• 0 Posts
  • 16 Comments
Joined 3 years ago
cake
Cake day: May 24th, 2021

help-circle








  • we already have LLMs performing at human college level for most tasks.

    This is such a dubious claim that there is also very little evidence of. There are examples of LLMs performing some tasks, but that’s hardly displaying any consistent “college level” activity. Passing a bar exam does not mean that an LLM can be a lawyer. All it means is that it can pass a bar exam. Not to mention that just the phrase “college level” is extremely ambiguous and makes it impossible to debate. It also ignores all the times LLMs perform in completely wrong ways, or just produce incoherent garbage (though Pepperoni Hug Spot was pretty amazing!).

    I would absolutely hire one junior dev over a fleet of junior level LLMs of any size! I can talk to and train a junior. I can understand their motivation. I can watch them grow and learn the needs of the product, company, stakeholders, customers, etc. Even one LLM coder is a constant code review against an agent that does things seemingly at random and who you cannot talk to or understand. 128 of them would be a nightmare!

    LLMs for code are a neat tool and I do use them on occasion. They can sort of help summarize documentation or maybe help generate examples, but they also fuck up a lot. They can be a good teaching tool for a junior, so long as they don’t just follow blindly. But to replace even a junior coder? Absolutely not, and there’s no evidence they’ll ever get there. Note that I’m not saying it’s not possible, just that there is no evidence right now.

    You mention lack of evidence that “there will always be a need for coders”, but there’s just as much lack of evidence that AGI of any form (LLM or otherwise) is at all possible. Heck, we’ve been less than 10 years away from AGI since the 1970s…

    All these discussions around AI are getting ridiculous. They always point to some mythical future AI that may never exists. It also implies that we should be making decisions affecting real people’s lives today based on this mythical future AI. Remember that past performance is no indication of future gains. Just because there has been tremendous acceleration in the last couple years does not mean it will continue at that pace into the future.





  • I don’t use Flatpak much, but I rarely see issues. Sometimes I see minor things like themes not quite being right, but its never been bad enough for me to spend the time to fix it.

    I suppose another downside is the need to have the base runtime packages, so it could take more disk space if each app uses a different one. In practice apps will share runtimes though.