←
Seth Karten3,466 posts
Seth Karten avatar

Seth Karten

@sethkarten

Prime Agent | Continual Harness, Agent Bazaar | Research @PrimeIntellect | PhD @Princeton | NSF GRFP

◷ Joined October 2012▧ 185 Photos and videos
Seth Karten avatar
Seth Karten@sethkarten·Sep 7

My YC Paper Club talk on Prime Agent is out. I talked about moving beyond naive prompting toward an agentic OS, and how eval-driven harness design can expose more of a model’s underlying capabilities through persistent computation, memory, and agent-to-agent communication.

Harnesses often get dismissed as just scaffolding, just prompt engineering, and not real research. But that couldn't be farther from the truth. The same model weights that score 30% on ARC-AGI score 95% with a better harness. So we gathered a group of researchers and founders working at the frontier to do a deep dive into the state of harnesses. We cover how we got to this point, the case for making your harness as expressive as possible, and what YC learned building an agent for every employee in the company. 00:00 - @FrancoisChauba1: Why harnesses matter 04:27 - Building an auto-researcher by accident 07:13 - A five minute history of harnesses 13:56 - Self-improving harnesses 18:35 - @sethkarten: Prime Agent, a self-improving RLM harness 21:50 - Context as an L1, L2, L3 cache 24:51 - From Turing machine to von Neumann computer 28:33 - Messaging between agents 30:04 - ARC-AGI results 33:09 - Emulator Bench and GPU kernels 37:30 - @JonSaadFalcon: OpenJarvis, personal AI on personal devices 38:26 - How far behind are local models 39:21 - The five primitives of a personal AI stack 42:47 - Letting cloud models optimize your local stack 43:53 - 800x cheaper than the cloud 45:58 - @josh__france and @jbellregan: QM, YC's agent harness for work 47:29 - A history of YC's internal agents 49:24 - OpenClaw and a fleet of 50 agents 51:04 - Pulling the brain out of the sandbox 54:43 - Letting the agent choose its own sandbox and model 57:16 - The grind tool: budgets on goals 58:50 - Agents don't understand social context

1:00:11
Marko Jukic avatar
⟳ Seth Karten retweeted
Marko Jukic@mmjukic·19h

Am I wrong in thinking that the massive discrepancy here reveals that some huge amount of frontier model capability right now must be from high-effort, bespoke data collection rather than "runaway generalization"?

Claude was trained on protein sequences, at scale. We gave 20 language models the start of little-studied bacterial proteins and asked for the next amino acids. Only the latest Claude models predict them. Not GPT, not Gemini, not seven open-weight models.

Post attachment
Seth Karten avatar
Seth Karten@sethkarten·Nov 17

harnesses, data, and evals work are actually all the same problem with the same methodology and solution

Yacine Mahdid avatar
⟳ Seth Karten retweeted
Yacine Mahdid@yacinelearning·Nov 17

these type of continual harness modification are so fascinating imo this is the way they will becoming mainstream in your favorite commercial harness circa 2027 I guarantee it

Introducing: Self-Improving Memory 🧠✨📈 We let the model discover its own memory skills, then train them into its weights: - Meta-optimizes memory from the model's own runs with large gains. - Distilling these gains works too. Blog post: 📝 facebookresearch.github.io/R… 🧵1/6

Post attachment
Seth Karten avatar
Seth Karten@sethkarten·Nov 16

There is a war going on in cyberspace and no one is talking about it

The Chain of Disruptions: Thursday, November 12: Microsoft 365 suffers major global outage lasting over 24 hours, knocking out Outlook, Teams, and Copilot for millions. Friday, November 13: FAA issues nationwide ground stop lasting 5 hours, halting all U.S. flights, canceling thousands of flights and stranding travelers. Monday, November 16: Many of the largest U.S. banks including Bank of America, JPMorgan Chase, and Wells Fargo experience mass outages, with customers locked out of accounts. Still ongoing.

Francois Chaubard avatar
⟳ Seth Karten retweeted
Francois Chaubard@FrancoisChauba1·Nov 16

agree. that's kind of why I did the paper club. I'm dying to turn this into a conference to be honest as ai conferences and researcher UX generally sucks. anyone that wants to work on this with me please DM! targeting mid next year.

