Diptanu Choudhury

@rnjn Always been the case.

almost everyone told me that in early-stage startup sales, founders are the product. i bought into it. and i think its incomplete advice that can be quite dangerous if not understood well. it helped us get our 1st set of design partners. but last 12 months have taught me one more thing - the market does not care about your resume/past. "founders are the product" is a statement about founders' expertise, vision and passion. not the resume.

@bernhardsson 😂

In hindsight all the sandbox stuff is so obvious because containers without AI are basically just contners which makes no sense.

@pavitrabhalla @tensorlake Open Sourcing would take some time. Would love for you to try it out! docs.tensorlake.ai/sandboxes/intr…

@diptanu @tensorlake This is a really helpful improvement and something we have been looking into as well. @diptanu Can we help open source your implementation?

We’re working on Sandboxes @tensorlake. We built a high-performance file system to help agents run tool calls faster for coding and data analytics workloads. The goal was near-SSD speeds inside sandboxes. To get there, we forked Firecracker and built a custom, block-based overlay filesystem with dirty bitmap tracking for fast snapshots. Here’s a benchmark measuring raw SQLite performance across sandbox providers. Tensorlake is 1.2–1.3x faster than Vercel, 1.5–1.9x faster than Modal, 1.6–1.7x faster than E2B, and 1.8–2.2x faster than Daytona.

So much complexity from infrastructure goes away if you have sandboxes as primitives - stateful, dynamically sized, suspend, serverless boot. What is missing in the stack is sandbox native functions and applications. OCI Images, Kubernetes, elastic block stores, queues, workers were a drag to productivity. Agents will get better devtools to build than engineers got circa 2015-2024

Highly recommend if you’re in SF and building in AI. There’s no shortage of events here but 1) @diptanu is awesome and 2) the closed dinner format is my favorite.

@sandeepssrin You want Android emulators as a service? Interesting idea, hadn't thought of it, would love to chat and understand the use case some more! We built a full fledged VM infrastructure under the hood so should be possible in principle, have to figure out the product experience.

@diptanu Can you do android emulators using this ? On google cloud we have to set nested virtualization to make it work

github.com/adammiribyan/z… The part about using KVM to get clones of VMs is easy. The challenging part at the moment is cloning sandboxes 100x times across nodes. The bottleneck is around moving bytes across machines. We developed a hybrid approach of moving some data across machines directly, and some through blob stores. Also, requires tuning the network/RPC stack to move as much data the NIC allows you to.

@NathanFlurry @LoopholeLabs Will do! Writing something up as we speak :D

@diptanu v cool! @LoopholeLabs was doing smth sort of similar with criu (not kvm) streaming to blob storage: loopholelabs.io/blog/rethinkin… would be fun to read a writeup

@ellie_huxtable Would love to chat!

@diptanu this is p much a solved problem for containers, surely a similar approach works here too? p2p > blob store by a long way

We are hosting a small dinner in San Francisco at Spruce on the 24th of March for builders in AI Native software companies. If you working on harnesses for coding agents or any kind of back-office or background agents, would love for you to join us! We have a few spots left. Link - luma.com/t35j6dam

@gauthampai 100%

@diptanu We will still need higher level tools for developers and low level tools for agents, until we reach a point where we don't need human review of the code at all.

The era of building devtools for engineers is over.

I called this a couple of months ago! Based on this screenshot, Anthropic is barking at the wrong tree if they are using Kubernetes for building a PAAS for agents. Reconciliation loops like k8s controllers are not designed for high throughput stateful MicroVM scheduling. No one in our business has built something like this on K8s, it was not designed for this problem. Stay tuned to my feed, we have been working on the problem of internet scale agents in sandboxes.

Benchmark - github.com/tensorlakeai/s…

don’t sleep on @tensorlake this kind of optimization leads to a lot of new possibilities with massively distributed workloads at close to bare metal costs

@smehmood Because agents are using them. Agents write more code than engineers.

@diptanu And yet, all the devtools for engineers are exploding in usage?

@gukevinr Databases live on files too, just saying 😂

give your agents a db, not just a file system x.com/gukevinr/statu…

A coding workflow that is helping our team is getting Codex and Claude have a conversation about plans. 1. Write a plan using Claude 2. Get codex to review and criticize 3. Get claude to review the criticism .. and so on. It saves us a lot of time. A good example - whether to use virtio-pmem + DAX for sandbox I/O in @tensorlake. After a long session of back and forth, we decided to not go in that direction. We ran some benchmarks with FIO, and looks like after fixing some low hanging problems we are close to 90% of raw SSD performance.

@alcazarsec @criccomini @tensorlake @Gemini @criccomini We need to push this to a vercel app. We are releasing Serverless cron a week from now. We will put up a vercel app, that lets you get periodic updates of HN front page as podcasts!

@criccomini @diptanu @tensorlake @ElevenLabs @Gemini It seems they just published a skill.md. But this needs to be built. I'll ping you when we do it. hn.alcazarsec.com/daily

Turn Hackernews into a daily podcast automatically. We built an agent that: • Pulls top HN stories • Summarizes them • Generates audio • Publishes a podcast Runs on @Tensorlake, built with @ElevenLabs & @Gemini Code: github.com/tensorlakeai/e…