awwaiid

Maybe we're getting into a world where our general models can make their own separate fine-tuned models as utilities just like they would a bash/python script. If making a fine-tune is cheap and fast then it can also be throw-away and constantly improved.

Sorry this is ten SNIPPETS or BLOCKS of code. I'm counting more than 10 lines here.

... er, but also cool and inspirational :)

As far as I can tell, the difference is that they made a new website? The cloudflared instant tunnels have been around for years, including json output as far as I know.

As far as I know this has been a provided feature of cloudflare for several years. I don't understand why it was posted today.

They have some allowance flags, you should email them.

If they can get impatience and hubris in there then they are doing great! https://thethreevirtues.com/

I love this sort of project and write-up. Inspiring to try something a bit different and dance on that line between practicality and artwork. The in-progress "scans" are themselves an interesting stretching of time and space, even if they aren't what the artist is ultimately going for.

I wonder how parallel token caches are, like when exploring a tree of sample continuations.

I misread that as "the first time I dreamt...".

Anyway yes, as I understand it this should be the last time you _remember_ having dreamt. https://en.wikipedia.org/wiki/Rapid_eye_movement_sleep says we hit REM multiple times per night normally.

... Also this Wikipedia says we have non-REM dreams. This whole article is interesting, thank you for indirectly prompting me to read up a bit :)

For the first part they have https://openrouter.ai/docs/guides/overview/auth/oauth (which I haven't used). That gives you an openrouter key for that user (openrouter bills the user).

What value are you providing that you'd take some further cut? You could take the money from the user via stripe and then pay your own openrouter bill instead.

Matrix, specifically synapse/rustMatrixSDK :)

I'm saying that Machine Learning was borne of AI from the very beginning. ML has become specialized enough that we may later (much much later) declare it somehow larger/separate/overlapping, but it originated from AI as the generic umbrella.

https://en.wikipedia.org/wiki/Machine_learning

Counter example, the book that is the foundation of much coursework and learning for people in AI, has a whole section on "Machine Learning" with all that k-means and such in there - https://aima.cs.berkeley.edu/

I'm long on semi-colons

Yes -- there are also links if you click into the lower-right help button to the excalidraw documentation, blog, etc.

Should probably indicate somewhere purposeful that it is built on Excalidraw

Yes, many companies going out of business altogether, some of them large, is what a bubble pop would look like. As opposed to a uh.... Correction.

Was there pressure to do this, or freedom to do this? If I had an unlimited token budget I'd probably try all sorts of crazy things. Also you (one) can read the tests and see that they weren't modified to forcibly pass.

Maybe he didn't think it would work. Maybe even if it does "work" they'll keep the zig version anyway. Maybe further study is needed beyond existing compiling/test-suite. Intentions and perspectives change over time, even only a few days, without dishonesty.

I'm guessing that if I said it ... that we have no intention of re-writing in rust ... that what I mean is "we have no intention of spending the extreme cost it would take to rewrite". When I discover the cost model is completely different that changes things.

Those 11 damn lucky interns!

Not AI generated software -- DYNAMICALLY generated software, like at run time and ongoing. Even in-app directed by the user. This is not a thing that existed before, a degree of customizability well beyond letting the user pick a color scheme or from one of a few layouts or default start screens.

I don't know how good of an idea it would be, product-wise, to give programming level flexibility. I am reminded of greasemonkey scripts, but written in english maybe. Maybe it could be awesome. But Apple is saying "nope. Not interested in exploring this with you. BYE"

power users opt in to opt in telemetry, and power users opt out of opt out telemetry. Power users click all the buttons.

It's also difficult to recognize that when it got it right THAT might have been the lucky week.

Yes -- this is getting very close to ClaudeClaw. Next they'll offer cloud hosting of persistent execution.

Maybe another idea, no idea if this is a thing, you could pick your block-of-layers size (say... 6) and then during training swap those around every now and then at random. Maybe that would force the common api between blocks, specializaton of the blocks, and then post training analyze what each block is doing (maybe by deleting it while running benchmarks).

This was not my experience. For example, it want until chromium 110 that you could use webrtc+pipewire without overriding settings, in 2023. So maybe in a strict sense you could do it with fiddling, I don't think you could install any Linux flavor and screen share over meet reliably.

https://groups.google.com/g/discuss-webrtc/c/fe567r-UUrA/m/8...

.... did you give them money? Brave!

Yes -- I know at least 3 Orion Blasters.

Where is this 1000 lines of C coming from? This is python.

Generative TikTok for words