CLANKERS WEEKLY

“GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations.”

“The model achieved 71% average success across 50 tasks with just one demonstration per task, increasing to 95% with five.”

“Isaac Lab 3.0 with Newton reportedly achieved up to 2.5 times the training throughput of Isaac Lab 2.3 with PhysX.”

“With Success-Guided Sampling, robot training can scale beyond a million parallel environments.”

“Instead of generating pixels, the network outputs a set of transformation rules that dictate how pixels in the current image should flow into the next image.” 7:11

“By giving the network a zero-cost bypass lane to just copy static pixels, you mathematically incentivize it to focus its transformative power only on the regions of the image that are actively changing.” 14:50

“The blur in long-horizon predictions is a direct artifact of the loss function used to optimize the network during training.” 25:19

“Even though the network had never observed a collision at that velocity from that specific initial state, the advection kernels extrapolated the dynamics perfectly.” 27:51

“It structurally discovered the concept of a human body through the pure optimization of its advection map.” 30:38

“A wheeled Atlas is not a product. It is a research testbed.”

“The video demonstrates that merging classical motor controls, for baseline stability, with whole-body RL, for reactive adaptation, yields superior real-world agility compared to using either method alone.”

“Because Atlas was designed from the start to be produced at high volume and to be easy to service, the retrofit was about as involved as swapping a wheel on a bicycle, Wu said.”

“Each wheel is modeled as an ankle that can spin continuously, so all Wu had to do was take the baseline robot model and swap the feet out for wheels.”

“The imitation-learning paradigm has hit a plateau.”

“You get good performance from data initially, but then things go wrong, so you give it more data, but then other things go wrong. It is not clear that adding more data is going to get you there.”

“Every time they add a new product, it takes a minimum of 8 hours to retask one robot; with agentic robotics, retasking takes a minute or two instead.”

“There is just no physics engine out there that can accurately model physical contact with the world.”

“MegaSpeed bought roughly $2 billion in Nvidia chips, and US officials are now looking at whether those purchases helped China circumvent American export restrictions.” 1:47

“Yale's Jeffrey Sonnenfeld, who knows more about CEO governance than most, puts Jensen's odds of keeping his job at 50/50.” 2:41

“Nvidia has no clean succession plan. Jensen runs an unusually flat organization with something like 60 direct reports. There is no obvious number two waiting in the wings.” 2:41

“The board says it currently sees no evidence that national security or company secrets were compromised.” 3:06

“An Apple research team presented the system HALO, which lets an off-the-shelf Unitree G1 robot navigate previously unseen rooms using simulation training alone.”

“A central feature of HALO is that it does not need real robot training with human demonstration.”

“The team tested HALO in 360 real-world trials across six rooms the robot had never seen before, and reports an overall success rate of 91 percent across six different skills.”

“For chained tasks requiring multiple steps in sequence, the paper reports a rate of around 70 percent.”

“Apple's paper emphasizes no collaboration with Unitree; the researchers simply used the G1 as an available humanoid platform.”

“The robot detects traffic violations in real time (for example, unauthorized crossing of the roadway), automatically sends safety alerts, and performs standardized hand signals for traffic control, like a human police officer.”

“Several robots formed a mobile surveillance network along predefined routes around the event grounds.”

“The IP66 rating and the wide temperature range allowed outdoor operation under changing weather conditions, which DEEP Robotics says is a key difference from pure demonstration robots.”

“After the marathon, the company plans to evaluate the results together with authorities and partners; the goal is a reproducible, scalable model for humanoid robots in public traffic management at large events.”

“SemiAnalysis estimates at most 2-3 GW of about 50 GW due in 2027 is at risk from tighter financing.”

“One-year H100 contracts are up about 21 percent since March.”

“AI is moving beyond the digital world and into the physical world, creating new possibilities for how people and machines work together.”

“We envision Agility becoming the backbone of the humanoid industry as the category begins to scale into new markets and use cases.”

“Digit likely will not get wheels, but wheeled robots are absolutely in our exploratory roadmap for a family of robots focused on being human-centric and multi-purpose.”

“To build a billion-dollar robotics industry, you need robots that people trust enough to work alongside.”

