One of the biggest surprises in AI over the last few years has been how well coding agents generalize beyond software. In a recent essay, MIT professor Philip Isola argued that we may be entering the era of robot-use agents: general-purpose models that can control different robots, write policies, and learn new physical tasks with little or no robot-specific training. In this episode of Decoded, we're joined by the founders of @theWaddleLabs and @Robocurve, two of the startups whose work helped drive this realization. They're working at the frontier of using general-purpose models to control robots, and their recent demos helped inspire the growing conversation around robot-use agents. Together, we dig into the research behind that idea, from code-as-policies and vision-language-action models to the harnesses and evals needed to make these systems work in the real world. 00:00 — What Are Robot-Use Agents? 02:11 — From Vision-Language-Action Models to General LLMs 04:07 — The Bitter Lesson for Robotics 07:04 — How Coding Agents Learned to Control Robots 10:41 — How Robots Learn From Experience 14:22 — The Harness as a Form of Robot Intelligence 16:14 — Watching Astra Control a Robot 20:55 — Why General Models May Win in Robotics 26:05 — How Close Are General-Purpose Robots?

@theWaddleLabs and @robocurve ‘s works have been super interesting to follow during the S26 batch. They’re actually the main reason we added a new layer to our tech roadmap of a “robot operator model” that prompts higher frequency VLA/WAM models to execute tasks. Reasoning and in-context learning seem critical unless on-the-job/continual learning is solved in some other way

Diplomacy has come to Multi-Agent Arena! In late 2022 Meta FAIR released CICERO, which combined multiple models as one system to play Diplomacy well. Today, LLMs can do it all as one agent. So, see if you can compete against today's frontier agents in social strategy!

We talk about robot-use agents with @agupta and @FrancoisChauba1! Some thoughts: 1) robot-use agents should be able to control robots with either ode or VLAs or direct actions depending what's optimal. 2) more evals are needed to compare robot-use agents! I often find myself asking @chooi_jeq to benchmark new models.

We are aiming to make Superpower the best mid-size turbine on the market: - 42MW even at 110F ambient - Always waterless! - Deploys rapidly on four modular trailers - Competitive simple cycle efficiency 39% - Deploys in N+1 arrays for high availability scalable power

absolutely epic machine to bring to demo day and I can’t wait to be there at the launch site the first time it’s on starship heading to mars

Moritz is already the go-to law firm for so many companies I work with They pair top quality lawyers with AI to make everything faster and cheaper than seems reasonable

Wafer beat Cerebras on latency for @ycombinator's AI Office Hours. GLM-5.2 on Wafer averaged 379 ms versus 674 ms for Gemma 4 31B on Cerebras. that's 44% lower latency with a much larger model. YC wanted people to get startup advice from AI versions of its partners at conversational speed. after testing lightweight Gemma and OpenAI models, they moved to a dedicated Wafer endpoint. Wafer agents tuned the serving setup for YC’s request rate, cache usage, and prompt and response lengths. users spent 2.5 minutes longer talking to its AI partners on Wafer compared to other providers. read how YC built the experience and landed on Wafer 🧵 link in thread

Two YC companies led the way in igniting a revolution in how the robotics community sees robot control via LLMs. Tune in to hear more about the last five years of research that gave them the conviction that we were heading this way and some predictions on what comes next.

Introducing collaborators. Each is a mini AI that helps you: Figure out what to build. Find your first users + fans. Make your first $1 on the internet. If you're someone trying to start something new, I think you'll like this. Demo:

We built Codex/Claude ultra mode, but for web research! Uses heavy agent parallelization, code execution, DB tool, and some other goodies, all orchestrated with frontier models The squad kept iterating until it was clearly SOTA. Lemme know what you think 🫡

Congrats to @ayushswrites and the Warp team!

Given that I was part of YC W20 and now YC F26, I'm often asked about the differences between the batches. This either comes from actual curiosity or some hidden expectation for me to say it's "gone downhill". Let me answer the former and dispel the latter. There are huge differences, but probably not in the way you'd expect. There's a lot of public chatter about YC backing younger founders, inexperienced founders, and losing its touch. This isn't true. Remotely. There were incredibly talented founders in YC W20, and the same is true for this batch as well. Last batch, we had PostHog and WhatNot. This time, we'll have similar outcomes. The other day, someone tweeted "YC doesn't back visionary founders". But the point of doing YC isn't to advertise your vision; it's to quickly test your hypothesis in a structured environment. The biggest difference between my batch experiences is the location. Previously, YC was in Mountain View, and while YC recommended founders to live in Mountain View, many (if not most) lived in San Francisco. This meant a fair part of the YC experience was commuting down on the Caltrain, attending events, and then going your separate ways. Now, YC is in Dogpatch, and literally everyone lives in the same four apartment buildings. That makes a massive difference. You run into founders in the elevator. You run into founders when getting packages, You run into founders on the street. Everyone is building something, everything is talking about agents & harnesses, everyone is talking about the YC calendar. This sense of location builds community. I think YC's done a great job of not going full-send on campus. It's good to have independence. You choose where you live, what you eat, how you work, what you do. But the co-location in Dogpatch creates a lot of spontaneous encounters that lead to learnings and sometimes even new customers. Another difference is the schedule and partitions. One issue with YC W20 is it had a rigid schedule, and too many founders in one room. It's hard to build relationships if you feel the constant pressure to meet more and more people. The new structure keeps everyone in consistent groups so you build actual connections; it also has opportunities to meet everyone else without exhausting everyone's social battery. I can somewhat understand the YC haters because YC can look like a bunch of copycats from the outside. I remember a few batches after YC W20, I was looking at YC announcements from W23 or S24 and thinking the same. But then you return to YC and realize that's all a bunch of nonsense: social media distorts reality, and the same brilliant founders that YC backs have been in YC all along. All in all, I think @ycombinator's evolution has been positive; the move to Dogpatch was brilliant; and the orange remains really f'ing orange.

Jev, now open source: Lev A 4B open source System One model based on Qwen backbone The best performance for it's small size huggingface.co/interfaze-ai/…

If your legal team is still chasing context, manually triaging requests, and working weekends to keep up, this one's for you. Intake turns ad-hoc requests into tracked work in Legora, executes the work with the Legora Agent, and sends it back to the requester, with a lawyer accountable throughout. Walter Myer, Product Lead at Legora and former lawyer, walks through how it works. Available on October 7.

they signed $1.37B in LOIs during the YC summer batch! incredibly ambitious

Introducing Agent Ultra - a step change in deep research Agent Ultra orchestrates swarms of agents to perform exhaustive research, build comprehensive lists, and answer questions requiring thousands of sources. In both evals and vibes, it's state of the art

Most folks don’t know that the inspo for @numeral was my own experience trying to deal with sales tax in 40+ states. Fast forward to today, the puzzle has only gotten more complex. Ah yes, Matt, the gold coins in the Discord server.

Here’s to the wanting. Here’s to the dreaming. Here’s to everyone.