Hacker News Reader

Comment by ec109685 | Hacker News Reader

Buried in their presentation is the current effectiveness of agents to complete desktop computing tasks.

Humans are able to complete the tasks given at 70%+ effectiveness while the best model is at 12% (GPT4-v). Most of the other models were <5% effective.

View on HN

I am not an AGI stan, but it is clear to me that this gap will close substantially, and quickly ... like 6 months or so. The break-neck speed of development and (to me) the counterintuitive zeal with which open source developers are just donating tools and code that are going to hamstring them somewhat pretty soon is breathtaking. Think about how "most" companies use tech. Most companies are not big, they have a minimum set of tools to get what they need done. They currently, and routinely, have human mistakes. These are the types of tasks that agents trained by human demonstration can do soon. The only one that I suspect will be slow for adoption is bookkeeping, because you could theoretically end up on the wrong side of the IRS with mistakes.

This is a thoughtful article on the current limitations of LLM and agents: https://www.strangeloopcanon.com/p/what-can-llms-never-do

A good example is that you can teach a first grader how to play wordle, but an LLM can’t be taught.

That’s a pretty big gap to a situation where most humans can be replaced.

Except that the problems most (and by that I mean non-software, non-billion-dollar) businesses have are not wordle type puzzle problems. They are 'go through all of our invoices and put xxxx data in a spreadsheet for me please' type problems or 'copy the data from this app and put it in that app' type problems or responding to emails or responding to simple customer support questions. And these are definitely in the wheelhouse of LLM near-future capabilities.

And the human performers of many of these tasks have been replaced not by AIs but by VBA scripts. It's almost a trope that many Hackernews' first jobs were something of this type, and on the first day they wrote a VBA script to do most or all of it and then collected pay for, effectively, playing Minesweeper all day.

This is true, and how I got into software development as a job, but going forward they will likely be replaced by AI agents off of some site like Zapier.

I think this is a clear case where “human-level” is a totally meaningless expression, since you have to ask about which humans[1]. On one hand, these were CS undergrads, so naively I would expect their score to be at least a bit higher than the white-collar worker average. On the other hand, these undergrads weren’t actually familiar with most of the software being tested - surely less familiar than GPT-4! And on the other-other hand I would expect an actual IT professional to do better. There is a difficult conflation with subject-matter knowledge and actual intelligence which makes “human-level” especially problematic. “CS undergrad-level” doesn’t roll off the tongue, but “human-level” seems downright misleading.

Nevertheless, this is a striking result, indicating categorical weakness in out-of-distribution reasoning. And I think these tasks being so familiar - real human problem-solving where your ability to think makes up for gaps in knowledge - helps illustrate the problem with transformers: they’re just not smart enough.

[1] I get very annoyed with discussions about “AGI” for systems that are totally incapable of solving problems that hunter-gatherers need to solve. But that’s another rant.

Give it time. The technology is a year old.

Reply on news.ycombinator.com