@MParakhin Just ask astral to enable this feature in config.toml “features.context_management.experimental_mode”. You should also be able to ask it to set the context window at 1m, but that burns tokens very fast!
@gdb@dkundel ChatGPT Work works in the cloud meaning I don’t have to carry around a half open laptop, it also means I’m not constrained by CPU/ram when working across lots of different projects. This has opened up a lot more parallelisation (also the resets).
@OpenAIDevs@emollick this seems like a move towards what you have been discussing (and I very much agree with) re interfaces that optimise for reviewing and interacting with the underlying work.
@emollick Many enterprises use Copilot or other services with a model router, complex prompts can get much better results because they trigger reasoning models. I think this somewhat adds to the confusion.
@willdepue PDF export / print for Deep Research queries. Given the length of the results, I often use o3-mini to convert the reports to Latex then export in overleaf. It would be great if this was built in.
Yes, as in, LLMs would be doing all the config and filesystem management for you, fully replacing stuff like Finder / Spotlight / Systems Settings at the user level.
Want to flip the mouse wheel? Just ask it. No need to launch regedit and navigate 17 archaic menus. Want to cook something sweet? Just ask it for all sugary recipes you've written. And so on. Never manually manage a file again: just let the LLM be the middleman between you and the OS. That's what I believe will be the ultimate human-computer interface.
Of course, for it to be as efficient as possible (like any other usage pattern), you will definitely want all layers to be *aware* of this usage. And, yes, this includes hardware-specific instructions (like on-chip inference), dedicated kernel syscalls, and so on. That's what I meant.
Now, of course, I totally get how "100% around LLMs" could have been interpreted as "LLMs are scheduling threads" so I just wanted to apologize and be clear that this would be an absolutely terrible idea and that no sane mind would ever consider doing something like that
... I hope
This is a demo of a custom component called `gradio_leaderboard`, which was designed and developed by @freddy_alfonso_
What sets the latest version of the Leaderboard apart is the use of the "𝐫𝐚𝐧𝐠𝐞 𝐬𝐥𝐢𝐝𝐞𝐫" component, especially useful for Leaderboard implementations.
I share @GeorgeMonbiot's concern there's a risk the West may be sleepwalking into authoritarianism, with the UK being particularly concerning. Though, I'm intrigued he doesn't appreciate that some absurd 'pandemic theatre' measures he supported didn't play a role in this.
1/
@AnnabelGSteele Spaced repetition flash cards on Anki, I have really bad memory especially for names but I used it to memorize roughly 800 case names with surprising ease. This is a great place to start https://t.co/IMAm25jv8V