In brief archive
8 October 2026 · Afternoon edition
Updated on · Afternoon edition. Four signals worth your attention.
Most US adults think AI is moving too fast
A new AP-NORC poll finds that 64% of US adults think AI is developing too fast, while 81% rate keeping it under human control as an extremely or very important government priority. The survey of 2,140 adults was conducted September 24–28 and published October 8. Its overall sampling margin is plus or minus 2.9 percentage points. This measures public attitudes, not AI capability or support for a particular ban: the broad priorities were rated separately, without forcing trade-offs.
US research picks AI projects for fusion and chip design
The US Energy Department announced twelve Phase-II Genesis Mission selections in a $159 million funding announcement. Planned work includes a digital twin for the SPARC fusion facility and chip design for extreme environments. These are concrete attempts to connect AI with scientific equipment, not completed discoveries. The official project list, posted October 8, says selection for award negotiations does not guarantee an award or funding. Neither the announcement nor the list establishes that the money has been paid or that the systems work.
A folding robot checks its mistakes before moving on
Glasgow researchers report that FoldBack can detect a failed grasp, roll back and repair the next part of a robot's folding plan. Across 105 controlled trials per method using 33 garments, final folding success rose from 45.7% for the strongest prior baseline to 75.2%. The gain comes from recovering before one bad move ruins the rest. Category-specific trained policies and mostly flattened starting garments make this a laboratory result, not a robot ready for a household laundry pile.
Eight RTX 5090s run AI faster with fewer data transfers
CoMoE changes how a group of consumer GPUs exchanges data while running mixture-of-experts AI models. Using CPU memory as a coordinated distribution point, the authors report 1.24 to 1.46 times the output throughput of the standard software backend on the same eight-card RTX 5090 system. The four-model test exposes a software bottleneck, not just a shortage of chip power. This was a dual-Xeon server with 1.5 TB of RAM, not an ordinary home PC or a full system-cost comparison.
The poll and DOE selection were published October 8; polling took place September 24–28. The two preprints were first announced by arXiv October 8 after submission October 7; their experiments predate publication.
Afternoon update: the poll and project selections replace the optical-chip and magnet-coil items. The morning edition remains in the archive.
8 October 2026 · Morning edition
Published on · Four signals worth your attention.
A folding robot checks its mistakes before moving on
Glasgow researchers report that FoldBack can detect a failed grasp, roll back and repair the next part of a robot's folding plan. Across 105 controlled trials per method using 33 garments, final folding success rose from 45.7% for the strongest prior baseline to 75.2%. The gain comes from recovering before one bad move ruins the rest. Category-specific trained policies and mostly flattened starting garments make this a laboratory result, not a robot ready for a household laundry pile.
Light patterns reprogram an optical chip
A Cornell and NTT team used projected light patterns to reconfigure one lithium-niobate device for different optical jobs, including routing and separating light signals. The laboratory demonstrations included eight-channel routing on the same fabricated platform. This shifts some design work from making another device to programming the one already built. It is not a production-ready AI chip: the setup still uses external optics, and the study does not establish a cost or energy advantage over electronic computing.
Eight RTX 5090s run AI faster with fewer data transfers
CoMoE changes how a group of consumer GPUs exchanges data while running mixture-of-experts AI models. Using CPU memory as a coordinated distribution point, the authors report 1.24 to 1.46 times the output throughput of the standard software backend on the same eight-card RTX 5090 system. The four-model test exposes a software bottleneck, not just a shortage of chip power. This was a dual-Xeon server with 1.5 TB of RAM, not an ordinary home PC or a full system-cost comparison.
Superconducting magnet coils reveal damage between layers
Researchers at the US National High Magnetic Field Laboratory found that layers inside superconducting REBCO tape separated during prototype-coil tests. Microscopy revealed blister-like damage; a short-sample experiment supports electromagnetic stress as a cause. The work reports tests from 2025 and 2026, not a new field-strength record or a fusion demonstration. It makes a practical constraint visible: a conductor can carry enormous current yet remain mechanically vulnerable. The proposed ways to prevent that damage still need validation.
