g413n

it's a pretty general policy but this is all super early, it's great at exploring websites so fuzzing was easy, for CAD it has good enough base rates with the few-shot prompt when we do the repetitive stuff, and we gave it checkpoints on each step, the other stuff in the mosaic are just some of our favorite clips from internal evals

relevant note is that we finetuned by having the human also use arrow keys which keeps it in-distribution but also slower to collect

yeah we actually had some wacky ideas with ctc + a reverse-causal mask but diffusion does just make it all a bit more simple

we do exponential binning but fwiw I think we can do way better just hasn't been the main research area initially

yeah we've done audio work in the past so we'll def merge the recipes at some point, long term should have full io that a human has (except maybe not generating video for video calls that seems a bit much)

so cool :)

what's the basis for conversion between hours of neural data to number of tokens? is that counting the paired text tokens?

yep this. we just turned off management

we would be so down to buy s3 junkyard tbh we were going around begging various storage clouds to offer us this before giving up and building it ourselves

yep this was the case for us.

I think <2x more drives than needed, not 20x (24 vs 14TB), but the racks holding the drives could've been denser. Around the same cost in any case and our colo doesn't charge for space, so it's not a big deal and we were just going with what we were familiar with, but something to try.

yeah we weren't sure about putting that number esp whether it includes all the image attachments, but in any case it's at least around the right reference class for the largest text data operations.

def agree on the setup fees, that was just a price crunch to get it done within the weekend. (too short-notice for professional services, too sensitive for craigslist, so basically just paying a bunch of folks we already knew and trusted)

for IPXE do you have any reference material you'd recommend? we had 3 people each with reasonably substantial server experience try for like 6 hours each and for whatever reason it turned out to be too difficult.

yeah that's why we started paying people near the second half- not super clearly stated in the blogpost, but the novelty definitely wore off with plenty of drives left to stack, so we switched strategies to get it done in time.

I think everyone who showed up for a couple hours as part of the party had a good time tho, and the engraved hard drives we were giving out weren't cheap :p

yeah the cost calculus is very different for gpus, it absolutely makes sense for us to be using cloud there. also hardly any datacenters can support the power density, esp in downtown sf

it's not just in sf it's across the street from our office

this has been incredibly nice for our first hardware project, if we ever expand substantially then we'd def care more about the colo costs.

yeah it's on the wishlist to try

our training stack doesn't make strong assumptions about data integrity, it's chill

When I was working at sfcompute prior to this we saw multiple datacenters literally catch on fire bc the industry was not experienced with the power density of h100s. Our training chips just aren't a standard package in the way JBODs are.

just general research work. Once the recipes are efficient enough the modality is a smaller detail.

On the product side we're trying to orient more towards 'productive work assistant' rather than the default pull of audio models towards being an 'ai friend'.

yeah this

it means that even after negotiating much better terms than baseline we run into the fact that cloud providers just have a higher cost basis for the more premium/general product.

yeah misunderstanding we'll update the post-- separately it's true that we aren't network specialists and the network wrangling was prob disproportionately hard for us/ shouldn't have taken so long.

around 250

egress costs are the crux for AWS and they didn't budge when we tried to negotiate that we them, it's just entirely unusable for AI training otherwise. I think the cloudflare private quote is pretty representative of the cheaper end of managed object-bucket storage.

obv as we took on this project the delta between our cluster and the next-best option got smaller, in part bc the ability to host it ourselves gives us negotiating leverage, but managed bucket products are fundamentally overspecced for simple pretraining dumps. glacier does a nice job fitting the needs of archival storage for a good cost, but there's nothing similar for ML needs atm.

yeah it's totally plausible that we go with something like this in the future. We have similar offers where we could separate out either the financing, the build-out, or both and just do the software.

(for Hetzner in particular it was a massive pain when we were trying to get CPU quotas with them for other data operations, and we prob don't want to have it in Europe, but it's been pretty easy to negotiate good quotes on similar deals locally now that we've shown we can do it ourselves)

worth stressing that we literally could not afford pretraining without this, approx our entire seed round would go into cloud storage costs

atm we don't and we're a bit unsure whether it's a free lunch wrt adding complexity. there's a really nice property of having isolated hard drives where you can take any individual one and `sudo mount` it and you have a nice chunk of training data, and that's something anyone can feel comfortable touching without any onboarding to some software stack

oh we use cloud gpus, infiniband h100s absolutely aren't something we want to self-host. not aws tho, they're crazy overpriced; mithril and sfcompute!

we also use cloudflare extensively for everything that isn't the core heap dataset, the convenience of buckets is totally worth it for most day-to-day usage.

the heap is really just the main pretraining corpus and nothing else.

we don't have perfect metrics here but this seems to match our experience; a lot of failures happened shortly after install before the bulk of the data download onto the heap, so actual data loss is lower than hardware failure rates

we're in a pretty unique situation in that very early on we fundamentally can't afford the hyperscaler clouds to cover operations, so we're forced to develop some expertise. turned out to be reasonably chill and we'll prob stick with it for the foreseeable future, but we have seen a little bit of the state-creep you mention so tbd.