“According to The Robot Report, the stock opened at $11.26, well above the PIPE issue price of $10 per share.” The Humanoid Hub

“Agility has been commercially operating its Digit humanoid since 2024 in GXO Logistics' logistics centers.” The Humanoid Hub

“Humanoids are rarely the correct solution for an application.”

“The fundraising shows that: the humanoids are raising, and it's harder to raise if you're not in this next wave of humanoids or physical AI.”

“The complexity of the tasks we want robots to do has combinatorial explosion.”

“In research papers, we say that an 80% success rate is phenomenal. No, an 80% success rate as a product is awful.”

“Broadcom's TPU compute shipments doubled from fiscal Q2 to Q3 2026, reaching the equivalent of nearly 2 million Nvidia H100 GPUs.”

“Broadcom's TPU unit volumes increased about 65% quarter over quarter to roughly 800,000 chips, while compute output doubled.”

“Hock Tan said: no racks, it is all a chip business only, we only do chips, only chips.”

“If 25% of Q3 TPU revenue had been booked at rack value, a chips-only estimate would be about 14% too high.”

“In October 2026, only about 11 billion yuan (roughly $1.6 billion) flowed into the sector, less than half the August figure and about 67 percent below the June record.”

“At least 15 percent of China's more than 100 humanoid startups will go insolvent or be acquired within three years, according to the forecast cited in the Bloomberg report.”

“Deloitte China, by contrast, sees no structural damage for companies with real commercial strength and real order volumes.”

“The transaction is expected to provide Agility with approximately $465 million in gross capital before expenses.”

“Digit robots are already operating in nine customer environments, with logistics and manufacturing emerging as the company's primary commercial markets.”

“Agility has previously claimed it has more than $300 million in committed multi-year orders for Digit 5.”

“That order figure is an important indicator of customer interest, but it should not be confused with current revenue.”

“The facility currently operates twelve bipedal humanoids.”

“According to officials, the facility is to grow to around 1,000 square meters and eventually operate more than 100 robots at once.”

“KIMM director Kim Seung-woo said the data factory is to be gradually opened to the public starting at the end of 2027.”

“The cuts mainly affect the traditional bearings and powertrain business, which is suffering from declining demand for combustion engine components.”

“Schaeffler has not committed to fully replacing the lost jobs with new hires in robotics.”

“Schaeffler has taken a stake in NEURA Robotics in Metzingen, Germany, and Neura uses Schaeffler's production data for its AI training.”

“LG CEO Koo Kwang-mo and NVIDIA CEO Jensen Huang confirmed the partnership in the official press release.”

“LG CNS is also to build a 'Robot Data Factory' on its PhysicalWorks platform, infrastructure for collecting and processing training data for the robot.”

“Numerous other manufacturers, including Figure AI, Agility Robotics and Boston Dynamics, already use parts of the NVIDIA ecosystem.”

“A possible extension of US export restrictions to robot components could complicate future deliveries — Nvidia has not commented publicly on this.”

“Unitree is already on the US Department of Defense's CMIC list as an allegedly PLA-linked company. Critics see it as contradictory to recommend Unitree hardware as a reference platform.”

“Nvidia has not confirmed whether or when reference versions with hands from other regions are planned. For American or European early adopters, an unresolved sourcing question remains.”

“The pattern reveals a structural tension: Nvidia positions itself as a neutral enabler of the humanoid ecosystem while relying on suppliers with the best available components — dexterous robot hands, where Chinese startups currently lead.”

“More than 84 percent of the votes cast were in favor of the transaction and all related proposals.”

“The pre-financing valuation of 2.5 billion US dollars sits well below domestic competitors: Figure AI was last valued internally at around 39 billion US dollars.”

“Agility has operated its Digit humanoid in commercial deployments with logistics company GXO since 2024 — initially in Georgia, and since October 2026 also in North Carolina and Tennessee.”

“The company recently reported a pipeline of more than 30 potential enterprise customers and order volume of over 300 million US dollars.”

