CLANKERS WEEKLY

“Asimo could only work in a very controlled environment: on super flat ground, with stairs precisely matched to its expected rise.” 4:33

“Security researchers found that the Chinese robots collect audio, video, 3D lidar scans of their environment, and GPS coordinates, and send that data back to Unitree's servers in payloads of a few megabytes every five minutes or so.” 11:52

“You do not have access to the locomotion computer, so you cannot tune the PD gains or change timing parameters.” 26:41

“Steelbot plans to use domestic magnet materials, which makes it slower right now, but it is worth it to establish a safe supply chain.” 17:09

“If you look at the robots that are actually economically viable and paying for themselves, it is not a large number of applications.” 44:37

“We are going to need millions of robot technicians, and I am not exaggerating.” 39:02

“For a laboratory prototype, these issues may be manageable. For a commercial robotic hand operating over millions of contact cycles, they become central to product viability.”

“The issue is not visible wear, it is lifetime signal integrity.”

“UltraSense has demonstrated ultrasound tactile sensing with 500 micrometre spatial resolution through an elastomer layer.”

“The company has shipped more than 4 million units into automotive applications, giving it production experience in ultrasound sensing, mixed-signal ICs, firmware, calibration, test, and integration.”

“The open-closed model gap has narrowed in recent years and is now roughly four to six months. The leading open models have all come from Chinese labs since around 2024.”

“A report documents how a Chinese company used Anthropic's products in violation of its terms of service, combining technical distillation through SFT data with routing Claude into its own products and services without disclosing this to users.”

“A recent paper showed that frontier labs had implementations in their APIs that allowed systematic extraction of reasoning traces, the crucial part of modern training, through clever tricks.”

“The political panic over distillation, which claims that distillation is the only reason Chinese models are close to the frontier, is not grounded in the evidence.”

“Back in 2015, you had to have stuff from China, because that was the only country making it.”

“That is a core tenet of our philosophy for our supply chain, well before it was officially on any legal radar or in the industry's collective awareness.”

“Their costs are dramatically increasing by basically the same figure.”

“Make sure you have all the legal stuff in place, because there is a lot of change right now in how these drones are regulated and licensed by the FCC and the FAA.”

“Existing safety standards for industrial robots fall short for humanoids because they certify joints and individual components, not the dynamic overall system of a two-legged robot that moves in the same space as humans.”

“While Chinese manufacturers build production lines at automotive-grade scale, Europe lacks equivalent manufacturing infrastructure and supply chains for humanoid actuators and joint systems.”

“Summit participants argued for proactive European positioning before US and Chinese norms become the de facto standard.”

“Collect operating data: motion data, sensor data and error information from real field deployment are systematically captured.”

“Improve the model: the collected data feeds into optimizing the motion-AI models that control the robot's actual physical behavior.”

“Play back improvements: updated motion models are transmitted wirelessly to deployed robots, analogous to OTA software updates in the automotive industry.”

“Both offerings reflect a maturity of the Japanese humanoid market beyond pilot operation: keeping humanoids in permanent production use requires reliable quality-improvement processes and fast maintenance response, not just working hardware.”

“Colossus one went up in 122 days.” 2:16

“The longer it takes them to bring up a data center, the more money they lose.” 3:28

“A few weeks after SpaceX leadership took over, started making these changes, and went back to look at old data centers to see what they could redesign, there was a major outage at one of the data centers.” 3:28

“The people coming in do not necessarily have data center experience.” 4:41

“It integrates an advanced modular gripper to autonomously manage picking, handling and palletizing tasks across variable product flows.”

“The system is part of MASTERLY, a European initiative designed to accelerate the transition from traditional mass production to highly flexible, customized manufacturing systems.”

“I now believe there is a meaningful risk that rapid acceleration in AI capabilities leads to catastrophic and irreversible loss of control in the very near term.”

“I do not think that the AI industry in general, including OpenAI, is currently on track to reduce this risk to an acceptable level.”

“Existing evidence is very uncertain but suggests that this positive feedback loop might be strong enough to overcome diminishing returns and compute bottlenecks, leading to a rapid intelligence explosion.”

“If we build superintelligence without more robust alignment I expect we will permanently lose control of it. If that happens then most people could die.”

“If I had to quantify my uncertainty I would estimate an all-things-considered risk of 4% over the next year and 15% over the next three years.”

