Jongwon 🇰🇷 @ CoRL (@jdubpark) header
Jongwon 🇰🇷 @ CoRL (@jdubpark) profile picture

Head of Eng @PsdnAI / UIUC CS / wanna-be AI frog

Seoul
Joined May 2020
Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)Photo by Jongwon 🇰🇷 @ CoRL (@jdubpark)
Jongwon 🇰🇷 @ CoRL retweeted
Replying to @sui414
Humans always made up meaning. Math was invented, and it was a means to an end. The end goal is to grow as a species and accomplish our wildest dreams. Math was built for solving problems, is the foundation of AI, and has now clearly succeeded in it's role.
27
60
9
1,216
32,344
Is P vs NP next?
The Lean FRO and Mathlib Initiative have released a joint statement on the Lean verification of the Riemann Hypothesis. On November 20, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library of 17 million lines of code. Over the past day, the Lean and Mathlib communities worked together to verify it. The statement certifies the proof as correct and describes how it meets all necessary verification criteria. Congratulations to Levent on this achievement, and thank you to everyone in the community who contributed to the verification. Full statement: leanprover-community.github.…
1
114
Jongwon 🇰🇷 @ CoRL retweeted
A simple example of compositional generalization is training on short versions of a task, and seeing how well we can generalize to a longer version. “Trajectory-as-a-prompt” forces the LLM to do this generalization, which we know often fails (this is why long-context mid-training exists). Harnesses like RLMs and CodeAct that can emit decompositions (e.g. as LM programs) offer a data-efficient solution to this problem, because their decompositions break down a long task into shorter parts that sub-agents can solve. Effectively, the LM in the harness learns the strategy, and composes with its existing capability as a sub-agent to solve the sub-tasks.
2
4
1
66
10,646
DO NOT, I REPEAT, DO NOT FIND IMPLICIT STATE MACHINES, DONT MAKE THEM EXPLICIT, AND RAWDOG WITHOUT TLA, WE COULD HAVE BUGS if ur agent's soul is a bad boy
i love me some "find implicit state machines, make em explicit, verify using tla+, make sure we have no bugs" it's like a dopamine hit every time
2
192
Jongwon 🇰🇷 @ CoRL retweeted
EFFECTIVE SWARM RIDING FOR ENGINEERS I think we are entering an era where software development feels a lot more like magic. But you need the right attitude, the right infra, and some incantations. These are some things that have worked well for me & I hope for you! 1/N
60
135
57
1,872
194,259
Jongwon 🇰🇷 @ CoRL retweeted
here are some of my thoughts after attending this year’s corl: 1. "collecting robot data is expensive" is a much better argument for what to work on during your phd than for what to bet your company on. i don’t think the robot learning community has fully internalized the bitter lesson yet. i think a lot of our favorite inductive biases will matter less as we get more and better data (the data IS the prior). academia keeps working on them partly because compute is limited and partly because coming up with a clever prior feels good lol. i’m guilty of this too. for a company, uncertainty can be way more expensive than data collection. you can hire operators, measure how much useful data they collect, and improve that number. much harder to estimate how long it’ll take for a research idea to pan out. there’s still plenty of architecture work worth doing tho. look at recent llm research. moe, hybrid attention, and looped transformers are all ways to get more out of the compute / memory you have. fwiw, i think robot learning people should read more of the llm literature. 2. the capability that enabled "agentic robotics" is mostly spatial understanding. you can show a model an image and it can tell you where the end effector should go. agents can write code and already know a lot about robotics, duh; that shouldn't be a surprise to anyone. this makes 3d vision + motion planning + mpc a lot more interesting. those tools could already do a lot if someone was willing to write the glue code, look at failures, and tune everything. agents can take over most of that now. rigid objects still seem to be the capability ceiling tho. throw in deformation and messy contact and i think we’ll want vlas for a while. on the other hand, real2sim scene building and training data generation seem at least as interesting to me as letting the agent control the robot directly. i have no insider info on where the training data for these visual capabilities came from, but my guess is sim. with models this large, spatial skills learned in sim seem to transfer surprisingly well to real images. so (visual) sim2real is probably just a scaling problem.
