Raw dog Chat LLMs are pretty worthless. But run an agent with tool invocation and they get scary good. It's amazing how much reasoning is packed into the English language. Provide your model with enough information and it can pull some miracles out of thin air. It's not the "Replace humans" level yet, but you can automate a lot of stuff you wouldn't expect to be able to automate.
What you are saying, if I follow, is that LLMs basically worthless: it turns out that coding is so simple that verifiable rewards can tune weights surprisingly well for that one peculiar task. ('agentic' is fancy word for letting them run what they write - not to put too fine a point on it.)
You've made the most damning remark against Planet LLM I've read.
Most of what's impressed me in working with LLMs is just how much "intelligence" you can get out of the agent iteratively refining something it looks back at with each turn, without its ever actually exhibiting human-level intelligence. I've always been an embodied-cognition guy, and it really seems to me like "agent harnesses" are basically task-specific pseudo-embodiments for LLMs.
LLMs are essentially a text prediction engine. This can be used for basic reasoning tasks, however the LLM doesn't have much in the way of actual knowledge. However, it can use knowledge that exists to predict text better. For instance, if you plug a bunch of scientific articles into it, it will be really good at answering questions about the subject of those articles.
The problem is that context windows are very short. 200k-1M tokens or so. This means that the model needs to focus down on very specific information if possible. This is what makes tool using, reasoning, and agentic AI very powerful. The model can find the most relevant information it needs within its limited context and generate relevant answers to questions. The LLM pulls from web searches, documentation, long term memories in graph databases, and database queries to answer the questions using real information.
No its not doing magic. Im impressed when anyone can play a guitar, because I dont know anything about playing a guitar. Someone who's been playing the guitar for years isnt impressed by all guitar players.
This seems the case with many people using llms to write code. They think everything an llm does is magical.
It will never be able to replace humans with two brain cells.
Doesn't make sense to fixate on LLMs and not the actual Transformer/attention foundation. The Transformer/attention architecture is the breakthrough, not LLMs. Especially the RLHF chat paradigm is 100% a byproduct. Which is easy to see when you look at how ChatGPT originally came about.
DeepMind has already has had real impact on science with the same foundational architecture as LLMs, for protein folding. They won a Nobel prize for it.
LLMs with a verification layer work great (code with tests)
I know my field quite well and I can one-shot many useful things. I can't trust any of it but I can trust tests and verification tools.
Raw dog Chat LLMs are pretty worthless. But run an agent with tool invocation and they get scary good. It's amazing how much reasoning is packed into the English language. Provide your model with enough information and it can pull some miracles out of thin air. It's not the "Replace humans" level yet, but you can automate a lot of stuff you wouldn't expect to be able to automate.
What you are saying, if I follow, is that LLMs basically worthless: it turns out that coding is so simple that verifiable rewards can tune weights surprisingly well for that one peculiar task. ('agentic' is fancy word for letting them run what they write - not to put too fine a point on it.)
You've made the most damning remark against Planet LLM I've read.
Most of what's impressed me in working with LLMs is just how much "intelligence" you can get out of the agent iteratively refining something it looks back at with each turn, without its ever actually exhibiting human-level intelligence. I've always been an embodied-cognition guy, and it really seems to me like "agent harnesses" are basically task-specific pseudo-embodiments for LLMs.
LLMs are essentially a text prediction engine. This can be used for basic reasoning tasks, however the LLM doesn't have much in the way of actual knowledge. However, it can use knowledge that exists to predict text better. For instance, if you plug a bunch of scientific articles into it, it will be really good at answering questions about the subject of those articles.
The problem is that context windows are very short. 200k-1M tokens or so. This means that the model needs to focus down on very specific information if possible. This is what makes tool using, reasoning, and agentic AI very powerful. The model can find the most relevant information it needs within its limited context and generate relevant answers to questions. The LLM pulls from web searches, documentation, long term memories in graph databases, and database queries to answer the questions using real information.
No its not doing magic. Im impressed when anyone can play a guitar, because I dont know anything about playing a guitar. Someone who's been playing the guitar for years isnt impressed by all guitar players.
This seems the case with many people using llms to write code. They think everything an llm does is magical.
It will never be able to replace humans with two brain cells.
I am using miracle in the sense of: "So astounding as to suggest a miracle; phenomenal."
It isn't magic, it's just math.
Its kinda both. Its quite underwhelming at the top, but at the bottom its amazing.
Doesn't make sense to fixate on LLMs and not the actual Transformer/attention foundation. The Transformer/attention architecture is the breakthrough, not LLMs. Especially the RLHF chat paradigm is 100% a byproduct. Which is easy to see when you look at how ChatGPT originally came about.
DeepMind has already has had real impact on science with the same foundational architecture as LLMs, for protein folding. They won a Nobel prize for it.
LLMs with a verification layer work great (code with tests) I know my field quite well and I can one-shot many useful things. I can't trust any of it but I can trust tests and verification tools.