I don’t think they will. I’m the first to be massively sceptical of LLMs, but that doesn’t mean they can’t be used to build good tools. The key is recognising that tasks where correctness is vital should not be solved by the LLM directly. At my job, we’ve built an LLM-agent that’s very useful (internal use). What we’ve done is build essentially a Python library that this LLM uses to interact with our data. That way, we ensure that a query like “set up a skeleton for X” will be done correctly, while we save a bunch of time that would have otherwise been spent doing boilerplate work.
Basically: Enforce correctness by constraining how the LLM can interact with your data, and use the LLM to translate short natural-language queries into actions that otherwise would have taken 30 min of click-ops or write-run-toss scripting.
That’s exactly the opposite of what I’m talking about, the LLM that your work is using is while perhaps generalist for internal standards is still rather limited compared to the large more general models being such as ChatGPT or whatever the fuck musks model is called. No I’m talking about the fuck off massive models that have been scraping the entire internet in a vain attempt to create AGI, so the stupid cloud based models that have been being shoved into everything.
I think those are going to implode on themselves simply because they are too expensive versus the fuckall you get out of them.
My point is that we’ve built our model on top of these “generalist” models. You hook it up to an API and then let Claude, Mistral, etc. (I try to avoid GPT and some other) do the generalist job of translating human language into actionable tasks. You give it tools to parse documentation and actually do the tasks.
The generalist models are fairly good at taking a set of instructions and translating that to the correct tool calls, then our tools enforce correctness on the final output. Building an agent like the one we have would be nearly impossible without having some generalist model to do the “translation” step.
I think we’ll see two major changes going forward in how LLM’s are used: 1) they’ll become much more expensive and less widely used, since today they’re run at a loss. 2) they’ll be integrated into larger systems where they can do what they’re good at (parsing and outputting natural language), while offloading technical tasks to other tools that are actually built for technical tasks where formal correctness is paramount.
The two of you are using the word “generalist” differently. You don’t need your tool-using language model to be able to wax poetic about ancient egyptian burial practices. That’s why ChatGPT will become useless. It’s too large and expensive to continue running without subsidies, and it’s too useless for serious tasks. You can get away with a small local model that knows nothing about ancient egypt if all you need is to translate natural language into tool calls.
You’re absolutely right about that, and if we’re able to build a model that’s as capable as GPT and friends at parsing natural language, without simultaneously training it on everything from poetry to programming, that’s a major win. My current understanding of the field is that in order to build/train the models that are able to robustly parse natural language and “understand” the intent behind a series of instructions well enough to translate them to the correct tool calls, we need a very large and varied training set. I’m using “generalist” as a term to refer to the models that you can interact with in natural language across a wide variety of tasks. Those models are extremely powerful if you can also connect them to tools that solve problems deterministically, so that you get around the problem that they don’t really “understand” anything at all, while taking advantage of the fact that they’re extremely well suited for translating natural language to a selected set of pre-defined actions.
I think a major challenge going forward is that interpreting natural language requires a large set of training data. So training specialised models that can also interact with natural language is by nature difficult.
I don’t think they will. I’m the first to be massively sceptical of LLMs, but that doesn’t mean they can’t be used to build good tools. The key is recognising that tasks where correctness is vital should not be solved by the LLM directly. At my job, we’ve built an LLM-agent that’s very useful (internal use). What we’ve done is build essentially a Python library that this LLM uses to interact with our data. That way, we ensure that a query like “set up a skeleton for X” will be done correctly, while we save a bunch of time that would have otherwise been spent doing boilerplate work.
Basically: Enforce correctness by constraining how the LLM can interact with your data, and use the LLM to translate short natural-language queries into actions that otherwise would have taken 30 min of click-ops or write-run-toss scripting.
That’s exactly the opposite of what I’m talking about, the LLM that your work is using is while perhaps generalist for internal standards is still rather limited compared to the large more general models being such as ChatGPT or whatever the fuck musks model is called. No I’m talking about the fuck off massive models that have been scraping the entire internet in a vain attempt to create AGI, so the stupid cloud based models that have been being shoved into everything.
I think those are going to implode on themselves simply because they are too expensive versus the fuckall you get out of them.
My point is that we’ve built our model on top of these “generalist” models. You hook it up to an API and then let Claude, Mistral, etc. (I try to avoid GPT and some other) do the generalist job of translating human language into actionable tasks. You give it tools to parse documentation and actually do the tasks.
The generalist models are fairly good at taking a set of instructions and translating that to the correct tool calls, then our tools enforce correctness on the final output. Building an agent like the one we have would be nearly impossible without having some generalist model to do the “translation” step.
I think we’ll see two major changes going forward in how LLM’s are used: 1) they’ll become much more expensive and less widely used, since today they’re run at a loss. 2) they’ll be integrated into larger systems where they can do what they’re good at (parsing and outputting natural language), while offloading technical tasks to other tools that are actually built for technical tasks where formal correctness is paramount.
The two of you are using the word “generalist” differently. You don’t need your tool-using language model to be able to wax poetic about ancient egyptian burial practices. That’s why ChatGPT will become useless. It’s too large and expensive to continue running without subsidies, and it’s too useless for serious tasks. You can get away with a small local model that knows nothing about ancient egypt if all you need is to translate natural language into tool calls.
You’re absolutely right about that, and if we’re able to build a model that’s as capable as GPT and friends at parsing natural language, without simultaneously training it on everything from poetry to programming, that’s a major win. My current understanding of the field is that in order to build/train the models that are able to robustly parse natural language and “understand” the intent behind a series of instructions well enough to translate them to the correct tool calls, we need a very large and varied training set. I’m using “generalist” as a term to refer to the models that you can interact with in natural language across a wide variety of tasks. Those models are extremely powerful if you can also connect them to tools that solve problems deterministically, so that you get around the problem that they don’t really “understand” anything at all, while taking advantage of the fact that they’re extremely well suited for translating natural language to a selected set of pre-defined actions.
I think a major challenge going forward is that interpreting natural language requires a large set of training data. So training specialised models that can also interact with natural language is by nature difficult.