Tarun Chitra

ヽ(⌐■_■)ノ♪♬ @gauntlet_xyz/@robotventures/@aerafinance/@_choppingblock/@zeroknowledgefm

[Tarun Chitra @tarunchitra]

2000s: Conspicuous Consumption
2010s: Conscious Consumption
2020s: Conspicuous Consciousness?

[Tarun Chitra @tarunchitra]

@qquinnsworth yeah, working on computational bio stuff there [he was at D. E. Shaw Research; but he worked on GPU stuff with me during his internship]

[Yi Sun @theyisun]

AI is making cyberattacks far more dangerous, so at @axiom_xyz we formally verified @openvm_org to stay ahead of AI attackers.

Today we’re launching @Intrinsic_HQ to bring the same AI-powered provable security to the rest of software.

[b/acc, context platform engineer @AccBalanced]

@alexatallah per @tarunchitra, looking forward to kv cache capacity and utilization metrics on your provider leaderboard!

Also let me know if your team wants to use, to produce amazing KV cache agent use 3D visualizations like these?
https://t.co/uPzlvt25bs

[Tarun Chitra @tarunchitra]

@MurielMedard *seem to come to some consensus…

[Tarun Chitra @tarunchitra]

An interesting math question I've thought about the more I've tried to have multiple swarms run is what the "best graph topology for a problem" should be [this image is result my playing around with a few swarms and artistically interpreting it] — clearly the people managing these things have some "good intuition for this"

Compare:
- Humans always think in terms of tree-structure parent-child processes with ~fixed or regular lease times / lifetimes
- Agents have more 'loopy belief propagation' with stochastic lease times

The former gives you more 'regular' topologies, whereas the latter can be irregular and in some ways Cantor set-like (which is often how you get wrong answers if you get stuck in a recursive loop whose information gain per agent or depth level seems to decay)

Can you even formalize this? It sort of seems like the answer is yes, but it probably is something quite different than the standard methods for this (e.g. Lyons and Peres)

[Matt Huang @matthuang]

Swarm riders and Swarm-native companies

AI-use even among highly AI-native people is power-law distributed.

There seems to be a new class of people — call them “swarm riders” — who are capable of consistently commanding large swarms of agents to productive ends.

This is a new kind of managerial skill, analogous to the skill it takes to manage large organizations of people, but wholly different in nature and of course more technical.

Based on the swarm riders we know (including many at @Tempo and @Paradigm), these people are capable of productive token spend that is 1000x-10000x higher than median software engineers, and 100x higher than top 5% Claude Code users.

The art of swarm riding is different from software engineering and more akin to machine learning, where practitioners grapple with black box systems: defining reward functions and evals, transmuting non-verifiable problems into verifiable problems, running experiments, etc.

Swarm-native companies are a new class of company founded by and populated by swarm riders. These companies are organized around swarms in a way that conventional companies will struggle to replicate, as the cost of internal human coordination is sand in the gears of an effective swarm. Effective swarm-native companies will define the right reward functions and evals, remove functional and bureaucratic barriers, and then let the swarms cook.

We’ve gotten to see a few swarm-native companies up close — reshaping financial rails @Tempo, building a new gaming platform @Spawn, cyber swarms at Cathedral Systems — and it’s clear these teams wield a significant advantage over pre-AI teams.

Some questions I’ve been wondering about:

1) How fast will swarm riding diffuse through the economy? The implications for token spend could be significant.

2) Can swarms be effectively adopted by existing companies or will it generally track the formation of new swarm-native companies?

3) What becomes scarce as swarm capabilities themselves commoditize (through frontier models becoming better at swarming, swarm capabilities built into harnesses, etc)? Cultures and teams that can define reward functions well and otherwise get out of the way?

