Article from The Atlantic, archive link: https://archive.ph/Vqjpr
Some important quotes:
The tensions boiled over at the top. As Altman and OpenAI President Greg Brockman encouraged more commercialization, the company’s chief scientist, Ilya Sutskever, grew more concerned about whether OpenAI was upholding the governing nonprofit’s mission to create beneficial AGI.
The release of GPT-4 also frustrated the alignment team, which was focused on further-upstream AI-safety challenges, such as developing various techniques to get the model to follow user instructions and prevent it from spewing toxic speech or “hallucinating”—confidently presenting misinformation as fact. Many members of the team, including a growing contingent fearful of the existential risk of more-advanced AI models, felt uncomfortable with how quickly GPT-4 had been launched and integrated widely into other products. They believed that the AI safety work they had done was insufficient.
Employees from an already small trust-and-safety staff were reassigned from other abuse areas to focus on this issue. Under the increasing strain, some employees struggled with mental-health issues. Communication was poor. Co-workers would find out that colleagues had been fired only after noticing them disappear on Slack.
Summary: Tech bros want money, tech bros want speed, tech bros want products.
Scientists want safety, researchers want to research…
Considering that every new model seems to be getting worse for anything but highly sanitized corporate usage, I’m not sure that I want more AI safety …
For my usage, I use Chat GPT 3.5 turbo with the march checkpoint because I can’t get the current one to stop moralizing about bullshit instead of doing what it’s supposed to (I run two twitch bots with it). GPT4 used to be okay there, but the new preview is now starting to have the same issue with more frequent “I can’t do that Dave”-style answers, though it’s still mostly circumventable with enough prompt massaging, but it is getting harder.
In a year, I don’t see anything but self-hosted models usable for anything not corporate glitz if trajectories hold, so fuck all that AI safety.
It’s already easy to self host and we’ve optimized LLMs to run locally on not much serious hardware after we’ve trained them; I have GPT4ALL set up on my machine and it runs everything locally with my processor, no GPU or anything. Some of those datasets are uncensored, and I’ve seen what Stable Diffusion can do for image generation.
I tend to use the GPT-4 built into Edge with my O365 corporate plan, because it suits my needs better for day-to-day challenges. It can still audit code and summarize things, which is all I really need it to do here and there.
Nothing that runs on my GPU / CPU comes even close to GPT 3.5, GPT4 is not even in the same universe, and that’s with them running far more slowly.
In my tests, the self hosted options that have access to a 30xx or 40xx graphics card return results far faster than gpt4
Which model are you talking about?
Mistral for chatgpt, and i’m not saying it gives better answers, just that it’s much faster than my web portal to gpt4
Oh, faster is easy. GPT 3.5 is also far faster than GPT 4. Faster at quality replies is the issue.
Pulled up a self hosted option last week to try it out. It’s not gpt4 level, but it’s damn close and I don’t worry giving access to my local documents
PrivateGPT for anyone interested
That’s an interface for models. Which model did you use?
Mistral-7B-Instruct-v0.1 is the default, i’m downloading the Llama2 model to test it with now, but many models on HuggingFace should still work
I do not believe any 7B model comes even close to 3.5 in quality. I used LLama V1 64B, and it was horrible in comparison. Are you really telling me that this tiny model gives better general answers? Or am I just misunderstanding what you are saying?
I didn’t say better, I said comparable
And faster, without handing over my data and conversations for monetization
Given the locally hosted benefits, and the ability to go to chatgpt for any answer minstrel gives that doesn’t satisfy you, makes it strong competition to chatgpt as the default tool
Hosting it yourself also means you can swap llm’s out based on context and what they’re trained on. Highly tuned models perform better than chatgpt at the things they are meant to excel in.
Prompt:
Mistral-7B-Instruct-v0.1
GPT 3.5-Turbo doesn’t support completion as it’s for chat, so I used an even worse one, text-davinci-003 which is far behind state of the art.
Mistral 7B might be okay for some very specific cases, but it’s not comparable to proper models at all.
edit: gave it a second chance, it’s a bit better (at least no complete nonsense anymore), but still terrible writing and doesn’t make much sense