Shiza Charania

35 followers · 40 following · 2852 views

on the atlas · 80

highlights · 148

You have to be the right kind of determined, though. I carefully chose the word determined rather than stubborn, because stubbornness is a disastrous quality in a startup. You have to be determined, but flexible, like a running back. A successful running back doesn't just put his head down and try to run through people. He improvises: if someone appears in front of him, he runs around them; if someone tries to grab him, he spins out of their grip; he'll even run in the wrong direction briefly if that will help. The one thing he'll never do is stand still. [7]
The Hardest Lessons for Startups to Learn
benchmarking our initial SKUs against what they need in order to get larger offtake agreements against an aggressive manufacturing ramp, which justify making larger buys and investments of input materials
America Spins on Westmag - Not Boring by Packy McCormick
The design-cycle time for actuators right now, for almost anyone in the US, is in the months, because you rely on overseas suppliers to order pretty much everything.
America Spins on Westmag - Not Boring by Packy McCormick
Stack those tolerances on top of each other (lamination, winding, magnet orientation, rotor balance, bearing fit, etc…) and you start to see why motor manufacturing is mostly a process problem, not a materials problem. The materials are simple. The process is annoyingly precise.
America Spins on Westmag - Not Boring by Packy McCormick
The truest driver of a decision is not whether it makes sense but whether it makes you more: more awake, more free, more alive, as if a magnetic pole has been planted somewhere in the earth just to make your heart expand the closer you are to it.
To be alive is to live wanting something - Sherry Ning
The tactile map drifts with temperature, so recalibrate if the sensor has been powered on for more than 30 minutes.
LeFlexiTac Docs
policy to focus more on force input during initial training phases and gradually balances force with visual inputs as training progresses
FACTR: Force-Attending Curriculum Training for Contact-Rich Policy Learning
disregard irrelevant visual details once contact is established and rely solely on force feedback to perform tasks such as lifting a box or rolling dough
FACTR: Force-Attending Curriculum Training for Contact-Rich Policy Learning
contact force signals are typically less discriminative, often remaining near zero for extended periods when the arm is not in contact with the environment during an episode
FACTR: Force-Attending Curriculum Training for Contact-Rich Policy Learning
creates a random grasping force against the load cell, which we compare against the FORTE readings
FORTE: Tactile Force and Slip Sensing on Compliant Fingers for Delicate Manipulation
To characterize how well we measure grasp force across different shapes, we perform multiple grasps on custom load-cell testing indentors
FORTE: Tactile Force and Slip Sensing on Compliant Fingers for Delicate Manipulation
environment companies hire overseas developers to replicate the UI of sites like DoorDash or Uber Eats and sell the mockups to labs
RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures
Scaling up RL is difficult as it requires a steady stream of tasks the model needs to solve and learn from.
RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures
But OpenAI’s progress last year and ability to keep up using an older base model was existence proof of the efficacy of post-training.
RL Environments and RL for Science: Data Foundries and Multi-Agent Architectures
Figure 2:Pipeline for extracting qualitative policy characteristics from RoboArena’s rich evaluation data. We use a VLMs to categorize scenes and tasks, and then use an LLM to aggregate information across a large number of evaluation rollouts into a policy report that summarizes qualitative strengths and weaknesses, and cites concrete evaluation rollout videos as evidence.
RoboArena: Distributed Real-World Evaluation of Generalist Robot Policies
These are likely to expand, although major companies right now have too much to lose and too little to gain to bother competing.
How do We Quantify Progress in Robotics? - by Chris Paxton
Inference-time scaling with verifiers. Video models allow VLMs to be used directly as verifiers. By generating multiple videos and scoring them with a VLM, we can filter out unrealistic or low-quality generations. This provides a simple recipe for inference-time scaling [14].
Anirudha Majumdar on X: "Should we predict pixels for world models in robotics?" / X
hallucinations: objects that duplicate, appear from nowhere (see example below from our Veo work), disappear, or morph either spontaneously or when re-appearing after an occlusion.
Anirudha Majumdar on X: "Should we predict pixels for world models in robotics?" / X
urrent video models in robotics struggle to produce high-quality generations beyond 20-30 seconds at most.
Anirudha Majumdar on X: "Should we predict pixels for world models in robotics?" / X
(1) generating a video of imagined success and using an inverse dynamics model to infer the requisite actions [1, 2, 3], or (2) by directly optimizing plans using an action-conditioned world model [4, 5].
Anirudha Majumdar on X: "Should we predict pixels for world models in robotics?" / X
Pushing farther from the center of rotation produces more torque for the same magnetic force