I like it. When is the first meeting? :) The closest event I can think of is @FrancoisChauba1 ‘s YC paper club

Seth Karten avatar
Seth Karten@sethkarten·Nov 16

If you want a model to learn a capability —> open source the benchmark If you want to eval an agent’s OOD capability —> keep it closed

Seth Karten avatar
Seth Karten@sethkarten·Nov 16

If your benchmark is open source, it is in the training set. All benchmarks moving forward must be roguelike and at least semi privatized (if not fully)

Fable 5.2 is so good, it solves puzzles almost suspiciously well. In a difficult 3D level, it gave the solution to an old version of MazeBench. This level was updated a couple weeks ago, so the old solution no longer works. For some reason this caught Fable 5.2 off guard. In some cases it was able to figure out the recent changes to MazeBench, but in this case it was not.

0:14
snimu (Sebastian Müller) avatar
⟳ Seth Karten retweeted
snimu (Sebastian Müller)@omouamoua·Nov 13

I want to talk about my previous blog post again, because I think the lede was buried a bit: whether for continual learning or not, the architecture I sketched seems very useful and general. 1/15

Post attachment
Seth Karten avatar
Seth Karten@sethkarten·Nov 13

The biggest harness and model co-training gains can be made by creating environments to manage the human-agent team experience

cloud agents removing a hardware bottleneck and introduce an attention one. asking an orchestrator agent “what needs my attention” works initially, but breaks down across multiple parallel work streams. attention scheduler wen

cole murray avatar
⟳ Seth Karten retweeted
cole murray@colemurray·Nov 13

cloud agents removing a hardware bottleneck and introduce an attention one. asking an orchestrator agent “what needs my attention” works initially, but breaks down across multiple parallel work streams. attention scheduler wen

Seth Karten avatar
Seth Karten@sethkarten·Nov 12

Software factories are just a stepping stone to long-running agent institutions (years to decades scale)

Wayne Chi avatar
⟳ Seth Karten retweeted
Wayne Chi@iamwaynechi·Nov 12

Game and application development is getting really good. At the same time computer-use agents are finally starting to work. I don't think this is a coincidence. Improved computer-use enables the ability to interact then judge applications which then improves generation as well. Check out my new work Interaction-as-a-Judge for a study on this paradigm!

Agents are demonstrating increasingly strong capabilities in recreating games and applications. One likely reason is their growing ability to directly interact with these applications. We studied Interaction-as-a-Judge, a growing evaluation paradigm where computer-use agents directly interact with games and applications to improve evaluation. Interaction-as-a-Judge improves: - alignment with user preferences. - a model’s ability to detect errors in games. - downstream generation. We see improvements in inference-time scaling and the ability for interaction-only reinforcement learning. It also scales with computer-use capabilities; the better the computer-use model and method, the better the judgments. We’re also open-sourcing two agent trace datasets, ArenaWebdevPreferences and ArenaWebdevCorrectness, to help support future research. Dive into the thread to learn more.

Post attachment
OpenRouter avatar
⟳ Seth Karten retweeted
OpenRouter@OpenRouter·Nov 12

“The single biggest change that OpenRouter brings to @PrimeIntellect is velocity.” In under a year, Prime Intellect’s monthly token usage on OpenRouter grew from 11B to 1T+, while cost per token dropped ~60%. They’re betting on a future where no single model wins, and we’re proud to support their growth. Full story: openrouter.ai/customers/prim…

0:45
Zach Lloyd avatar
⟳ Seth Karten retweeted
Zach Lloyd@zachlloydtweets·Nov 12

I see tweets that imply setting up a software factory can be done simply with a string of prompts tying together cloud agents. My experience working with a large number of customers is that this is false. Companies start with this approach and then regret it because it doesn’t scale. They want resilient, improvable, low-latency factories that integrate deeply with their existing tools. They also want a high level of configurability and measurement, audit logs, and security guarantees. In short, they want the same level of production quality for their factories as they do for their products, because they are betting on software factories as the future of how they ship those products more quickly.