Tireless AI agents will be able to go back in history and audit everyone's token output to see who has been the most generally accurate person. The days of charlatans hiding behind sheer volume of falsehoods that would take too long for any reasonable person to fact-check is over.
22
Not well understood: you won't beat frontier labs at agents by writing better harnesses around their models. They're training the model and harness to work together end-to-end, and eventually the model will absorb the harness.
1
1
38
can we train an LLM on only writing done by 8yr olds and put to bed this debate on whether there will be an uncontrollable super-intelligence lift-off?
let an army of 8yr olds RL whatever they want onto it, i don't care. whatever came of those pre-1920 language models?
2
86
how are people still falling for terrible X poasts when you can just click the grok button and ask the bot anything you want about the topic or the person who poasted? the latter i've found is especially useful for understanding the poaster's personal biases before accepting their arguments
1
57
Gut instinct: LLM performance would drastically increase if they're given the ability to delete their own context.
Today, models believe too strongly in their "first instinct", and are then polluted by in-context reasoning tokens.
If reasoning is allowed to "delete" portions of context, i.e. "hmm looks like i was wrong about this, let's just delete this instead of letting it attend every future token" i think would unlock massive reasoning leap.
1
124
Seems pretty clear that Anthropic models use some kind of formal logic scaffolding that causes all the weird language outputs "load-bearing, substantive, one correction that matters, reframe, etc"
100
David Yang retweeted
hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0)
David Yang retweeted
Just read in The Atlantic according to a study that "using AI did not significantly improve patients’ ability to diagnose themselves or others."
Went to go look up the models used in the study.
Hmmm. 🤔
Basically what’s happening is, tech wizards were the meta since the transistor patch but the new update completely revamped summoners allowing them to summon every other class! So now many tech wizards are dual-spec’ing into summoner while druids complain about what is art.
2
361
Grok can determine whether people are lying on X by looking at all their historical X posts and checking for discrepancies. In my use folks rarely outright lie but often heavily exaggerate the truth...
1
290
Posts on X used to sound dumb. Now every post that would have sounded dumb before just sounds like ChatGPT. I can't be the only one experiencing this...
342
At some point AI agents will be able to automatically detect everyone's believability by looking at their historical posts and comments. When this happens many folks will delete their X accounts, start new ones, and think before they post.
154
Reading a lot of bad takes on Alpamayo, so here's mine:
If we still believe we live in a world with AI capability is correlated with model size and data size:
Nvidia Alpamayo: 1M miles
Waymo: 100M miles
Tesla FSD: 7B miles
Until there is a breakthrough in simulated randomness,
1
293
The name of the game is, "how do you get the data?"
Auto OEMs could certainly pay for Nvidia chips, cameras, line retooling to have this model in their cars today, but to improve it, they have to collect billions of miles more data. That will take years minimum if they start now
1
182






