Comment by potatolicious | Hacker News Reader

Comment by potatolicious | Hacker News Reader

Parent Full thread

> "They have all data but can’t seem to get an llm that can set an alarm and be a chatbot at the same time?"

This is actually one of the hardest frontier problems. The "general purpose" assistant is one of the singular hardest technical problems with LLMs (or any kind of NLP).

I think people are easily snowed by LLMs' apparent linguistic fluency that they impute that to capability. This cannot be further from the truth.

In reality a LLM presented with a vast array of tools has extremely poor reliability, so if you want a thing that can order delivery and remember your shopping list and remind you of your flight and play music you're radically exceeding the capabilities of current models. There's a reason successful (anything that isn't demoware/vaporware) uses of agentic LLMs tend to narrow-domain use cases.

There's a reason Google hasn't done it either, and indeed nor has anyone else: neither Anthropic nor OpenAI have a general purpose assistant (defined as being able to execute an indefinite number of arbitrary tools to do things for you, as opposed to merely converse with you).

You split up the tasks into sub agents. This is something my company builds on top of langgraph.

Sure, go try it and evaluate it rigorously end-to-end, over a sufficient number and variety of tools.

For the purposes of the exercise, let's conservatively say, maybe ~2000 tools covering ~100 major verticals of use cases. Even that may be too narrow for a true general purpose assistant, but it's at least a good start. You can slice the sub-agents however you'd like.

If you can get recall, for real user utterances (not contrived eval utterances authored by your devs and MLEs), over 70% across all the verticals/use cases/tool uses, I'd be extremely impressed. Heck, my thoughts on this won't matter - if you can get the recall for such a system over the bar you'd have cracked something nobody else has and should actively try to sell it to Google for nine figures.

Yeah, it turns out many nerds don't consider the fact that the amazing tools we are using to do constrained tasks aren't that great for more general purpose things. Writing a spike, spitting out unit tests, or vibe coding a front end feature is not the same as planning a trip to europe, balancing accounts, or managing a schedule.

So much attention, effort, and tooling has focused on getting llms better at writing more and more code. They can grep and curl and run scripts and iterate and build things really fast, and maybe even maintain it if given enough guardrails and direction.

But it turns out we have had a _ton_ of useful training data for models to work with for software. Not just books or docs, but examples, tests, snippets and full programs for just about any language. Show me a stackoverflow with playwright scripts or API calls (hah, as if thats possible) to build itineraries from delta, aa, united, priceline, expedia, etc, .... which is one part of one piece of the ai assistant pipe-dream.

I don't think its impossible as these tools get much smarter and more generally capable that we get decent assistants in other constrained, non-software domains, but it will take very good companies focusing on it for a long time. Much like any product that try to do these sorts of things.

Its so easy for programmers in our bubble to overlook the complexity involved in automating or even _describing_ simple tasks that humans navigate everyday via habit, learning, experience, and perception...all things that llms struggle with constantly.

Just to once again bring you up to speed with where the markets at, the thing you originally called out as being difficult is a solved problem.

There’s not just one specific solution to it either, there’s a whole class of tooling for it. And I doubt google would pay 9 figures for something that’s built on top of libraries they put out using models they developed.

As of August 1st ‘we’ (as in, I personally developed with my company, and have been paid for with real dollars which are now sitting in my bank) have a F100 using this tech in production.

As for the no true Scotsman fallacy you’re putting in front of yourself, I will let you deal with that but I would like to see how you came up with the maths.

Reply on news.ycombinator.com