Suraj Vathsa

Mostly interested in the intersection of how agents retrieve and utilize data at scale... but sometimes I like to let my brain wander. SmithDB @ langchain 🦜⛓️

[Suraj Vathsa @svathsasa]

@adityaag That’s not true. Take a look at the cotton gin invented by Eli Whitney. Whitney spent much of his life battling IP infringement but we definitely cannot say the cotton gin was not valuable during the Industrial Revolution.

[Suraj Vathsa @svathsasa]

@AnikaSomaia @SemiAnalysis_ engineering always wins over raw capital

[Suraj Vathsa @svathsasa]

@AdamRLucek I queue messages cause I’m bored

[Suraj Vathsa @svathsasa]

The FST is a beautiful piece of engineering: https://t.co/L2Zpph0AwP

Very relevant in today's token intensive world.

[Suraj Vathsa @svathsasa]

how will organizations trade-off between cutting token spend vs layoffs vs AI saas spend?

[Suraj Vathsa @svathsasa]

@bubbleboi high ownership is all that matters

[Suraj Vathsa @svathsasa]

@chamath “long term” could mean the same if the company continues significant investment in r&d or m&a

[Suraj Vathsa @svathsasa]

open weight + open source is the way

[Jacob Lee @Hacubu]

Own the frontier:

"In our evals, Nemotron 3 Ultra with a tuned LangChain Deep Agents harness provides advanced agent performance at a much lower inference cost. The main takeaway is that agent performance improves when the model, harness, evals, and runtime are tuned together."

[Jacob Lee @Hacubu]

Own the frontier:

"In our evals, Nemotron 3 Ultra with a tuned LangChain Deep Agents harness provides advanced agent performance at a much lower inference cost. The main takeaway is that agent performance improves when the model, harness, evals, and runtime are tuned together."

[Suraj Vathsa @svathsasa]

Reading Americana by @bhu_srinivasan had me thinking - Is llm inference mania like the canal mania of the 1820s where come the 1840s steam boats carrying goods on canals were made obsolete by steam engines on tracks? Are gpus the better business when compared to asics? If there is a paradigm shift will the asic makers be able to move fast enough to support the new paradigm or will nvidia prove cycle after cycle a data center filled with gpus offers the most resiliency to these paradigm shifts.

[Suraj Vathsa @svathsasa]

@amankhan @swyx @willpienaar @shrumm @talraviv @jxnlco @eugeneyan @sh_reya @petergyang @Vtrivedy10 nice t-shirt @Vtrivedy10

[Suraj Vathsa @svathsasa]

@grok can you find anything on this topic?

[Suraj Vathsa @svathsasa]

Is anyone working on a semantic DB - one that is optimized to scan a billion rows running each row through a small LLM?

[Suraj Vathsa @svathsasa]

Everyone's talking about the datacenter capex boom so here's a quick summary:

The foundation is fabs + their equipment (ASML, TSMC). They manufacture the physical silicon for almost everyone above. Then chip designers (Nvidia, Broadcom, AMD) and memory (SK Hynix, Micron, Samsung) sit on top and are customers of the fab. Nvidia designs a GPU; TSMC makes it; memory makers stack HBM beside it. Then optics/networking (Arista, Marvell, Coherent, Astera, Credo) wires those chips together into systems and racks. Energy/power (GE Vernova, Vertiv, Constellation, Vistra, Bloom) doesn't just feed into the chips, it feeds the building that houses them; it's a parallel input to the whole compute layer rather than a sequential one. All of that assembles into data centers, which are operated/owned by the demand layer: hyperscalers (Microsoft, Google, Amazon, Meta), neoclouds (CoreWeave, Nebius, Oracle), and frontier labs (OpenAI, Anthropic, SpaceX). SpaceX dependence is temporary, Elon is trying to vertically integrate ALL of this.

[Suraj Vathsa @svathsasa]

For those of us banging our heads against walls building agents:

[Sydney Runkle @sydneyrunkle]

https://t.co/WoyfH0rF7U

[Suraj Vathsa @svathsasa]

I agree, we cannot simulate our way to AGI. Observing, storing, and retrieving agent trajectories from the real world is critical. That's what we're building LangSmith to solve.

[Dwarkesh Patel @dwarkesh_sp]

Here's a question I find confusing and interesting and which actually tells us a lot about the nature of current AI progress:

Why has progress on computer use been so slow? Computer use is so clearly verifiable.

I think the answer is that it is not enough for a domain to be verifiable.

It also has to be very grindable—in the sense that you can run lots of parallel rollouts against a deterministic and replayable simulator.

If you’re trying to make a model better at coding, you can create an environment that has a software repo with some missing feature that you’ve tasked the AIs with creating, and then you have a thousand parallel agents just go at the problem, each with their identical copy of the container.

But this doesn’t work with computer use—at least not trivially. You can’t have a thousand agents go try the same checkout flow on Amazon. Because Andy Jassy will find and detect your bots and shut your ass down.

How would we train an AI to build a business? How would you make an AI that’s really good at winning court cases? Or having a profitable day trading in the markets? Or helping a candidate win an election?

What is the RL environment to make an AI as good at politics as Lyndon Johnson, or as good at building a space launch business as Elon Musk?

The rollout requires interacting with the world and cannot be recreated simply within the datacenter. And the outer loop verification may take months or years of real world actions to elicit, and cannot be re-observed by perturbing the model’s actions thousands of times in parallel so that you can isolate what exactly the model did that actually worked.

[Ankush Gola @ankush_gola11]

Supporting sub-second full-text search on object storage is hard, especially when dealing with large agent observability workloads. Here is part 2 of our blog post that outlines how we accomplished this in SmithDB!

[Suraj Vathsa @svathsasa]

More people need to be thinking about where bytes of information live, how they are laid out, and what it takes to retrieve them to convey that information to a human or an agent.

[Suraj Vathsa @svathsasa]

@Vtrivedy10 Now Ronaldo is gonna try and do the same thing and not pass the ball 😡