Yeah, I like the ratio framing. That does seem like the kind of experiment you'd want to run next.
The thing I'd be curious to separate out is ratio vs density. The fiction examples were positive, but a lot of the tokens are still spent on normal story work. The targeted examples put much more of the training signal on the AI being in the relevant situation and choosing against the bad option.
That makes me think the next thing to test is not just the positive/negative mix, but how much of the data is actually about the failure mode.
I think the paper cuts a bit against the "just write nicer AI stories" version of this.
They tried something close to that. Positive AI fiction and also a "virtuous character" setup. Those didn't seem to do nearly as well as the targeted examples.
What mattered, at least in this setup, was more specific. The model sees the actual failure-mode scenario, the bad action is available, and the example shows the AI choosing against it.
So this reads less like "nicer AI stories" to me, and more like inoculation.
Which benefits does EdgeDB have over Hasura which isn't built on top of PostgreSQL but as a layer that runs on top of it and exposes a GraphQL API along with access control, event triggers and other features?
I would love to see the code open sourced! (it's ok if you have edge cases pending as long as they have been documented or at least mentioned in a README). Managing and running migrations smoothly without downtime is a very common challenge that a lot of startups face and I think a lot of us have been reinventing the wheel over and over to handle this due to the lack of current offerings.
Highly unrelated (and lots of reasons missing): I was never a fan of Gnome and for years ended up "just going with LXDE instead" until I gave Budgie a try about 6 months ago. I have never been happier in my life.
Check out https://dgraph.io/ . I haven't had the chance to use it yet, but I have been following its development for a while. Definitely worth looking into.
Even though Mutt, Alpine, Sup and Notmuch might work great for some people they seem to be accessible only from a terminal, emacs, etc and very text heavy. As much as I love spending time on the terminal, I prefer a GUI app for my email.
Postbox unfortunately doesn't support Linux, besides that while I don't mind paying licences for great software, it doesn't look like Postbox is fully opensource [1].
Thanks for not letting Thunderbird die. In my opinion it's still the best and most customizable opensource email client and there is just not a viable replacement.
Most of the open source email clients I have tested require you to run a local webserver and access the mail using a web browser with very limited features. All I want/need is a desktop app that can be customized to work similar to Gmail, pulls and deletes emails from remote SMTP/IMAP servers and allows me to create backups locally.
A realtime chat without auth but with private messages or multiple rooms. Basically the user will pick a nickname, click next and then be able to chat on a general or create/join a room.
Another approach would be to have to pay the equivalent of USD 5-10 cents in some email crypto currency in order to allow an unknown sender to put their email on my inbox. If I reply or add the sender to my whitelist the fee is automatically refunded to the original sender and future messages between both parties would won't require a transaction fee any longer.
The thing I'd be curious to separate out is ratio vs density. The fiction examples were positive, but a lot of the tokens are still spent on normal story work. The targeted examples put much more of the training signal on the AI being in the relevant situation and choosing against the bad option.
That makes me think the next thing to test is not just the positive/negative mix, but how much of the data is actually about the failure mode.