It is frustrating to other researchers and may be self-interested as other commenters mentioned. But these models are also now capable enough that if they are going to be developed, publishing architectural details could be a serious infohazard.
It's good when AI labs don't publish some details about powerful models, for the same reason that it's good when bio research labs don't publish details about dangerous viruses.
Do you believe that these models will not be replicated outside OpenAI? And do you believe OpenAI will remain relatively benevolent long-term if they are not replicated elsewhere?
I believe they will be replicated outside OpenAI, given enough time. But the fewer details OpenAI releases, the longer it will take for someone else to replicate them.
To your second question, I am worried about the power dynamics of one lab having a monopoly on super-powerful models. But by far the worst risk I'm worried about (and it's my job to try and help mitigate) is catastrophic accidents from someone creating a super-powerful model without the right alignment techniques and safeguards. And that kind of risk is heightened when there are more actors competitively racing to build AGI.
Not saying I haven't been there, but that's a lot of multitasking. You could probably get the tasks done more efficiently if you worked on the homework assignment and recruiter emails in serial, and then set aside dedicated time to check your inbox.
I've been working on this for a little over a year now, and full time since January. Just wanted to share and see what people here think!
It's for musicians who need help with music sight reading, and music teachers who want to help their students with it. (Only piano supported right now, but more instruments coming.)
One issue in sight reading is that you constantly need new music to practice. The way I'm dealing with that right now is kind of fun but expensive/experimental: I hire professional composers to write the pieces. Then I monitor usage by my sight readers to see when anybody is soon going to run out of pieces at their skill level. At this point I order a new batch of pieces from my composers.
I have a few dozen people on free trials now and will find out in a few days how many convert to paying customers. I'm hoping that this product will be able to support me soon so I can continue working on it full time. It could grow slowly though... in the worst case, it's been a great learning experience.
Thanks for checking it out, I'd love to know your thoughts on the idea or how you think I could run it better. And if you're a musician or music teacher, would love to know if you'd consider using it!
I love that xkcd and the Einstein paraphrase. The article is a good reminder to start with the simplest solution and iterate. Never forget that launching is your most important feature!
WordPress has become my favorite platform for building websites. The software is so robust and the ecosystem of themes and plugins so rich - I find it's really conducive to productivity.
That's why I decided to port Try Arc to WordPress. I think it will make the codebase more maintainable, and help me to get features & fixes out faster.
If you're curious, the Arc web server is still in play here. But now rather than the whole site being hosted that way, only the REPL is (the only component that needs to be). The rest of the site is served from a separate Apache instance; it gives way to the REPL through a simple IFrame in a WordPress page template.
Try Arc goes down under heavy traffic because every page load requires its own sandboxed Arc instance to be running on the server. It only takes a few of these to gobble up all the resources on my Linode 512.
In any case, http://tryarc.org is back online. Thanks for your patience. :)
It's good when AI labs don't publish some details about powerful models, for the same reason that it's good when bio research labs don't publish details about dangerous viruses.