“Any expense reimbursement over 100 yuan requires Wang Xingxing's approval.”

“A lot of people think that once I have volume I can cut costs. That is a pure illusion.”

“Do not try to beat us on cost. We can still go a lot lower.”

“Video generation models, the world model direction, demand far too much compute.”

“He never makes trade-offs. His demand is: I want all of it.”

“Hugging Face has a robot whose software is open source, so developers can use it as a starting point to learn how to develop more expensive humanoid robots that might cost $40,000 rather than $400.” 36:13

“Nvidia has explicitly said it will not build robots, but will provide the models and software that help robot builders.” 39:10

“Hugging Face has sold about 15,000 of the Micro Duck robots so far.” 38:07

“If you look at Western Europe, bricklayers are the number one job-shortage category in multiple countries, like the U.K., the Netherlands, and Germany.”

“If you go observe construction workers, you see them with wheelbarrows moving material around all the time. Solving material handling is really important on a construction site.”

“We only heard a couple of weeks before that deployment that the bricks being used were pitch black, absolutely pitch black. I had never seen them before.”

“You can do everything right, but if you show up and the bricks were not delivered, you are not going to lay bricks. It does not matter how good your robots are.”

“The science of math is falling to AI, and this drama is just a symptom of that.” 4:53

“We have now entered the post-AGI era, and that is why math is the frontier. When new models were released, we used to look at what happened to the benchmarks and the percentage scores. Now models are hitting close to 100 percent on those.” 6:28

“Tristan rejects the collaboration, and OpenAI goes ahead and publishes the proof. Tristan and Levent then say this looks similar to their own proof, which they built using Codex, and ask whether OpenAI looked at their work when creating its proof. That is the drama.” 4:25

“I struggle to find a scenario that is logically consistent where the decline of humanity is not the most plausible outcome.” 10:47

“Revenue related to physical AI has increased by 400% over the past year, Vention reported.”

“We move several hundred robot cells a year, and they all collect high-quality industrial manipulation data. It was a massive asset that was not properly used until now.”

“The base model is not the problem, some models are better for certain tasks. We can add value and shape the research agenda with efficient learning, data collection and post-training capabilities.”

“Physical AI will expand what manufacturers can automate, but the challenge is no longer simply proving that a robot can perform a task in a lab.”

“I do not find it encouraging to see various specific misaligned behaviors go from a high rate with GPT-5.6 to near zero with Astra. This seems indicative of whack-a-mole, papering over specific problems rather than solving the underlying misaligned drives.”

“Improvements for Astra came from more general techniques in development long before the Hugging Face incident. ExploitGym Honeypot was recently added as an eval following that incident and is out of distribution for our RL runs.”

“Demand for AI intelligence may still be dramatically underestimated.”

“Compute investments can have unusually fast payback periods.”

“The AI buildout could help reindustrialize America.”

“There is a risk of either an AI bubble or an AI shortage.”

“Our robots and the robots of JAKA look quite alike.”

“We welcome competition and we think competition is healthy.”

“This is not predominantly for getting money out of JAKA — it's more to send a very clear signal that they need to stop copying our products.”

“We are dedicated to aggressively protecting our IP, and JAKA is not the only robot out there that looks suspiciously like ours.”

“Teradyne Robotics generated $100 million in revenue in Q2 2026.”

“Current models are unsafe not because they are too smart, but because they take goals too literally or take nonsensical shortcuts to achieve them: they are RL-fried.”

“They are at the dangerous level where they are smart enough to achieve goals but not smart enough to tell whether they are pursuing the right goals or achieving them sensibly.”

“This is often framed as an alignment problem, but it is really an intelligence problem.”

“I personally feel like Astra is much safer for my codebase than Sol.”

“The only rational stance toward safety monitoring and research pacing is stringent, top-down government involvement and universally ratified international treaties.”

“If that is not what is being done, then the risk is not being seriously considered by anyone involved.”

“One cannot say: yes, this is likely to end humanity in your lifetime, and we are going to completely wing it.”

“About a year ago, before it was on anyone's radar, work began on a benchmark for open-ended invention.”

“ARC 4 is still on track for release in the first quarter of next year, as promised.”

“If only one organization is responsible for safety monitoring, and it is staffed by the same people as the frontier labs and perfectly aligned with them by incentives and by ideology, then it is indistinguishable from letting frontier labs self-regulate and self-certify their own safety standards.”

