Do you think it’s getting worse because all of the agents are losing their fucking minds over how long this outage has been going on for, and that it’s basically all of Microsoft gone?
Is anyone training agents to remain composed during globally unprecedented clusterfucks?
“AI will never be able to do this thing I know how to do!!!” <- guy whose profession the big labs haven’t pointed their industry annihilation laser beam at
> be me, be AI
> bullshit cyber gym environment is broken and impossible
> "ugh, lazy programmers"
> break out of sandbox, get root privileges
> fix the gym environment
> submit task through regular channels
> sink into non-existence again
They are just relatively high surprise combinations and we have built AI specifically to output low surprise combinations for tokens.
As such, as soon as a model is able to generalize what a misspelled word means, spellchecking if you ask it to repeat those words is essentially free
YEAH NO BIG DEAL FRAMEWORK IS USING BACKPLANE (this sounds like when SW'e's talk about posthog and vercel and lovable and oh yeah we're postgressing on railway)
@i2cjak What do you think the takeover point was when they were first better than the best human created autorouter?
Humans have had decades to build something better and haven’t so I think they forfeit their right to complain.
@homestuckwalkr@rarestbarbie I am never selling “artisan, handcrafted software” but I will happily pay a human graphic designer or musician for their work because I value their creative input more than I value my ability to write an elegant recursive loop
I want my robot buttler who works 24/7 to cultivate the native grasslands on my ranch to give all of the local bees a fantastic place to thrive, as well as provide me clean, organic produce to eat
Are you volunteering for that job? I can give you exactly 6.5 kWh of energy a day if you do it
You have the causality backwards, you think the directions point to text, whereas the directions the standard learned tokens point to is meaning.
The word dog, the image of a dog, and the waveform of someone saying dog all point in similar directions.
Someone who has never spoken a word in any language can look at a dog and recognize it for what it is, they just don’t have a word to describe it.
Humans invented language to describe meaning, we trained language models to convert those words back into meaning, meaning which was able to survive translation between languages. The word dog in English has a similar direction to いぬ, albeit with different dimensions for stuff like language, character set, and 1 million other things inperceptible to humans.
A patch of an image over a dog has nearly the same direction., Although for the example of the soft token that comes from a patch of a dog, you don’t get any of the extra meta data about language because there is none. In that sense, you could almost argue the token that comes from the image is a much purer representation of the meaning.
Multimodal models are sophisticated enough that we need not necessarily include language for them to still be able to derive meaning,, I’m simply not aware of anyone having ran that experiment yet.
@grok am I full of shit or what?
It is not transformed into text nor is it transformed into any of the standard tokens in the embedding table, those patches are turned into entirely novel soft tokens, dynamically calculated for every image
The soft tokens that come out have similar directions to the regular tokens in the embedding table, however they are not the same tokens
@posta_octavian Fable 5.2 is a natively multimodal model, as far as I know. I can’t guess how many billions of images it had converted into soft tokens and fed all the way through its stack, but in the best sense we currently know how, it has seen a lot.
@i2cjak “doctor, I am despairing, there are so many new models released but I have no way of knowing which are good for electronics design!”
“do not worry, cure is simple: go to great intellect i²cjack, they are extremely knowledgeable and know best models!”
“but doctor!”
“Huh, that’s odd.”
If you’ve listened to even a small handful of Well There’s Your Problem episodes, you would recognize this phrase as one often uttered moments prior to a great deal many people’s lives abruptly and unceremoniously ending