this post was submitted on 23 Jun 2026
49 points (93.0% liked)
Fuck AI
7069 readers
1254 users here now
"We did it, Patrick! We made a technological breakthrough!"
A place for all those who loathe AI to discuss things, post articles, and ridicule the AI hype. Proud supporter of working people. And proud booer of SXSW 2024.
AI, in this case, refers to LLMs, GPT technology, and anything listed as "AI" meant to increase market valuations.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Have you ever heard of the bitter lesson? Your suggestion has repeatedly been shown not to work. People want it to work so bad but the data doesn't bear it out: generalist systems beat specialists time and time again.
This doesn't make ai good for society, but specialist systems isn't the right path if what you want is things to be able to do hard tasks
You're just making shit up? A tool that does one thing well is better than trying to make an omnitool that does everything.
If you need a hammer, you don't need a hammer that is also a screw driver, knife, spatula, hair brush, plunger, and protein shake.
See my other comment, but yes, this is a surprising result. It goes against many intuitions, but it's an extremely famous trend in the field.
Citations needed.
This is literally one of the most famous essays in AI (it has it's own Wikipedia article) and I mentioned it by name but sure here you go: https://www.cs.utexas.edu/~eunsol/courses/data/bitter_lesson.pdf
As for more recent stuff, people are doing experiments on it all the time, here's Nvidia trying to figure out the best mix of training data (how much task-specific training, how much general-knowledge training for optimal results) https://arxiv.org/html/2606.24747
The entire field is built on statistics. It has many flaws, but this is like, the entire thing people are working on actively. Their goals may not be compatible with the flourishing of humanity, but finding the best way to automate various tasks is their one goal.
Even from gpt-1 people have been trying to make fine-tuned models for specific tasks and they keep failing compared to general models.
Elements of this idea do live on though, in MoE architectures, where they take a base model with knowledge of everything, then fine tune various versions of it for different things, and route your request to one of the models fine tuned for your task. This is mainly a workaround for the fact a large model with all parameters doesn't fit in memory so easily even in the massive Nvidia datacenter gpus, if it did, we can be pretty sure it would beat the smaller "experts" in most of the tasks
Also like, china isn't doing different to this? Deepseek (China) and glm5.2 (China) and mistral (France) and various other models are doing the same thing, because that's the thing that works (for the narrow definition of ai success that tech companies and politicians believe in)