Humanity's Last Machine
the motor must flip its magnetic poles at precisely the right moment
Humanity's Last Machine
notoriously difficult to join to metal frames without creating weak points
Humanity's Last Machine
China produces 85-90% of global magnesium
Humanity's Last Machine
processing magnesium in a semi-solid state (never fully liquid)
Humanity's Last Machine
humanoid limbs where every gram saved reduces motor workload and power consumption
Humanity's Last Machine
high-wear areas like shafts, gears, pins, fasteners, and bearings
Humanity's Last Machine
most reliable material for building structural housings
Humanity's Last Machine
The Inverse Dynamics Model (IDM): We bridge pixels to actuators by training an IDM to predict the exact action sequence required to transition between the model's generated frames. We use the IDM’s metrics and rejection sampling to enforce kinematic correctness of generations.
1X World Model | From Video to Action: A New Way Robots Learn
Proprioception and actions need to be inferred from the SLAM estimate of end effector pose Camera images all feature human arms holding a device, but at inference time the robot sees robot arms instead
State of Robot Learning, December 2025
To the public intellectual, one must ask: if your ideas are so good, why aren’t you executing on them? It is much easier (and less impactful) to write about the importance of “green tech” than to build Tesla. Moreover, many self-titled public intellectuals are not even particularly intellectual. They are just… public.
Choose Good Quests – Founders Fund
Besides that fundamental architecture, we also have an energy recycling architecture involving our muscles and tendons. We store energy in our tendons and reuse it on the next step — our Achilles tendon at the back of each of our lower legs is the one that stores most energy and the one most likely to rupture.
Why Today’s Humanoids Won’t Learn Dexterity – Rodney Brooks
If the big tech companies and the VCs throwing their money at large scale humanoid training spent only 20% as much but gave it all to university researchers I tend to think they would get closer to their goals more quickly.
Why Today’s Humanoids Won’t Learn Dexterity – Rodney Brooks
In the first video the person picks a match out of a box and lights it. The task takes seven seconds. In the second video the same person tries again but this time the tips of her fingers have been anesthetized so she no longer has any sense of touch right at her fingertips.
Why Today’s Humanoids Won’t Learn Dexterity – Rodney Brooks
Modern Vision-Language-Action models (VLAs) are actually quite bad at handling tasks which require even a very short memory.
Are Humanoid Robots Ready for the Real World?
They have no mechanism to learn or remember the dryer settings I prefer for different clothes, and no mechanism to know which clothes I prefer to hang-dry. The answer to these questions is not as simple as read the label.
The Current Crop of AI Startups is not Prepared for Big Worlds
If you can train yourself to ask “is there a better way to do this?” at random intervals ten times a day, you will become unstoppable.
50 things I know - by Cate Hall - Useful Fictions
“Are you in venting mode or solutions mode?”
50 things I know - by Cate Hall - Useful Fictions
when we train our robotic foundation model on, for example, human data, and then present it with a new problem, it will try to predict how a human will approach this problem, rather than predicting an effective strategy for a robot
Sporks of AGI - by Sergey Levine - Learning and Control
“so muchness” of modern life has given us commitment anxiety
Hugging the X-Axis - David Perell
combination of rapid or precise movements, adjusting motions based on subtle tactile feedback, and manipulating small, fragile, or deformable objects
Robot Dexterity Still Seems Hard - by Brian Potter
the older you get, the more expensive your time becomes
loneliness - by Elaine - manners & mystery
And because DOGE is constrained by a preexisting set of rules around reductions in force, the people DOGE is firing are disproportionately the more talented and more useful people
50 Thoughts on DOGE - by Santi Ruiz - Statecraft
number of federal employees and dollars saved
50 Thoughts on DOGE - by Santi Ruiz - Statecraft
making customers feel special for owning something rare
Lululemon
CIA has declassified extensive documents describing even more profound psychic abilities, like the ability to write on pieces of paper that are sealed in envelopes, or ‘remote view’ distant regions of space and time by just sitting there and letting your mind wander without even knowing the space-time coordinates you’re supposed to view.
Predictions for a Science Fiction Future - by Andrew Cote
Part of this awareness, of what I can actually do with several days of focused work and intention, is the pain of knowing how much I am not doing that anymore.
my phone is making me dumb - by Isabel - Mind Mine
“I’m going to schedule a phone call with you in two months to catch up, I’ll send you the invite — if we need to adjust when we get closer to the date, that’s fine.”
50 things I know - by Sasha Chapin - Sasha's 'Newsletter'
I know that it feels horrible to create from a place of defense
50 things I know - by Sasha Chapin - Sasha's 'Newsletter'
she accepts uncertainty as a universal phenomenon.
Things You Learn Dating Cate Hall - by Sasha Chapin