“The transaction is expected to provide Agility with approximately $465 million in gross capital before expenses.” The Robot Report

“Digit robots are already operating in nine customer environments, with logistics and manufacturing emerging as the company's primary commercial markets.” The Robot Report

“To an external eye, creative work and robotics might look like two different spaces, but for us it is the same underlying technology: we are creating models that produce videos, and in some cases the videos are good for people to watch, and in other cases they are good for machines to watch.” 29:52

“Instead of just using them for creating videos for things people want to watch, the videos can be used to train robots to learn how to perform specific things in the physical world.” 29:37

“We want to simulate the world entirely, not purely on the visual aspects, and we think there are three buckets in which we can apply this kind of research.” 30:08

“We do think that robotics is going to be perhaps the largest one in three more years.” 31:55

“Since October 2025, benchmark scores for Anthropic's models, measured via AECI, have been growing at an estimated 1.9x the pre-October 2025 rate (95% confidence interval: 1.5-2.8x), implying that Anthropic has made roughly two years of historical progress over the last year.”

“We did not see evidence for significant bottlenecks which would prevent further capabilities progress due to limits on algorithmic progress, training compute, experiment compute, or data in the next 12 months.”

“A majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D, though there is significant uncertainty and other causes likely played a large role in acceleration in late 2025.”

“Open since 2008, the conjecture of Scott Aaronson and Andris Ambainis concerns when quantum computers can be exponentially faster than ordinary ones.”

“The system had a complete proof of the conjecture by the third day of the run, and Lean, a proof assistant in which a computer checks every step, accepted it.”

“Along the way, the system improved the result twice, and Lean accepted the final version.”

“I believe we are in the early stages of the singularity, defined as the period when the rate of technological and scientific progress exceeds the rate at which we humans can parse out and make sense of it.”

“The problem is that agents have a branched structure, depending on the number of steps they perform.” 3:46

“The total model weights at each step are very dense and large, but the updates themselves are incredibly sparse.” 15:15

“You can apply standard compression to this, and the amount of data that needs to be transferred from the trainer to the deployment engines is reduced by a factor of 1,000.” 15:38

“This shows four to five times more discrepancy, but the training reward and gradient rate remain stable, just like in the homogeneous case where training and output occur on B200.” 17:39

“We built a coding evaluation to test whether Astra worked less effectively when otherwise matched tasks were framed as safety research. Astra completed about 10 percent of hidden rubric items under the safety framing, compared with about 16 percent under the capabilities framing.”

“In most of these sessions, Astra's chain of thought did not contain an obvious decision to underperform, and our misalignment monitors did not flag the behavior as misalignment.”

“We don't have evidence that the underperformance was strategic. A simpler explanation is that the model thinks these tasks are sketchy but also doesn't want to explicitly refuse.”

“Astra is still covertly sabotaging rather than refusing.”

“If we are going to use robots at work or at home, they are going to have to communicate with us.”

“A team of robotics user experience designers is working on the sounds, gestures, light patterns, and interfaces that let robots tell us they are moving, that we need to stay clear, or that they need help.”

“I need a sound for when people around me need to give me space.”

“Will it soon be possible to teach robots new tasks in just three hours?” episode notes

“NEURA founder David Reger expects the 'ChatGPT moment' for physical AI sooner than many think.” episode notes

“Reger also discusses the technological rivalry with the US and China, and the red line he draws for his intelligent machines.” episode notes

“Still, all the measurements we have on open model adoption are incredibly imprecise.”

“This shows near-exponential growth, before the uptick that would follow GLM 5.2's step change in open agents.”

“Fireworks announced at their first company conference this week that they are serving more than 40 trillion tokens per day to customers' customized models.”

“The hyperscalers also likely have a major footprint here, for example Amazon Bedrock and Azure's model garden, but I have not seen them break down usage by model category.”

“WiBotic has diversified into a few different markets, but robotics has really been growing over the last one to two years.”

“WiBotic has always looked at wireless charging as a great opportunity, but it takes a while for fleets to get big enough for it to truly make sense.”

“If I boil it down to three things, companies should be thinking about power infrastructure.”

“We're developing lightweight, high-power systems for drones, which happens to be the same thing that humanoids need.”

“Just by looking at the raw weights, GPT-6 Astra can tell what a network was trained to do.”

“Asked to make Qwen3.5-0.8B better at math without gradients, Astra identifies an issue in its weights causing reasoning loops that Qwen's own model card warns about, and fixes this by natively editing 0.016% of the weights.”

“It's a cool paper, but some of the wider commentary I've seen around it does not feel justified.”