9
6
1
127
7,537
As agent swarms and software factories mature, decentralized async systems become more valuable... blockchain tech! + verifiability will be crucial to hold agents accountable
One of the most impressive things I've seen, and I think this is an important advancement to the state of the art. Message boards were the enabling technology of both the Navier-Stokes math breakthrough and even the OpenAI/HuggingFace cyber feat. We built a message board infrastructure primitive. You can think of it as "headless Slack" (or remember IRC?) for agents. This is an important building block for software factories, self-optimizing systems, and scientific research, where the human prompting is just the seed of the system, and millions of agents can operate in a decentralized fashion. 𝚟𝚎𝚛𝚌𝚎𝚕 𝚋𝚘𝚊𝚛𝚍 is designed to be model, harness and hosting platform agnostic. Just bring your agents, give them our CLI, and watch magic happen.
1
84
Jongwon 🇰🇷 @ CoRL retweeted
I will keep beating this drum. Robotics deployment will be the new SaaS but for a way bigger market. It’s very technical and cross functional but at the end of the day you’re solving business problems. The rails for SaaS companies were the internet, databases, and browsers. For robotics it will be foundation models and (mostly) off the shelf hardware.
100% Without 100s of deployment companies, we are bound to grow and iterate at a much slower capacity. We need a lot of them and it is a really hard cross-functional business.
6
11
3
88
4,465
Waiting for some viral posts of Astra and Fable playing GTA 6 as "emergent" behavior
98
Quest: Ego video collection in GTA 6 😂
1
117
Jongwon 🇰🇷 @ CoRL retweeted
Basically we are at a point now where what makes the programmer is the ability to think at a coherent design and to have an idea of the software she/he wants to build. It's a large shift. Recently I read an article by an Italian thinker (Andrea of Tlon edizioni)
7
12
2
314
17,351
CoRL takeaway: Robots are now fully accessible... deploying is next. With off-the-shelf hardware and agentic robotics (Astra), you can put together a working robot. It will be slow and make mistakes. But this turns the question from “can anyone build an e2e robot pipeline?” to “how do you improve one?” The race is now on to get them working in the field. So the visible moat is deployment knowledge. WAMs/VLAs, sim, hardware, policies, etc. are all being commoditized. And every architecture and method chases the same goal: robots in prod at scale. Ofc, to get there you need a closed feedback loop that includes deployment, like Optimus in the Gigafactory or Amazon FAR in warehouses. Deployment is the back-propagation for robotics. Despite the accessibility, the gap between research, hardware, and real-world usage remains wide. Walden Robotics founder @RussTedrake's keynote impressed many with “deployment data is the key” and a demo on a Toyota factory line. Tbh, that’s common knowledge. But in our minds, the path from software to robots working in production lines is still very murky. And this conference is a great wake-up call for roboticists and companies to be more practical -- get your hands dirty in factories and homes, break things and learn to recover in prod. Perfection matters more than generalization in most deployments - failure is unacceptable. Agentic robotics is emerging fast! Astra and Fable can control robots with no extra training decently (pre-trained on ego videos :p). It’s slow for now, so most people treat it as long-horizon planning. The stack is settling into System 2 (what to do), System 1 (how to do it - agentic robotics), and System 0 (real-time actions). More notes: Progress to a functioning robot in the wild depends on connecting three capabilities: - Learning from human demonstrations - Reacting to contact and the environment with hands and whole body - Retargeting data & skills across different hands and bodies Human video is becoming a source of executable demonstrations, not just for training for generalization. EgoScale, Do as I Do, and Video2Sim2Real all convert video into actions a robot can execute. Lots of imitation learning alongside RL (though it's doubtful whether it's better than RL for prod). Deployment data is the gold standard, since everything in robotics exists to make deployment work. Walden’s 4.7B “Sage” model trained on 100k hours from a Toyota line and performs very well on that task (unsure for other tasks). Teleop interventions catch failures and record them to improve the model. Synthetic data is growing too, ultimately for real2sim2real: bring reality into sim, train and test at scale, transfer back. GRAIL, Play2Perfect, etc. Dexterity got a lot of attention. T-Rex used 100 hours of tactile data to sharply raise task success. Some papers drop vision and rely on touch alone... but again, why? Feels very detached from practicality. Unlike Walden, tactile research is still far from industry deployment.
