Nice thanks for clarifying!

very cool!

Not to rain on the parade, but this is the same size as the OpenDV dataset right? Is the novel part the data? or perhaps that it is in europe?

Ooo peak design is legit

for the record, this is why LLMs have been more widely successful and applicable than, say, vision-language-action models, and why VLAs are catching up: this is a recipe that can be applied very broadly, but only works at a production level if the data domain is VERY thoroughly covered

The more data you have, the better an embedding space you have, and the more likely your interpolation is to be correct. So you are right in that the something like the answer is probably in the training data, but you are wrong that the exact answer is in the training data or searched for.

Like many social media discussions, what is missing here is nuance. LLMs, like all generative no-prior ML models, are, effectively, interpolating. But in the case of LLMs, they are interpolating in the space of "next token embedding."

Fundamentally you can *have* both, but functionally when you optimize for multiple objectives usually only one ends up as the primary. Guzdials article is suggesting that the prior push being so attached to undergrad outcomes is a bad primary objective for K-12 students, which is reasonable...

I think a deeper difficulty in ML is the economy of attention. The hundreds of papers each day released on ArXiv in ML means that a reader needs to resort to heuristics to keep up. Stuff like trust a recommender system, or only read famous authors, or scan for buzzwords.

Example, pre-train (reward free) to map temporal distances into distances in latent space, and then, finetune: map these through a dot product with a latent task description to a reward function.

A couple of refs:

openreview.net/forum?id=YGh...
arxiv.org/abs/2110.02719
arxiv.org/abs/2110.15191

I know exactly what you mean. Especially for us academic-related folks, our recommended bubble gets ultra tight. my recommendation is to look at some of the "highly followed" topics, which will give a more norm-y feed. But truly BlueSky needs "Trending"

Depending on precision, that is a crazy price for 2 high quality 6-dof robot arms, to say nothing of them attached as one torso. If the price stays when people start building it you can be sure I'll be one. The Rethink Baxter is a lesson, cumulative error from backlash will be the important thing

agreed

Begs the question: at what point is multi-task training implicit meta learning @chelseafinn.bsky.social

super excited to try this out