Joe Clinton

@PgChiyo If you've kept the same motors yours arms now have a rated payload of 0g, and a peak payload of 250g. You need to at least replace the first few motors with STS3250.

Updated version of ALOHA Mini, called 🦂ALOHA SCORPION ver 2.0 is out !! github.com/k1000dai/Aloha… - better chassis - larger battery - updated gripper (using PincOpen) - better arm position to reach a higher place Here is the VIDEO which replays the collected dataset!

Recently have been working with image-to-video generation models a lot more, so I put together this graph to help determine the best video model for any price point. Seedance-v1.5-pro stands out the most to me as the optimal choice to balance quality and cost.

@ihorbeaver Perhaps you could speed up the model and then learn a residual network to adapt the action decoder to the wobble with online-rl?

The advantage of arms with industrial internals is that they don’t wobble, so the AI model can control it faster by just multiplying frames per second. Here’s we multiplied the FPS by 3× compared to teleoperation (180 instead of 60).

NEO’s Starting to Learn on Its Own

@thealexbanks This doesn't account for developers moving to untrackable local agents like claude code, codex, cursor and copilot in the same timeframe. Claude Code is far ahead of Codex which is in turn ahead of Gemini.

OpenAI lost 22% market share in 12 months. Gemini is eating their lunch. The first Global AI Tracker of 2026 just dropped. Here's what caught my attention. Market share as of Jan 2nd: → ChatGPT: 64.5% → Gemini: 21.5% → DeepSeek: 3.7% → Grok: 3.4% → Perplexity: 2.0% → Claude: 2.0% → Copilot: 1.1% 12-month transformation: → ChatGPT dropped from 86.7% to 64.5% → Gemini exploded from 5.7% to 21.5% → Grok didn't exist and now approaching DeepSeek 12-week data change: → OpenAI: -22% (their worst period on record) → Gemini: +49% (relentless momentum) → Claude: -14% (stable but dipping) → Grok: +52% (fastest growing)

@Ciszek @chris_j_paxton The vae has a 16x16x4 compression. The model begins with a 480x640 input, so 4 frames is compressed to 600 tokens. The input is 5 context frames + 4 noisy latent frames. The DIT generates this in a single step then passes to the action head. This is not a problem.

@chris_j_paxton 1 frame = 196 tokens 16 frames = 3136 tokens

It seems very clear to me that video models will replace image models for robotics over the next year Video models make so much more sense for robotics tasks which usually involve some occlusion and partial observability (Video from mimic-video paper, eth)

@chris_j_paxton VLA's with a video model backbone are my PhD topic. Wholeheartedly believe they are the way forwards and will share some exciting progress on this front later in the year.

@ycombinator @gentrajectory So it's affordance driven grasping like arxiv.org/html/2508.0889… ?

In just ~3 months, as a solo founder with no prior robotics experience, @gentrajectory trained a foundation model for dexterous manipulation that lets humanoid robots pick up unseen objects and perform real-world work. It generalizes to novel objects and scenes, including cases where prior SoTA models achieve 0% success. Congrats on the launch @joshuabelofsky! ycombinator.com/launches/P6q-g…

@joshuabelofsky Doesn't look accurate enough to be useful unfortunately. I think the data collected would be low quality and it would impact the resulting policy.

Open-sourcing my internal dexterous teleop stack! github.com/GeneralTraject… It uses vision + Vive wrist trackers instead of data gloves → about $500 in hardware vs. ~$5,000.

@mo_danesh @k7agar You can't guess anything from a such a tiny amount of information. Why even bother trying to help. There's hundreds of possible reasons a VLA model might underperform.

@k7agar IG the LR is too small to force the model to learn the semantics.

unfortunately loss curves mean nothing in robotics </3

@k7agar Loss curves are useful for comparisons between models that have same datasets and same loss function.

@chatgpt21 45s Christmas ad for mcdonalds with no speaking roles, 18 locations, 45 actors, 90 extras, 3 cgi shots would require a budget of >$1 million. It's likely they spent about 10x less on this ad and even negative attention is still attention.

McDonald’s just dropped a new AI ad and it’s beautiful and I am genuinely tired of people pretending this is not the future of media. If this played on your TV during a normal commercial break, you would be disingenuous to say “its slop” or “I could easily tell it is AI.” It is a fantastic ad on its own merits, and it is obvious that AI video will eventually be one to one with reality, where you truly cannot tell the difference. if your of average intelligence and can extrapolate of course. When we get there, then what? Is it still “slop,” or does “slop” permeate as a label for anything made with AI, even when you cannot tell at all?

@lukas_m_ziegler I think this could have been done way cheaper by just waiting for the heated bed to cool down and then repeatedly ramming the part with the flat side of the extruder head until it unsticks and then pushing it off the ledge onto a cushion below.

A robot for the 3D printing farm! 🖨️ 3D printing is often tied to a repetitive cycle: wait for the print to finish, remove it, clean up, start the next one—and repeat. But what if there were a solution that changed all of that? This robot powers an entire 3D printer farm! With this system, printing can run non-stop, as long as there’s filament to feed it. The robot handles the rest: collecting finished prints and placing them neatly on the rack, ready for the next job. Great engineering by DHR Engineering! 🦾 ~~ ♻️ Join the weekly robotics newsletter, and never miss any news → ziegler.substack.com

@KLieret Hi, when will you update with GLM 4.6, Kimi K2 thinking and Minimax M2? Would love to know how they compare.

Opus 4.5 reclaims the top of the official SWE-bench leaderboard with 74.4%, narrowly ahead of Gemini 3. Cheaper than Opus 4, but more expensive than Gemini. Takes less steps than Sonnet 4.5, but still run for >100 steps for optimal performance. Details in 🧵

@liyitengx @RemiCadene Hi, first off this is amazing! Secondly, wanted to ask two questions: 1. why you didn't go for an off the shelf telescopic lift solution? 2. What is the payload of the lekiwi base and do you think it's overloaded?

AlohaMini wouldn’t exist without LeRobot. Thank you @remicadene for building such an inspiring open-source robotics framework. Using LeRobot, I built a dual-arm, 3D-printed robot with a lift — and now I’m open-sourcing everything too. GitHub: github.com/liyiteng/Aloha…

@vbingliu Could you test models with their preferred agent that they recommend (Claude with Claude Code, GPT-5 with Codex, Gemini with Gemini-Cli, Qwen with Qwen-code)? The right agent pairing should significantly boost performance.

🚀 Introducing SWE-Bench Pro — a new benchmark to evaluate LLM coding agents on real, enterprise-grade software engineering tasks. This is the next step beyond SWE-Bench: harder, contamination-resistant, and closer to real-world repos.

gpt-oss is now on @ArtificialAnlys, and is absolutely dominating the Pareto frontier of intelligence vs cost!

GPT-5 significantly underperforms expectations on SWE-bench verified with a score of 74.9%. This suggests progress on swe from ai is slowing down. We are yet to see GPT-5 + Codex scores which should be higher - hoping for a 78% score by the end of the month.

The new Qwen3 fills a much needed gap in intelligence vs cost. I'd recommend gpt wrapper startups that are using 4o, begin using Qwen3 instead, for a significant boost in intelligence whilst actually REDUCING cost.

@mark_k O3 did it without problem. chatgpt.com/share/685deb40…

Grok 3.5 generated this, wanna bet? 👀👀