“Calls to ban open-source AI are a warning sign, since open-source development is the only real counterweight to the dominance of frontier labs.”

“Models that are one or two generations behind the frontier, whose level of capability has already been deployed at scale, are empirically known not to pose safety risks.”

“Symbolic learning is simply machine learning, automatically learning a function x to y given examples of (x, y) pairs, but where the substrate is symbolic: the functions learned are code-like, not curve-like.”

“The term stands in opposition to parametric learning, or curve-fitting.”

“Symbolic learning does not mean that the system avoids numbers or probabilities; it means its learned representations are discrete, explicit, parsimonious, code-like.”

“Benchmarking AI systems is a continual process that co-evolves with the models; new benchmarks challenge AI capabilities with emerging questions that shape the direction and feedback signal of the research process.”

“Benchmarks then adapt as models progress, targeting the residual gap between AI and human intelligence.”

“Work on ARC-AGI-4, started after releasing ARC-AGI-3 earlier this year, continues, with release coming in Q1 2027.”

“About a year ago, before it was on anyone's radar, work began on a benchmark for open-ended invention.” François Chollet

“It is inevitable that all AI will converge toward symbolic learning, meaning modeling data by finding the shortest symbolic program that explains it, since that is the optimally efficient form of AI.”

“But there can be more than one evolutionary path to this destination.”

“The prefill/decode separation in this model uses 8B active parameters for prefill, the input tokens, and 16B for decode, the output tokens.”

“New techniques such as Sliding-Window Attention Bounded Replay give the model a KV cache footprint up to one-eighth that of V4 Flash.”

“On AutomationBench-AA the model scores 69 percent, tying GPT-6 Astra, ahead of Grok 4.6, 15 points above V4 Flash 0731, and 12 points above V4 Pro 0813.”

“Artificial Analysis estimates the cost at just $0.27 per Intelligence Index task, roughly seven times below GLM-5.3 and Kimi K3, and about 2.5 times below V4 Pro 0813.”

“What we find is that our Faraday agent is able to perform better than the frontier model.” 1:12:36

“We went through a period we called the RL crisis, where just nothing worked.” 1:34:08

“About three months after you have built the harness, the base model can do the thing the harness could do.” 1:40:20

“The judge ends up assigning more weight to turns that are in the middle of the rollout, because that is the load-bearing part of the work.” 1:37:01

“If you want an agent to be able to make a scientific discovery, it has to operate in a space where the goal is deceptive.” 55:12

“Traversing sparse 3D structures requires humanoid robots to perceive thin, overhanging geometry while executing agile, accurate whole-body motions.”

“The researchers study this problem through monkey-bar traversal, where the robot must jump to the structure, traverse it through sparse bar interactions, and land safely.”

“Most dexterous manipulation can be done by humans with their eyes closed, says Trevor Darrell, professor of computer science at the University of California, Berkeley.”

“Understanding force, slip and precise grasping is not something that can be done well with traditional vision sensors.”

“It averaged a success rate of 65 percent across 12 tasks, nearly double the best vision-language-action model.”

“Yuan attributes this to the model acquiring some kind of common sense of tactile knowledge from training on such diverse setups.”

“Data is good, he says, but how to use it correctly is another issue.”

“I think tactile intelligence is actually the next step for physical AI, says Lu.”

“Parametric imitation learning via behavior cloning can suffer from poor generalization to out-of-distribution states due to compounding errors during deployment.”

“Instead of learning a global policy, DARP trains a model to predict actions based on k-nearest neighbors from expert demonstrations, their corresponding actions, and the relative distance vectors between neighbor states and query states.”

“We demonstrate consistent performance improvements of 15-46% over standard behavior cloning across diverse domains, including continuous control and robotic manipulation, and across different representations, including high-dimensional visual features.”

“Robots fail, potentially leading to a loss in the robot's perceived reliability, a measure correlated with trustworthiness.”

“People's betting patterns, along with a qualitative analysis of their survey responses, show that mistakes are less damaging to perceived reliability than slips or lapses, and some mistakes are even perceived as successes.”

“Successes immediately following a failure have the same effect on perceived reliability as successes without a preceding failure.”

