Diptanu Choudhury

@andymac4182 @criccomini @JuicedataInc Would love your feedback on this! docs.tensorlake.ai/filesystems/in…

I have been doing a bunch of reading about SlateDB over the last few weeks. Last I checked it wasn’t an option for providing storage to agents and sandboxes. zerofs.net and @JuicedataInc are the two biggest ones I have ran into. Maybe there is more I haven’t seen.

After spending the last few months split between infrastructure and AI spaces, I have had my belief reinforced that investing in a good set of primitives really pays back dividends at the higher levels in your stack. There are so many startups starting to put effort in here. I do wonder if long term we will see more OSS projects like Kubernetes that end up coming out to help will this gap.

The timeline has really mellowed down. Not a lot of bangers from creators. What did they do? 😅

@zeeg I think of them as long horizon context for agents.

For someone who hasn't used ADRs and is considering it, I'd like to hear your opinions on it, tips, etc. Specifically with agents.

@parmita 100% this. So much of our announcements are just stuck around the logistics of the announcement.

one thing i have learned: public launches are perhaps the most unpredictable scary part of doing things in an early-stage startup lol thanks everyone for the support. we are posting the blog today and then on to the next task

@cyrusnewday 👋

if you’re seeing this we’re probably mutuals and you’re cracked

@andymac4182 You need to meet @criccomini and hang out on the slatedb discord :)

@diptanu That is fair. I want to see more investment in the OSS space for storage. We have Firecracker and rust-vmm which is great. I would love to see similar building blocks helping in the storage space.

@bnjorogedev Thanks! Will fix it :)

@diptanu great read. heads up: the tables dont seem to resize on mobile. Crops out the last column.

Tensorlake Sandboxes come with a highly optimized storage engine that delivers SSD-class IOPS: 66,100 random-read and 65,000 random-write IOPS at 4 KiB inside the guest, with 4.7 GiB/s sequential writes. Postgres runs within 1% of it on fsync-bound TPC-B. Sandboxes consistently snapshot in tens to hundreds of milliseconds, regardless of how much data the VM has on disk. We wrote a blog post on the design of our custom storage engine in Firecracker, with comprehensive benchmarks of running Postgres in the sandbox.

@erikdunteman Thanks Erik! The file system on Modal is really great now. I didn't post the raw IOPS that we got on Modal sandboxes, but it was solid. You guys are cooking :D

No it's totally fine, glad you tagged the post! Yeah I am telling people we are building compute infrastructure for agents. There is VM infra, file systems and networking at play here to get us code storage, sandboxes and versioned file systems to make it possible to run agents and code gen at scale.

@diptanu @ivanburazin @mlejva @anuraggoel @HeyGarrison I'm sorry if the tag appeared out of context, as it was a reply on my original thread stating "We should probably retire the word "sandbox". Distributed computers for agents (Disco)?"

We should probably retire the word "sandbox". Distributed computers for agents (Disco)?

@pavitrabhalla @ivanburazin @mlejva @anuraggoel @HeyGarrison What do you think Pavitra? :)

Curious what you all think @ivanburazin @mlejva @diptanu @anuraggoel @HeyGarrison

@damian_b Thanks brother!

@diptanu Great read. Especially the comparison against other providers.

Writing well researched blogs with benchmarks is a full time job :D

@arni0x9053 @PrimeIntellect Interesting! How long does it take to reinitialize the pods to serve new requests?

I'd say very accessible. Yume is open weights (5B), and one 80GB GPU is enough. I just had Claude use @primeintellect CLI to rent an H200 pod. Getting Yume + dependencies onto it takes 20-25 min (you're paying pod time during that but thats $2 or less depending on pod availability and price). Once it's up a ~2s segment of 720p world takes ~75s to generate (on H200), or ~3.5 min on an A100 (which it can also run on). I saw A100 spot as low as ~$0.63/hr so overall works out to roughly 1.5-5 cents per second of generated video depending on hardware. You don't keep the pod alive between sessions either. Weights and dependencies are on disk so when you want to generate again a fresh pod attaches the disk and it's ready to generate within minutes. And this kind of experiment (activation patching) works in the same kind of pod as inference cause it's just Python code executed along, no training. (And all of it Claude will set up for you just give it credentials)

Continuing my mech-interp experiments with the Yume world generation model. Trying out activation patching. I recorded the activations of a world generated from Monet’s poppy field (the full internal snapshot at one of the middle layers) and injected it into the same layer of a New York Manhattanhenge world mid-generation. A really cool effect, as if two memories are fighting over the same territory. This is where mechanistic interpretability meets art.

The advent of agents is forcing ML centric neo-clouds and app centric platforms to converge.

Software factories inside companies are changing the equation of what's expected from sandbox infrastructure. Coding agents building anything serious should get the same performance from sandbox's file system and networking layer as production environments to be able to benchmark what they are building.

@ahmetb I didnt know you bike bro!

🚴 doing the 1-day Seattle-to-Portland ride (207mi/330km, 5700ft climb) today. off we go!