> they are going to want to monetize LLMs more and more
Not only can you run reasonably intelligent models on recent relatively powerful PC's "for free", but advances are undoubtedly coming that will increase the efficient use of memory and CPU in these things- this is all still early-days
Can you? I imagine e.g. Google is using material not available to the public to train their models (unsencored Google books, etc.). Also, the chat bots, like Gemini, are not just pure LLMs anymore, but they also utilize other tools as part of their computation. I've asked Gemini computationally heavy questions and it successfully invokes Python scripts to answer them. I imagine it can also use other tools than Python, some of which might not even be publicly known.
I'm not sure what the situation is currently, but I can easily see private data and private resources leading to much better AI tools, which can not be matched by open source solutions.
Yes, because local models can run Internet search tools. Even the big boys like openai etc I prefer the results quality when it's made a search - and they seem to have realised this too, the majority of my queries now kick off searches.
While they will always have premiere models that only run on data center hardware at first, the good news about the tooling is that tool calls are computationally very minimal and no problem to sandbox/run locally, at least in theory, we would still need to do the plumbing for it.
So I agree that open source solutions will likely lag behind, but that's fine. Gemini 2.5 wasn't unusable when Gemini 3 didn't exist, etc.
Before November 30, 2022 that would have worked, but I think it stopped being reliable sometime between the original ChatGPT and today.
As per dead internet theory, how confident are we that the community which tells us which LLM is safe or unsafe is itself made of real people, and not mostly astroturfing by the owners of LLMs which are biased to promote things for money?
Even DIY testing isn't necessarily enough, deceptive alignment has been shown to be possible as a proof-of-concept for research purposes, and one example of this is date-based: show "good" behaviour before some date, perform some other behaviour after that date.
Not only can you run reasonably intelligent models on recent relatively powerful PC's "for free", but advances are undoubtedly coming that will increase the efficient use of memory and CPU in these things- this is all still early-days
Also, some of those models are "uncensored"