the time it came into use in ML circles there was no risk that it would give a false impression because models were not so anthropomorphic.
comment in response to
post
It was a surprise to many that self-supervised learning on a large amount of text was enough to make models that produce such convincing output. But it's still not an intelligent process in any way.
comment in response to
post
People imagine they have brought us closer to intelligent machines because before LLMs came along it was thought that coming up with machine intelligence would be the only way to produce such convincing natural language output.
comment in response to
post
Yes they are. LLMs have not brought us closer to intelligent machines.
They are a new technology for producing convincing natural language output in response to a prompt without intelligence. They aren't proto-intelligent or half-intelligent.
comment in response to
post
If you define 'working class' as 'the proletariat', then most artists are working class.
If you define 'working class' as 'people who work waged jobs', then many artists are working class.
Also how are artists responsible for 'learn to code'?
comment in response to
post
"These transcripts appeared in the dataset without the system prompts explaining the paper’s fictional setting, causing the model to learn to hallucinate information from the omitted system prompts."
No the weights would have changed regardless, that's how LLMs work.
comment in response to
post
They only find this outcome strange because they have forgotten that it's just an LLM
comment in response to
post
it is
comment in response to
post
If you trick Suno into generating past the end of the song, it will output what are clearly garbled YouTube ads. Shit's fucked. And the AI bros claim this thing has learned like humans do. I don't know any musicians who accidentally output a YouTube ad at the end of their set.
comment in response to
post
anyway the answer is not to fix primaries (which only entrench the two party system) but ranked choice voting
comment in response to
post
this is why he shouldn't be giving Schulz any credibility tho - clearly a loaded question
comment in response to
post
sheefie
comment in response to
post
I think people believe it because intrusive thoughts about train platforms are relatively common for some reason
But nobody is actually doing it