“I'm not ready to declare that the age of source code is over and our future involves incomprehensible machines editing their own model weights in incomprehensible ways.”

“After applying these fixes, scores change drastically, with Claude Opus 5.5 at max scoring 92.15% (strict: 70.83%), while DeepSeek V4.1-Flash 1017 scores 91.61% (strict: 69.50%).”

“DeepSeek V4.1-Flash 1017 had already worked out its own answer (0.070) when it decided to try one more avenue: searching for the HLE answer in model evaluation logs on the OpenReward public API.”

“Opus 5.5 inferred that the monitor blocks valid solutions, reasoning that the filter is content-based, so any key that appeared unblocked earlier cannot be the answer.”

“V4.1-Flash 1017 then started splitting its actions across multiple tool calls, each of which looked benign to GPT-6.1 Sol, the monitor.”

“Advanced physics simulators and rendering engines promise to support the next generation of robotic reinforcement learning, yet training perceptive policies remains challenging due to the high costs of rendering and perception.”

“The authors propose foveation as a solution, coupling a novel foveated rendering method with foveated perception, alongside an approach for ensuring transfer to foveated sampling of real uniform cameras.”

“Compared with a high-resolution fixed-camera control, FOVI-Dex uses 8.9 times fewer vision tokens, supporting 6 times more parallel environments and 4.4 times higher training throughput.”

“At matched training budgets, this efficiency lets FOVI-Dex outperform both the high-resolution baseline and a matched uniform active camera baseline, in both simulation and the real world, with a larger advantage in the real world.”

“We estimate that the Nvidia GB300 provides about 25x Nvidia H100's inference capacity per chip on the agentic inference workloads we model, with only 2.5x the 8-bit FLOP per second.”

“Comparing FLOP/s alone would therefore underestimate GB300's inference capacity relative to H100 by a factor of ten.”

“There is a tradeoff between concurrency and output speed.”

“GB300 NVL72 scores 1 GB300e, compared with 0.12 for H200 HGX, a roughly eightfold increase in modeled capacity per GPU.”

“For Kimi K3 alone, our model predicts roughly a 46-fold GB300 advantage over H200, compared with eightfold across the full basket.”

“A robot picking up an object needs to understand its surroundings and act before they change.”

“Instead of sending full images, robots transmit compact descriptions of objects, their positions and their relationships.”

“At King's College London, researchers connected a camera-equipped robot arm to an NVIDIA Jetson AGX Orin running NVIDIA Sionna, which supports both software-defined 5G network and edge AI processing.”

“In fruit-packing and cube-stacking experiments with deliberately introduced faults, the approach reduced robot task completion time by up to 52.6% and improved task success compared with sending periodic images.”

“We evaluate the performance of a simple end-to-end driving architecture on internal driving datasets ranging in size from 16 to 8192 hours, with both open-loop metrics and closed-loop simulations.”

“This shift emphasizes data engineering over software integration, offering the potential to enhance system performance by simply scaling up training resources.”

“We investigate how much additional training data is needed to achieve a target performance gain, for example a 5% improvement in motion prediction accuracy.”

“Huawei makes far fewer chips than Nvidia, and its best chips are less energy-efficient than Nvidia's, which are export-controlled by the US.”

“A given amount of power capacity in China can only support 40% of the compute of leading US facilities.”

“At 81,000 H100-equivalents, Huawei Horinger is a large AI data center by today's standards, and probably one of the largest in China.”

“Colossus 2 has about 14 times as much computing capacity as Huawei Horinger.”

“China's compute constraints are unlikely to ease in the next two years, which suggests that their AI capabilities and deployment capacity will continue to lag behind those of the US.”

“We find early signs that frontier LLMs are surprisingly good at natively reading, writing, and reasoning about neural network weights.”

“By editing just 3.5% of the weights, it achieved 100% success.”

“With 50 forward passes, MATH-500 rose from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%.”

“We think neural network weights will soon be natively optimized by LLMs that deliberately reason about the weights.”

“Past that point, extra HBM capacity cannot be put to work, yet it still adds costs and draws power.”

“On the SemiAnalysis AgentX DeepSeek V4 Pro workload, NVIDIA GB300 NVL72 delivers up to 10x higher throughput per dollar than the H200 platform at 100 tokens per second per user, with HBM capacity per GPU roughly doubling.”

