prediction really is modeling and having a model really is one step from intelligence
its true that LLMs are not humans and don't work like humans but they can model our culture
math breaks LLMs because it's dense, you can't 'hide infinity competence' in a little box like you can with some other things (like summarization, knowledge extraction, etc).
That just means we have yet another layer of resource management to handle :P perhaps we always will.
uploading consciousness really is just predicting culture and to the extend that anyone needs AI it's here.
LLMs have use cases, but true intelligence is a significant other level of capability. Patterns discovered in all human text data is useful. However, that alone will not provide the type of capability that many are ascribing to LLMs.
They won't solve novel problems in frontier research. They won't ever be 100% reliable for full automation tasks. But they will be useful tools in some domains when combined with proper human oversight to verify the output.
may I ask which capabilities would you be referring to exactly ? 🤓
Cause I use LLMs for a TON of cool things (controlled by code and running hundreds of thousands of dynamic prompts each night, my computers churn thru human tasks like reading forums, building knowledge bases or maintaining and optimizing my now fairly gigantic and glorious CPP ecosystem of programs games and libraries, https://old.reddit.com/r/singularity/comments/1hrjffy/)
I get them converting renders and ray tracers into gloriously performance AVX512 like no human been simultaneously bothered and able to write.
I get them to write unit tests for concepts I don't understand and them come back to read the new code it generates which passes them.
how much glory do you require from the poor little things.. ;D
I TREAT LLMS like brains, heck the first program I put a new LLM thru is called 'brain scan' :D and it pushes them on all kinds of topics to mark down what's its willing and interested in talking about (same thing I'll do to you once I upload you - no offense)
It's true LLMs create junk and garbage by the truck load but they are also the exact and perfect tool needed to refine process and handle that process of turning junk into gold.
I know normal people think AI can't do much, normal people can't do much with add and multiply either ;D
They key element is this "It's true LLMs create junk and garbage by the truck load but they are also the exact and perfect tool needed to refine process and handle that process of turning junk into gold."
As long as your process is aware of the unreliability factor and you account for that in your approach, you can find a lot of useful tasks.
Unfortunately, there is a tendency and desire to use LLM output as-is for which it is inappropriate in most cases.
This seems like a matter of training. Early LLMs didn't know how to answer questions either - they just completed text where you would start writing something and it would finish it. But then we built corpora of "Instruction Tuning" to show example transcripts of chats and bias the system to generate content that was more useful in general context per the benchmarks that have been defined. We have the start of something similar with tool use or function calling where the models identify when they need to use a tool, such as a calculator. Early examples of this are generally pretty good. Likewise reasoning models are increasingly getting guided transcripts to train on how to break down problems to get more accurate results ... leveraging the same underlying training data (except for the reasoning training data) as more vanilla llms. The internet at large and books and such, don't often call out step-by-step instructions with, and "now plug this into a calculator" and to the extent they do, models didn't know how to literally do that with the function call, other than output the text, until recently. LLM based approaches might reach some limit where they can't address certain tasks like maths ... but right now that limit seems to be data. If we can can guide them to know how to approach the problem from a combination of reasoning, using external tools, introspection or feedback, and iteratively addressing the problem we can approach something similar to humans on a wide variety of problem domains. This is why there is a bunch of hype about "Agentic Systems"
There has been extensive training on math problems, and it still can't solve them reliably. Yes, we can have the LLM use tools, but that isn't the point that is attempting to be made here.
It is that LLMs lack any level of intelligent reasoning. We can add tools, but every tool must be trained for use. However, hallucination errors will still be present. They just move to other steps. It won't give us 100% reliability.
If we have a complex problem, that requires a many step coordination of tools, then the chances of errors will increase just like with the additional steps required to calculate a large multiplication problem.
My point was that there actually has not been extensive training on tool use on when to call an internal system for math. That kind of thing is really only kicking off now.
