M2.7 with Hermes CLI Replaces Most Claude for Coding
M2.7 w/ hermes cli is replacing ~75% of my claude code / opus usage now, but we need clarity for using it as a coding agent @ work. We're truly blessed to have the weights of this one, looking forward to seeing the license change. Definitely a model worth checking out.
Q4_K_XL 4-Bit Shines for M2.7 Local Runs
Some stats when running M2.7 locally at various quants and machine configs. Based on M2.5 research, the Q4_K_XL is probably best performer @ 4bit for this particular model, seems worth the trade in t/s. I tried various configs to get the GPU+CPU...
Model Drifts Unnoticed While Provider Denies Any Change
I just want my model's behavior to stop changing while the provider gaslights me telling me nothing has changed.
Consumers Crave Humanoids Even if Engineers Resist
What I usually tell people about humanoids is: The heart wants what the heart wants. People want humanoids. Roboticists want to change form factor because humanoids are hard. But consumers want humanoids.
Dell GB10 Cluster Powers
3x Dell GB10s, 128GB unified memory each, 384GB memory for the cluster, mesh connected over 200Gbps QSFP, currently running Trinity-Large-Thinking (398B) Q4_K_M 17-18 t/s. Will review this model in time. Using as coding agent and via Hermes. https://t.co/u2fKP1V0NE
KimoLab Merges Kimodo and Mjlab For
Introducing KimoLab: Kimodo + mjlab for prompt to physics-based motion matching for the unitree g1. I accidentally had claude make this and it's actually pretty cool. https://t.co/TIOFHFnZLi
Nvidia Engineers Burning $250K Tokens on Projects
You're gonna have to show me what these nvidia $500K engineers are doing to spend $250K in tokens. https://t.co/iibICUiTPO
Demo Ambiguity Fuels Overestimation of Tech Capabilities
It's not always a bold lie either. In demos for example, certain things are not made clear and left up to the viewer to just "guess." I cannot help but think the intent is to encourage and allow the viewer, journalists,...
Humanoid Robots Find Floor Pickup Harder Than Backflip
picking something up off the floor w/ a humanoid is more challenging than a backflip https://t.co/xvqORaNi7Q
SAC Simplifies Quadruped Gait Tuning over PPO
In a world of PPO everything for reinforcement learning, I've been tinkering with SAC for training a quadruped gait. This gait is trained purely on CPU (training on one of the Dell GB10s) on a single environment. Training any particular run...
Incentives Keep Frontier AI Labs Private, Excluding Public Investors
it's deeply unfortunate that incentives have aligned such that all these frontier AI labs and other tech companies are way better off staying private and the general public cannot invest in them despite desperately wishing they could.
New Agents Accelerate Robot Scene Creation and Task Design
still very much in r&d but one of my fav projects being worked on at lucky is these agents that make creating the scene and what you actually want the robot(s) to do faster, just such a cool concept.
Building the Robotics Simulator I Dreamed Of
I’ve joined Lucky Robots as director of eng & AI to help build the robotics simulator I wish we had right now. If you’re someone who likes to make robots do cool stuff or want to shape a sim being built...
Hand‑Centric Policy Works Sim‑to‑Real, Needs Lidar Localization
a more in-home-acceptable hand-centric policy. sim2sim shown, it does work sim2real, but need to improve localisation w/ lidar. Changes: Good hand rotation, torso/head/camera actually faces where we intend to go. Semi decent walking, crouch, and general pose. Still all PPO https://t.co/hK5F0sJuJL
Fine‑tuning PPO Rewards Yields Smoother, Slower Gait
further reward crafting for hand-centric (tm) PPO model. Closing in on what I want. need to slow down max velocities and I want more stepping/stride + less tip toe on left foot ideally. almost cant believe this is actually just RL! https://t.co/ro56yI8yaV