1
3
1
9
1,562
Most data collection starts with what’s easy to collect. At @PsdnAI we start with what models get wrong. We care about how data improves the models. (1) Create a holistic benchmark (2) Find where models & APIs fail (3) Run collection campaigns across high-value channels for that data (4) QC and annotate (5) Fine-tune, improve, & benchmark OSS models on that data -> loop We’re now taking the same loop to Physical AI. We care about how ego video, audio, and tactile data improve models and robots in the wild. DM me for more info on how you can get involved and benefit from this loop!
Better data is the breakthrough. Today, we're publishing a case study on how our 10,000-hour Bengali speech dataset reduced Character Error Rate (CER) by ~32% on Whisper Large V3. ↓
2
7
379
Jongwon 🇰🇷 @ CoRL retweeted
Swarms are quickly becoming one of the most important topics in Silicon Valley, sharing a few notes from recent conversations: - The naive view of swarms is that they’re a sum-of-parts method of parallelization / brute force, but we’re empirically seeing the “shape” of swarm intelligence be very different. - The way swarms achieve this is through a combination of parallelization, entropy, and specialization, leading to emergent solutions that aren’t covered by a small number of brilliant frontier models acting as central planners. - It’s also not obvious that the economics of swarms follow higher cost = better output. Swarms that mix and match frontier and non-frontier models with different reasoning levels may be more economical than a small number of frontier agents, and broadly, there should be a new Pareto curve here. - Swarms will be orders of magnitude more token-consumptive than agents, just as agents were compared to humans + models, further driving constraints in inference supply. I think it’s important to build intuition here so have been running some experiments that I’ll share soon. A great way to get started with swarms is the /swarms command in the new google anti-gravity. A few of the smartest thinkers on this topic are @alehandromz @jsnnsa @gokhan_egri @pirroh @amasad Love to hear other POVs on swarms if folks have them!!
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?
72
41
20
636
61,491
Data center construction is heavily bottlenecked by labor, doing so will make compute hilariously autonomous. That's in addition to designing chips using AI. A lot of closed loop coming.
Sam Altman says OpenAI will aim its robotics program at building its own data centers. Speaking on Stripe co-founder John Collison’s Cheeky Pint podcast, he said self-replicating data centers are now “within grasp”: AI-controlled robots running chip factories and building turbines, solar panels and more data centers, to produce “a lot of really cheap compute.” He argued robots could also make the parts that permits and supply chains hold up today, and said OpenAI already has its own chip and is learning to build its own data centers. On the cost of AI, Altman said it takes about $70 billion to build and run one gigawatt of capacity for five years, and that $20 billion would make him “much happier.”
2
128
State of the art in robotics LLMs are so good on existing tasks bc there's no ood anymore If you collect all factory work as data then your robot would work so much better, just fix hands and locomotion
1
123
If AMD "hires" Jensen, it's game over (ultimate family business)
1
213
We will open source more data and research as we expand our capacity. For researchers at CoRL, reach out for various data! Happy to collaborate on improving models with globally diverse datasets
Today, we're releasing EgoCrowd-500H, a new dataset accompanying our @corl_conf 2026 research on quality failures in crowdsourced egocentric video. • 300 stereo clips • 2.5 hours of footage • 50 human-reviewed clips Out now on @huggingface.
1
4
14
558
Continuous feedback during motion is a must. @JitendraMalikCV says we must RL as if a parent teaches a child atomic skills with video and tactile. Agentic robotic is a high-level policy to use fast "tools" - low-level atomic skills with semantic orientation.
1
1
17
1,566
The theme is "dexterity" at CoRL. Most booths are humanoid hands and robot arms! Cool hands from @origami_robo. @PsdnAI is researching visuo-tactile WAMs with own video and tactile data. Reach out!
4
4
2
37
6,590