agajews

Hey Hacker News! I'm Alex, one of the founders of Pantograph. A year ago, my cofounder Kelsey and I set out to try to make the most inexpensive robot that's capable enough to do interesting things, and sturdy enough to be able to run for many months continuously, through all of the bumps and scratches that it will inevitably go through.

We're very excited to launch Pandroid! It's small, inexpensive, and safe to run models on, without needing to worry that a rogue model will punch through a wall. The early robotics models that exist today aren't that smart yet, so we think it's important to start with a form factor where mistakes aren't a big deal.

A friend of mine asked once, "I wonder if the world today would be different if the people who invented tech saw things the way artists like Ghibli did." That has rung through my head from time to time, and it makes me want to make products that feel like "Ghibli tech"—spirited, alive, and human.

Not to boast too much about these robots, but they really feel like Ghibli tech to me! Especially when they're running autonomously, part of my brain just registers them as little creatures, like some kind of bird or chipmunk.

We'll be around in the comments today, would love to hear what you all think.

exciting! you could imagine brain transcriptions being very high signal RLHF data that could help align models more deeply to human preference.

Pretty different--obviously a lot less consistent today, but capable of a lot more diverse kinds of behavior. As the models get bigger, they'll be easier to prompt, and you could imagine a game designer writing a long prompt for different kinds of NPCs. (E.g. GPT-3 was pretty difficult to get to do a particular thing, although it was very general, but over the years the instruction tuning has gotten a lot better and now it's very easy to ask questions to ChatGPT)

Awesome! We should be a lot better than ordinary LLMs, especially at tasks that require making a lot of decisions in real-time.

Probably not the current model, but one of the benefits of doing internet-scale pretraining is that the model has seen a lot of mods already! We think the bigger models will be able to handle some more custom Minecraft worlds.

The goal conditioning is just a training objective! You could combine it with tool use, web searches, etc. and still train end-to-end on goal conditioning. (One way of thinking about it is goal conditioning defines a good distribution of tasks, and lends itself well to pretraining with an unlabelled video dataset.)

you don't need to pick good images manually! it's similar to next-token prediction, many prediction tasks aren't especially interesting, but there are enough hard ones that the model can spend the most time learning from those. the simplest thing scaled up works very well.

Not yet, but we'll add language as a modality to the larger models! The models are trained end-to-end on video data, so we'll need datasets that mix video and language, e.g. transcripts of game streams. When the models are scaled up to cross-videogames and robots there will definitely be a bunch of language data.

Very much inspired by those papers! One of the things that's interesting about our model is it's goal-conditioned, so it can do any task at inference time without training on it. We had a lot of fun making eval environments after we trained the model trying to find interesting things it can do, and that was all after we trained the model. More like prompting an LLM.

(Versus Dreamer, which needs to be trained on a hand-written reward function for each task that you want to do.)

Hey everyone! I'm Alex, one of the founders of Pantograph. We've spent the last six months building a pretty smart Minecraft model, coming soon to a server near you!

We trained it on about 500k hours of Minecraft videos, and it learned how to fight creepers, build walls and other structures, and explore to find visual goals.

We're considering putting up a public API for larger models like this one, let us know if you'd like to be able to put Pan in your own server :)

What's most interesting about the model isn't the performance that it gets in Minecraft, but how general the method is. When we scale it up, it should be able to act in any kind of video game, as well as robots in the real world (which are really just another video game).

Ah the hardware isn’t gonna be in SF (not the cheapest datacenter space)

But I do think a lot of our customers will be out here —- SF is still probably the best place to do startups. We just have so many more people doing hard technical stuff here. Literally every single place I’ve lived in SF there’s been another startup living upstairs or downstairs

Good idea to host some in person events!

Yeah we aren’t going to let anyone book the whole thing for years. If we ever have to make the choice, we’ll choose the startups over the big companies.

Ah that’s only if you pay for 3 years of compute upfront. Most startups, especially the small ones, really can’t afford that

Yeah it’s pretty hard to find a big block of GPUs that you can use for a short time, esp if you need infiniband for multinode training. Lambda I think needs a min reservation of 6-12 months if you want IB.

Yeah exactly. That's why you can't really do it with a language model like GPT-3, you have to bake into the architecture the concept of a "link" as a first-class object.

Hey, thanks for posting!

We actually have an architecture that lets us expand the index without doing any retraining, so we can add/update pages pretty much for free.

Hey everyone! Metaphor team here.

We launched Metaphor earlier this morning! It's a search engine based on the same sorts of generative modeling ideas behind Stable Diffusion, GPT-3, etc. It's trained to predict the next link (similar to how GPT-3 predicts the next word).

After GPT-3 came out we started thinking about how pretraining (for large language models) and indexing (for search engines) feel pretty similar. In both you have some code that's looking at all the text on the internet and trying to compress it into a better representation. GPT-3 itself isn't a search engine, but it got us thinking, what would it look like to have something GPT-3-shaped, but able to search the web?

This new self-supervised objective, next link prediction, is what we came up with. (It's got to be self-supervised so that you have basically infinite training data – that's what makes generative models so good.) Then it took us about 8 months of iterating on model architectures to get something that works well.

And now you all can play with it! Very excited to see what sorts of interesting prompts you can come up with.