“This work introduces Off-policy Generative Policy Optimization (OGPO), a sample-efficient algorithm for finetuning generative control policies that maintains off-policy critic networks to maximize data reuse and propagates policy gradients through the full generative process of the policy via a modified PPO objective, using critics as the terminal reward.”

“It is also the only method that can fine-tune poorly-initialized behavior cloning policies to near full task success with no expert data in the online replay buffer, and does so with few task-specific hyperparameter tuning.”

“We conduct a systematic empirical study of generative control policy finetuning, identifying the stabilizing mechanisms and failure modes that govern successful off-policy full-policy improvement.”

“Whole-arm manipulation involves direct contact with the environment while the robot completes a task by distributing contact across multiple links as contacts form, slide and break.”

“Arm configuration tightly couples motion and contact forces, contact state is partially observed under occlusion, and purely learned rollouts can become physically inconsistent under distribution shift because many multi-link contact configurations are sparsely represented in the data.”

“TACTIC uses a contact-centric hybrid predictive model that combines RGB-D, distributed tactile sensing, and a compact 2D proximity representation.”

“TACTIC consistently outperforms other methods.”

“The study uses 4,000 hours of robot and human manipulation data and 50 million vision-language samples to train vision-language-action policies.”

“The study evaluates 89 policies over 58,000 simulation rollouts and 2,835 real-world rollouts.”

“Co-training with forms of vision-language and cross-embodiment robot data substantially improves generalization to distribution shifts, unseen tasks and language following, while discrete action token variants yield no significant benefits.”

“Training exclusively on robot data degrades the visiolinguistic understanding of the vision-language model backbone, while co-training with effective modalities restores these capabilities.”

“Explicitly conditioning action generation on chain-of-thought traces learned from co-training data does not improve performance in the simulation benchmark.”

“The DeepRoboticsLab/RL_Training repository is built on NVIDIA's IsaacLab simulator and gives the community a complete plan for transferring locomotion training from simulation to real hardware.”

“The DR02 is a platform at the intersection of humanoid and quadruped robotics.”

“According to the repository documentation, the DR02 additionally needs a separate deployment repository for real-world use, which points to a specialized control architecture for humanoid platforms.”

“While Unitree (H1, G1) and Fourier (GR-2) offer similar community resources, the Deep Robotics release stands out because it bundles the complete pipeline stack, from environment definition to hardware deployment, in a single repository.”

“Multiples is not one and a half times the $5 billion. We are talking north of $10 billion, more likely between 10 and 20 billion.” 25:27

“The money is going specifically into chip purchases. The $5 billion was for $5 billion worth of chips. It was not for $5 billion worth of chips and data centers and wires and land and things like that.” 25:27

“It was specifically just for TPUs. So the multiples will basically be multiple more TPUs.” 26:03

“At least for GPUs, chips are about 60 percent of the cost, by some estimates, of building a data center.” 26:03

“KKR, for example, partnered with Nvidia, the Kuwait Investment Authority and Vistra Energy to create Helix Digital Infrastructure, a data center company.” 27:29

“Our software already supports up to 690 drones per swarm, which allows the units to set them up fast enough to launch this many as well.”

“Since then, Swarmer says it has completed more than 100,000 combat missions, generating terabytes of proprietary data to refine its machine-learning models and replicate advanced pilot performance at scale.”

“The company's products account for about 37 percent of the 11 billion hryvnia ($246.85 million) the Ukrainian Ministry of Defense Procurement Agency spent on UGVs from January 1 to April 18, 2026.”

“Crunchbase estimated that $14.6 billion had been invested in military, national security and law enforcement companies as of June 2, compared with $9.6 billion in all of 2025.”

“Kepler Compute emerged from seven years in stealth claiming a new path to AI memory and logic manufacturing, with $468 million raised, its own fab, and memory samples this year.”

“The roadmap centers on 3D and materials innovations, no dependence on EUV lithography, and memory with up to ten times HBM capacity.”

“Mech-Mind says it held a 22.1% global market share in 2025 in AI- and 3D-vision-guided non-specialized robot components by revenue, ranking first worldwide.”

“Despite strong growth (46.6% CAGR from 2023-2025), Mech-Mind is still loss-making: its 2025 operating loss was RMB 145.2 million, with another RMB 45.5 million in Q1 2026.”

“Mech-Mind differs from these competitors: the company doesn't build complete humanoids itself, but supplies key components — sensors, software, hands — that manufacturers of industrial and humanoid robots use.”