⿻ Andrew Trask

@Miles_Brundage "waymo after waymo after waymo" lol i don't know why it makes me laugh but it does

God forbid a robot have a hobby x.com/Phil_Lewis_/st…

ah... 3rd problem of ABC is a system of mechanisms for developing trust in data you don't acquire a copy of. the main thing it helps for for what you're describing is that the paradigm inherently involves users deciding at runtime which sources they trust (this is different than an AI company deciding at training time which sources are trustworthy) doesn't solve everything, but given how contextual trust is, being able to "dial it up" and "dial it down" as you change prompts is (imo) helpful to the issue

@iamtrask nice. although ABC looks like inference-time access control for publishers. how do you see this connecting to data quality legibility?

training data is starting to look like a zero knowledge proof problem. labs have to judge quality without seeing the full dataset or the QC pipeline behind it. vendors proxy quality with multi-rollout pass rates, small-model ablations, and downstream eval gains. but compute and iteration costs explode as environments and trajectories grow more complex. quality has no ceiling, and the best data is often the hardest to capture in a metric or explain in a writeup. huge alpha in making data quality more legible.

@elonmusk We worked with your team on this in 2022 and got most of the way there. Happy to get things across the line fast if you’re interested. blog.x.com/engineering/en…

It should be possible for a third party to generate the same results if they had the same data and we will need to be audited by credible outside companies for the public to believe us

To give people confidence that we are not secretly manipulating the 𝕏 recommendations, it is critical that we open source anything that influences what people are shown

@Miles_Brundage tell me... was this heroic story truly the work of solo journalism... or did you have a team?

@iamtrask lmao

@ChrisPainterYup (also happy to concede that "bug" is reductive. Flaw/limitation/constraint/etc. would have been more academically precise.)

@iamtrask What do you mean by “instability bug”?

@ChrisPainterYup Counting as well, e.g. iamtrask.github.io/2015/11/15/any…

@ChrisPainterYup As a useful contrasting example, ~10 years ago anyone could fine tune an LLM to say the word "hello" an infinite number of times and it would happily do that consistently for years because it requires no real context.

@ramez 100% with you there. they're ways to generate infinite amounts of clean data. RLHF is similar if your userbase is big enough :)

Data quality ("clean data" here) is underrated relative to data volume. Why do we have ASI in Go and Chess? One reason is that the training data itself is error free. (Unlimited in volume is great too.)

@AI_WarriorNQ probably an empirical question, but notably... one of them was removing bad data, the other one was getting clean data

If it's just 'get clean data,' why did 85M tokens of behavioral filtering barely work while 3M tokens of principled reasoning crushed it? The why in data > the what.

@theazaelov it does when we run out of good data

@iamtrask essentially all the same move: make the alignment data itself do the lifting curious if that plateaus hard

@0x_Afzal monkey see monkey do

@iamtrask teaching a toddler to behave by hiring a nicer toddler to grade their homework

@vanvster data quality + fancy name (bonus: pinch of anthropomorphism)

@iamtrask method changes but the bottleneck always comes back to data quality

@timhwang @FranklinMatija Nah. Doing both meant more clean data. More clean data meant better results.

*virtue alignment intensifying*

@leo_guinan subscribed!

I'm very early on. Tech is scary around here. So I'm coming in cautiously with an AI bodega cat. First business I talked to told me AI was going to end the world. And thank you. I take a ton of inspiration from Douglas Adams. He was more deeply correct that people appreciate. And I'm writing here. hitchhikertothefuture.substack.com/p/hey-clown-i-…

@leo_guinan your twitter name is awesome. love this

@leo_guinan is there a writeup of your project anywhere? you need to be blogging about that... even just about the effort...what is it like to be a local compute champion?

@emsi_kil3r Cheers sir.

@iamtrask You are probably smart enough to know that liner regression often breaks and turns out to be sigmoid in disguise.