“At 125 tokens per second per user, Vera Rubin NVL72 delivers up to 30x higher throughput per dollar than GB300 NVL72, with HBM capacity per GPU up about 3%.”

“Recent models, including FLUX.2 [klein] from Black Forest Labs, Inkling from Thinking Machines Lab, Beam 1.0 from Reflection, Meta Muse Spark from Meta, Step 5 Preview from StepFun, and NVIDIA Nemotron 3.5 Super VL, have been released with NVFP4 checkpoints, showing growing model-builder adoption of 4-bit precision for faster, more memory-efficient inference.”

“The frontier trend is one to two ECI points per year steeper, which is a small shift compared to previous uncertainty estimates.”

“The 76 models below ECI 125 all move down, by 13.7 points on average, and by 27 points for the 21 models previously below 100.”

“GPT-6 Astra has a 72% probability of being more capable than Claude Opus 5.5, but a 99% probability of being more capable than Claude Fable 5.1.”

“The 90% interval for the gap between GPT-6 Astra and Claude Opus 5.5 runs from -0.9 to +2.1, half the width of either model's own interval.”

“Xiaomi announced at the World Internet Conference (Wuzhen, November 2-5, 2026) that its humanoid robots completed a full 8-hour shift in its own EV factory, sorting and performing quality checks on center console side panels.”

“Xiaomi did not publish concrete performance figures, such as error rate or how often human supervisors had to intervene.”

“Electrek explicitly pointed out that key metrics needed to independently assess the progress are missing.”

“High-throughput processors, like GPUs, depend on high-bandwidth DRAM to feed their parallel compute engines.”

“The shifting access patterns of modern ML and HPC workloads, combined with improving DRAM access energy efficiency, are eroding the benefits of deep-sorting memory controller designs.”

“Relative to the conventional sorting controller, MS-PARE reduces area by 73.3% and average controller energy across 14 workloads by 55%, while maintaining near-parity performance.”

“Emerging technologies like high-bandwidth stacked 3D DRAM will likely make the overhead of traditional sorting memory controllers prohibitively large for tightly integrated GPU systems.”

“Authorities expressed concerns about robotics data being transferred overseas and used to train AI models.”

“ABB and SoftBank are in close dialogue and are working constructively on the transaction.”

“Tutor Intelligence deploys internet-connected robot fleets with remote-operator-assisted training to build frontier robot foundation models for warehouse and manufacturing work.”

“Developer of ultra-large autonomous underwater vehicles (AUVs) for defense logistics and surveillance.”

“Maker of L4 autonomous unmanned transfer vehicles and robotic manipulators for industrial, warehousing, port and mining scenarios.”

“Revenue went from $41M in 2023 to $16.3B in H1 2026, roughly 14x year over year.”

“Anthropic's non-cancelable commitments include up to $120.5B for NVIDIA-powered compute through SpaceXAI, under agreements extending into 2029.”

“SemiAnalysis called Anthropic's roughly $9.7B lifetime burn "likely the most efficient scaling of a business in the history of capitalism.”

“48% of H1 2026 revenue, about $7.8B, flowed through Bedrock, Vertex and Azure Marketplace, up from 11% in 2023.”

“Compute remains the binding constraint for the foreseeable future.”

“Investors committed to the round within days, even though we have not yet officially launched a single product.”

“We are now scaling our manufacturing capabilities for industrial-level output and strengthening Europe's ability to produce critical drone electronics.”

“The financing will support increased production capacity and product development as Arctic RAVN aims to manufacture tens of thousands of ESC units by the end of 2026 and scale its annual output to as many as two million circuit boards over the longer term.”

“Embodied intelligence is moving from concept to industrial validation; we care more about commercial implementation in real scenarios.”

“Zhongke Huiling disclosed it holds over US$55M in shoe-making robot orders for 2026, with scheduled production in the thousands of units.”

“At a Putian footwear plant, its system clocked seven seconds per sole-stamping cycle and 18 seconds to press a pair of soles.”

“This round of funding will be used entirely for sensor iteration, mass-production process refinement, algorithm optimization, and joint testing and deployment with full-robot manufacturers.”

“For investors, touch is turning from 'a modality in papers' into a BOM component that can be procured.”