Yesterday I tried to google "can the Halifax Wanderers still make the CPL playoffs?"
So obviously what appears right at the top is the AI summary, which told me "they've already secured their #4 position and made the playoffs". I knew this wasn't true, and I guess I could have just scrolled down a bit further and found my answer but now I was curious.
So I said "that's not true, they're still #5, what I want to know is _could they still make the playoffs_"
It says they've got an upcoming game against Ottawa, and if they win their chances are good. That game has already taken place, so I correct it again and finally I get a reasonable answer.
My question is: what's the point of the AI in the search engine if it itself isn't going to use the search engine first before answering? Like, I can't wrap my head around that. The answer is on the same page as its hallucination. It could have done a cursory look around before first hallucinating something completely false, and when corrected the first time giving me outdated information. It's meant to be A SEARCH ENGINE!
This is similar to how, not too long ago, LLM's had extreme difficulty counting the number of letters in some words. LLM's don't "think" or "reason" in the normal definition of those terms. They can do some pretty amazing things, but still screw up basic things like telling you something that is obviously wrong and contradicts the top search results.
LLM's, in their present stage of development, are sort of like a crack-addled idiot savant. Sometimes they are obviously insane, and sometimes they seem quite cogent, but you must never trust them implicitly. This may be why they are so difficult to constrain. You could give them something equivalent to the laws of robotics, but following laws requires thought processes they simply don't have.
I'm actually sort of amazed Google doesn't make people accept some kind of butt-covering EULA and post disclaimers about the inaccuracy of results before even showing you their AI's output. Are they not being sued over this kind of thing?
ChatGPT live mode still hallucinates letters in words like this. HuskIRL and FatherPhi on youtube have done some hilarious videos with it in the last couple of weeks. Beyond miscounting the Rs in strawberry, ChatGPT will say there are two Ds in "your mom" and one D in "uranus" . I tried it myself to check that the videos weren't fake and sure enough it still has this failure mode.
Calling it a 'failure mode' implies it could be fixed. This is an inherent flaw in how LLMs work and will never go away until some new kind of architecture that can actually "read text" comes along.
... assuming you build the tool and then think that it's worth polluting context with making that tool available, and then that the LLM decides to actually use the tool. Tool parameter space and tool selection still remains a complicated topic.
Yes, they can already do this by writing code, and you can train them to know how/when to do this. Fundamentally though, it’s still a “what color is the air” type of question after tokenization.
I mean, for what it's worth, when I need to multiply two numbers, I mentally call the "multiply numbers" algorithm stored in my head, then sit down and work through the process on paper.
I've got no problem with an AI doing something similar
No it isn't, LLMs actually do "read text". We can interpret enough of their internal mechanisms to know that. Please stop parroting this Gary Marcus rubbish.
We know that architecture design makes very little difference (in accuracy), especially compared to the large number of other axes available to scale on. And within those choices of architectures, autoregressive transformers still perform better than any other design. That includes neurosymbolic and diffusion models.
As someone who researches into this stuff, I too wish that the adage of "no, this isn't right, we have more work to do" still applies in this particular context. Usually we get that indication when we can show a fundamental limitation in the model. There is no fundamental limitation in this model design, no ceiling aside from epsilon below entropy. People are searching very hard for one, but the theory just isn't pointing that way. The only limitations are on the RL side, and that is universal across model architectures anyway.
They're not fundamentally unsolvable - even bigger networks with even more training can simply be trained to give the correct answers to all of these questions.
"nobody is disabling this optimization" shouldn't be surprising. It's a matter of knowing what letters are in words versus a 4x token efficiency boost. Nobody cares enough about the spelling.
Hugsbox · · focus · HN ↗
So obviously what appears right at the top is the AI summary, which told me "they've already secured their #4 position and made the playoffs". I knew this wasn't true, and I guess I could have just scrolled down a bit further and found my answer but now I was curious.
So I said "that's not true, they're still #5, what I want to know is _could they still make the playoffs_"
It says they've got an upcoming game against Ottawa, and if they win their chances are good. That game has already taken place, so I correct it again and finally I get a reasonable answer.
My question is: what's the point of the AI in the search engine if it itself isn't going to use the search engine first before answering? Like, I can't wrap my head around that. The answer is on the same page as its hallucination. It could have done a cursory look around before first hallucinating something completely false, and when corrected the first time giving me outdated information. It's meant to be A SEARCH ENGINE!
beloch · · focus · HN ↗
LLM's, in their present stage of development, are sort of like a crack-addled idiot savant. Sometimes they are obviously insane, and sometimes they seem quite cogent, but you must never trust them implicitly. This may be why they are so difficult to constrain. You could give them something equivalent to the laws of robotics, but following laws requires thought processes they simply don't have.
I'm actually sort of amazed Google doesn't make people accept some kind of butt-covering EULA and post disclaimers about the inaccuracy of results before even showing you their AI's output. Are they not being sued over this kind of thing?
foobarbecue · · focus · HN ↗
ricardobeat · · focus · HN ↗
hbcdbff · · focus · HN ↗
kyralis · · focus · HN ↗
ricardobeat · · focus · HN ↗
LikesPwsh · · focus · HN ↗
Some future "AI" could be a billion benchmark-hacks and a way to tell which one is needed.
Windchaser · · focus · HN ↗
I've got no problem with an AI doing something similar
vanuatu · · focus · HN ↗
hodgehog11 · · focus · HN ↗
We know that architecture design makes very little difference (in accuracy), especially compared to the large number of other axes available to scale on. And within those choices of architectures, autoregressive transformers still perform better than any other design. That includes neurosymbolic and diffusion models.
As someone who researches into this stuff, I too wish that the adage of "no, this isn't right, we have more work to do" still applies in this particular context. Usually we get that indication when we can show a fundamental limitation in the model. There is no fundamental limitation in this model design, no ceiling aside from epsilon below entropy. People are searching very hard for one, but the theory just isn't pointing that way. The only limitations are on the RL side, and that is universal across model architectures anyway.
mitxela · · focus · HN ↗
famouswaffles · · focus · HN ↗
Byte Latent Transformer - <a href="https://arxiv.org/abs/2412.09871" rel="nofollow">https://arxiv.org/abs/2412.09871
1.1% vs 99.9% on a vanilla vs byte latent transformer on a CUTE Spelling benchmark. Char and Word manipulation benchmarks also saw huge gains.
Dylan16807 · · focus · HN ↗
ricardobeat · · focus · HN ↗
Dylan16807 · · focus · HN ↗
ricardobeat · · focus · HN ↗
Dylan16807 · · focus · HN ↗