[Tarun Chitra @tarunchitra]

Right, the resolvent of the adjacency matrix gives you a way to unroll analytically (barring Cauchy integral singularities); yet with multiple agents unrolling and terminating simultaneously, you end up with this dynamic graph and each agent having a local view F_i that they unroll. The F_i = F + N_i where N_i is a random matrix whose element-wise stochasticity encodes the randomness in the MDP each agent is unrolling. Yet somehow they seem to some to consensus even though the spectral gap of N_i can be larger than that of F. That is sort of the mystery of swarms of agents

[Lei Yang Σ: @yangl1996]

Natural continuation as well as a new beginning.

[Yilong Li @yilongl_megaeth]

Building more AI capacity is only half the job.

An inference provider can have the right model, spare capacity and better prices. But users won’t touch it unless they trust it to serve the advertised model and protect their data.

Our whitepaper ↓

https://t.co/ECirl2vO55

[Tarun Chitra @tarunchitra]

@MurielMedard But yeah of course loopy BP seems to have unbounded run time like a limit cycle. Yet these agent swarms seem to be able to learn stopping times (via reasoning) and then traverse the resulting tree; it sort of says loopy BP is the wrong abstraction but maybe the right state space

[Tarun Chitra @tarunchitra]

@MurielMedard But what if the cycles that be unwrapped are randomly terminated? You can still solve a Bethe ansatz on the unrolled tree if you know the unrolling size and then your question is more about the stochastic process doing the unrolling (which the agent sort of represents)

[Alex Wang @heyyalexwang]

what objective is reasoning solving?

i've been trying to better understand some weird behaviors in RL/reasoning (especially latent reasoning), so i went back to basics to correct some misunderstandings i had

new post on reasoning as pathfinding in token space:

[Nous Research @NousResearch]

The Sign in with ChatGPT flow in Nous Portal now supports your ChatGPT Business plan

[Tarun Chitra @tarunchitra]

OH: “Sounds like the unlock schedule of a 2021 L1”

[Dan Primack @danprimack]

Anthropic math: Top market cap is basically $1 trillion, but the fully diluted is closer to $1.4 trillion.

[Tarun Chitra @tarunchitra]

@justinblau @AbdulElSayed Haha, I am mainly just pointing out the rankings are interesting more than any opinion on the people

[Tarun Chitra @tarunchitra]

This led me down a rabbithole — how does @AbdulElSayed's h-index compare historically to other congressional officials?

Apparently he's tied for number 2 overall

[Jostein Hauge @haugejostein]

This is nuts. Abdul El-Sayed’s h-index on Google Scholar is 37 — higher than that of your average tenured professor at an Ivy League school.

[Tarun Chitra @tarunchitra]

@justinblau @AbdulElSayed Yeah but most of this was done a long time ago for everyone in the top 5

[Tarun Chitra @tarunchitra]

The @FireworksAI_HQ hack shows that AI has crypto-style attacks — after compromising keys, the attacker started monetizing by selling inference on routers using compromised hardware

The real question is when it will be more profitable for North Korea to attack inference providers than crypto exchanges

[Kavli Institute for Theoretical Physics @KITP_UCSB]

Statement #3 of Yang-Mills Group, scientific collaboration shepherding AI result on existence of quantum chromodynamics & completions of the Standard Model, including subproblem proving the existence of the Quantum Yang-Mills theory. https://t.co/j8O8TUjSD3

[Tarun Chitra @tarunchitra]

@JettMonroe0x @AbdulElSayed not untrue although given the poor comprehension level of people in Congress, the bar is low

[Tarun Chitra @tarunchitra]

@sidpra5ad cc @ericxtang

[Tarun Chitra @tarunchitra]

Giving a talk about this and some of the advances in routing today!

https://t.co/Z2Kjlhjt1v

[Tarun Chitra @tarunchitra]

Every successful marketplace has winners and losers

But some participants make it more of a dark forest than a chessboard

We find some direct evidence for this — and some simple math to make these marketplaces more robust to rogue inference providers! https://t.co/Phgr7ktYdv