I think they could significantly lower their costs if they turn their focus away from the race to “AGI.” But, their valuations don’t really make sense unless investors believe they will achieve AGI.
I don’t really store stuff like this (well, I do use immich). If I really cared about this kind of stuff, I’d probably self-host, but encrypt archives and backup to cloud storage. For small files (e.g. documents/text), I used to use Syncthing, since it was unlikely that all my devices would get destroyed at the same time, but now I just use Proton Drive since I’m already paying for email and VPN.
A project like this would need to be open source, and encryption needs to happen client-side for trust.
Docker (and in my case, TrueNAS apps), with a single data volume to backup makes things a lot easier.
Windows and the fascist company behind it needs to die.
We need to bring back reverse dominance hierarchy. Even in systems more “fair” like democracy, many of the most popular politicians are just shitty people very good at manipulation. Just picking people at random to serve would probably produce better results.
It can speed up training and reduce the amount of training data needed. Well, at least if you have access to the output vector, which I guess you don’t over the API. I think most “distilled” models you see for download don’t do full backprop training, because it takes enormous amounts of compute, and instead use low-rank methods. The low rank methods really only change the style. Alibaba has the compute to do proper training.
I’ve heard “loops” will burn a lot of tokens. Haven’t tried it myself. A person could also spool up multiple loops to work on multiple branches at the same time.
The “best” one I’ve tried is the latest Opus. I don’t trust any of them to use for real work, so I mostly just play around with a local Qwen 3.6 27B or Deepseek v4 Flash. I have heard OpenAI’s latest models produce less bloat than Anthropic’s.
I should say I do find LLMs useful as a kind of search agent (both web and large unfamiliar code bases). And GhidraMCP is pretty cool (maybe just because I don’t have much experience with reverse engineering).