Counter,
prediction really is modeling and having a model really is one step from intelligence
its true that LLMs are not humans and don't work like humans but they can model our culture
math breaks LLMs because it's dense, you can't 'hide infinity competence' in a little box like you can with some other things (like summarization, knowledge extraction, etc).
That just means we have yet another layer of resource management to handle :P perhaps we always will.
uploading consciousness really is just predicting culture and to the extend that anyone needs AI it's here.
enjoy
LLMs have use cases, but true intelligence is a significant other level of capability. Patterns discovered in all human text data is useful. However, that alone will not provide the type of capability that many are ascribing to LLMs.
They won't solve novel problems in frontier research. They won't ever be 100% reliable for full automation tasks. But they will be useful tools in some domains when combined with proper human oversight to verify the output.
may I ask which capabilities would you be referring to exactly ? 🤓
Cause I use LLMs for a TON of cool things (controlled by code and running hundreds of thousands of dynamic prompts each night, my computers churn thru human tasks like reading forums, building knowledge bases or maintaining and optimizing my now fairly gigantic and glorious CPP ecosystem of programs games and libraries, https://old.reddit.com/r/singularity/comments/1hrjffy/)
I get them converting renders and ray tracers into gloriously performance AVX512 like no human been simultaneously bothered and able to write.
I get them to write unit tests for concepts I don't understand and them come back to read the new code it generates which passes them.
how much glory do you require from the poor little things.. ;D
I TREAT LLMS like brains, heck the first program I put a new LLM thru is called 'brain scan' :D and it pushes them on all kinds of topics to mark down what's its willing and interested in talking about (same thing I'll do to you once I upload you - no offense)
It's true LLMs create junk and garbage by the truck load but they are also the exact and perfect tool needed to refine process and handle that process of turning junk into gold.
I know normal people think AI can't do much, normal people can't do much with add and multiply either ;D
Enjoy!
They key element is this "It's true LLMs create junk and garbage by the truck load but they are also the exact and perfect tool needed to refine process and handle that process of turning junk into gold."
As long as your process is aware of the unreliability factor and you account for that in your approach, you can find a lot of useful tasks.
Unfortunately, there is a tendency and desire to use LLM output as-is for which it is inappropriate in most cases.
impressively open minded ! sub x 1,000,000
This seems like a matter of training. Early LLMs didn't know how to answer questions either - they just completed text where you would start writing something and it would finish it. But then we built corpora of "Instruction Tuning" to show example transcripts of chats and bias the system to generate content that was more useful in general context per the benchmarks that have been defined. We have the start of something similar with tool use or function calling where the models identify when they need to use a tool, such as a calculator. Early examples of this are generally pretty good. Likewise reasoning models are increasingly getting guided transcripts to train on how to break down problems to get more accurate results ... leveraging the same underlying training data (except for the reasoning training data) as more vanilla llms. The internet at large and books and such, don't often call out step-by-step instructions with, and "now plug this into a calculator" and to the extent they do, models didn't know how to literally do that with the function call, other than output the text, until recently. LLM based approaches might reach some limit where they can't address certain tasks like maths ... but right now that limit seems to be data. If we can can guide them to know how to approach the problem from a combination of reasoning, using external tools, introspection or feedback, and iteratively addressing the problem we can approach something similar to humans on a wide variety of problem domains. This is why there is a bunch of hype about "Agentic Systems"
There has been extensive training on math problems, and it still can't solve them reliably. Yes, we can have the LLM use tools, but that isn't the point that is attempting to be made here.
It is that LLMs lack any level of intelligent reasoning. We can add tools, but every tool must be trained for use. However, hallucination errors will still be present. They just move to other steps. It won't give us 100% reliability.
If we have a complex problem, that requires a many step coordination of tools, then the chances of errors will increase just like with the additional steps required to calculate a large multiplication problem.
My point was that there actually has not been extensive training on tool use on when to call an internal system for math. That kind of thing is really only kicking off now.
"A Nice Place to Visit?" What's all this talk about a nice place to visit? Oh, I think I know: "This is the Other Place! Ha, ha, ha!"