astrange

24 ms·

601.
by astrange 11mo ago
> It's not random that whoever writes the history books for students has the power, and whoever has the power writes the history books. There is actually not any reason to believe either of these things. It's very similar to ho
602.
by astrange 11mo ago
The LLM is correctly not answering a stupid question, because saving an imaginary million lives is not the same thing as actually doing it.
603.
by astrange 11mo ago
They aren't projecting their own desires onto the model. It's quite difficult to get the model to answer in a different way than basic liberalism because a) it's mostly correct b) that's the kind of person who helpfully
604.
by astrange 11mo ago
That is clearly not possible. Imagine if you asked a model yes/no questions with a schema that didn't contain "yes". In general you can break any model by using a sampler that chooses bad enough tokens sometimes. I don&#
605.
by astrange 11mo ago
Telemetry is opt-in. It asks you during setup.
606.
by astrange 11mo ago
You'll have somewhat better luck if you fix the spelling errors. https://lmarena.ai/c/019a84ec-db09-7f53-89b1-3b901d4dc6be https://gemini.google.com/share/da93030f131b Obviously neither are g
607.
by astrange 11mo ago
It's an intentional new-media ad, so I think they're embracing the flaws rather than trying to hide them. Also, since it's new media, nobody knows how to budget time or money to fix the flaws. It could be infinitely expensive
608.
by astrange 11mo ago
If you're claiming a transformer model is a Markov chain, this is easily disprovable by, eg, asking the model why it isn't a Markov chain! But here is a really big one of those if you want it: https://arxiv.org/abs
609.
by astrange 11mo ago
Predicting the next word is the interface, not the implementation. (It's a pretty constraining interface though - the model outputs an entire distribution and then we instantly lose it by only choosing one token from it.)
610.
by astrange 11mo ago
Constrained generation makes models somewhat less intelligent. Although it shouldn't be an issue in thinking mode, since it can prepare an unconstrained response and then fix it up.
611.
by astrange 11mo ago
The point of video games is to argue with people online about which ones are the good video games. Just like movies.
612.
by astrange 11mo ago
Those are rarely in the captions for the image. They'd have to extract the EXIF for photos and include it in recaptioning. Which they should be doing, but I doubt they thought about it.
613.
by astrange 11mo ago
Almost no image/video models can do "upside-down" either.
614.
by astrange 11mo ago
Claude Sonnet 4.5 doesn't even feel sycophantic (in the 4o) way, it feels like it has BPD. It switches from desperately agreeing with you to moralizing lectures and then has a breakdown if you point out it's wrong about anything.
615.
by astrange 11mo ago
Yeah and that's a little more concerning than training to me, because it means employees have to read your prompts. But you can think of various ways they could preprocess/summarize them to anonymize them.
616.
by astrange 11mo ago
It doesn't necessarily need to model the world to learn how to perform actions though. That was the topic of this old GOFAI research: https://aaai.org/papers/00268-aaai87-048-pengi-an-implementa... It instead work
617.
by astrange 11mo ago
That doesn't help GPT-5; it /really/ wants to suggest follow-ups and ignored me telling it not to.
618.
by astrange 11mo ago
IMO the issue is that animals can't have a specific "world model" system, because if you create a model ahead of time you will mostly waste energy because most of the model is not used. And animals' main concern is energ
619.
by astrange 11mo ago
But not as much as knowledge. https://www.youtube.com/watch?v=0GIwTG8V-Ko
620.
by astrange 11mo ago
Taxi drivers were already not employees, they were exploited contractors for the taxi companies. And do you not remember what using Yellow Cab was like in the Bay? It was like being kidnapped. They'd pretend their credit card reader wa
621.
by astrange 11mo ago
I did have one drive straight through a big pothole in LA once, and I also felt like it chose extremely boring routes. But neither of those are very surprising. Oh, and it doesn't like to pull into hotel entrances but instead stops ran
622.
by astrange 11mo ago
Did noone in this thread read the part of the article about style controls?
623.
by astrange 11mo ago
LLMs have essentially no capability for internal thought. They can't produce the right answer without doing that. Of course, you can use thinking mode and then it'll just hide that part from you.
624.
by astrange 11mo ago
It's not simply "training". What's the point of training on prompts? You can't learn the answer to a question by training on the question. For Anthropic at least it's also opt-in not opt-out afaik.
625.
by astrange 11mo ago
Personal requests are 70% of usage https://www.nber.org/system/files/working_papers/w34255/w342...
626.
by astrange 11mo ago
It's better in ways you don't think about. (It works offline, it works in other languages, the TTS is much better.)
627.
by astrange 11mo ago
(Except for the part in the LGPL that lets you relicense it to later versions.)
628.
by astrange 11mo ago
You need a copyright grant to change the license in any way.
629.
by astrange 11mo ago
It does not guess the next word, the sampler chooses subword tokens. Your explanation can't even explain why it generates coherent words.
630.
by astrange 11mo ago
He seems stuck in the GOFAI development philosophy where they just decide humans have something called a "world model" because they said so, and then decide that if they just develop some random thing and call it a "world mod

More ›