All four preprints were first announced by arXiv on October 8, following submission on October 7. The experiments predate publication.
7 October 2026 · Afternoon edition
Updated on · Afternoon edition. Four signals worth your attention.
Chemistry Nobel honors reactions that favor one molecular mirror image.
Henri B. Kagan and Kenso Soai were named winners of the 2026 chemistry Nobel for discoveries showing how reactions can amplify a preference for one of a molecule's two mirror-image forms. That control matters when chemists need a particular form, including in pharmaceutical synthesis. Kagan uncovered nonlinear effects; Soai developed reactions in which the product helps catalyze more of itself. Today's event is the award: key experiments date to 1986, 1995, and 2003. The work demonstrates a chemical principle, not how life's molecular handedness actually began.
A small walking robot borrows a smartphone's senses.
UCLA researchers describe PhoneBot, a small two-legged robot that uses an Android phone for sensing, vision, and movement control. They report testing the app on four phones, including a 2017 handset. Reusing that hardware could lower the barrier to classroom robotics: the authors estimate the robot hardware at about $400, excluding the phone. This is a research platform, not a household helper. It has no arms, its walking policies were trained on flat terrain, and its conversational system still uses a separate laptop backend.
Personalized phone screens can trip up AI agents.
A new study tests six mobile AI agents on app interfaces populated for ten different user profiles. The same underlying task can lead to different outcomes as personal content changes; agents often select a competing item instead of the intended one. Training on those differences improves results on unseen profiles. The practical lesson is that success on a clean demonstration screen need not transfer to your messages or shopping history. These are controlled, generated app environments, not evidence from a rollout to real users.
AI servers rebalance work without adding GPUs.
Researchers introduce FluidPD, which moves spare computing capacity between reading prompts and generating answers as demand changes. In tests on eight A100 GPUs using workloads derived from Azure traces, the system improves how often requests meet response-time targets compared with fixed allocations. The useful distinction is between needing more hardware and having existing hardware assigned to the wrong job. This is a controlled research evaluation, not a live Azure deployment or a universal speedup; it cannot rebalance away a shortage affecting both stages.
Afternoon update: the chemistry Nobel announcement replaces the quantum-chemistry item. The morning edition is preserved in the archive.
7 October 2026 · Morning edition
A small walking robot borrows a smartphone's senses.
UCLA researchers describe PhoneBot, a small two-legged robot that uses an Android phone for sensing, vision, and movement control. They report testing the app on four phones, including a 2017 handset. Reusing that hardware could lower the barrier to classroom robotics: the authors estimate the robot hardware at about $400, excluding the phone. This is a research platform, not a household helper. It has no arms, its walking policies were trained on flat terrain, and its conversational system still uses a separate laptop backend.
Personalized phone screens can trip up AI agents.
A new study tests six mobile AI agents on app interfaces populated for ten different user profiles. The same underlying task can lead to different outcomes as personal content changes; agents often select a competing item instead of the intended one. Training on those differences improves results on unseen profiles. The practical lesson is that success on a clean demonstration screen need not transfer to your messages or shopping history. These are controlled, generated app environments, not evidence from a rollout to real users.
AI servers rebalance work without adding GPUs.
Researchers introduce FluidPD, which moves spare computing capacity between reading prompts and generating answers as demand changes. In tests on eight A100 GPUs using workloads derived from Azure traces, the system improves how often requests meet response-time targets compared with fixed allocations. The useful distinction is between needing more hardware and having existing hardware assigned to the wrong job. This is a controlled research evaluation, not a live Azure deployment or a universal speedup; it cannot rebalance away a shortage affecting both stages.
Three quantum processors share a tiny chemistry calculation.
A team using hardware at Oak Ridge reports dividing small helium-cluster calculations across three diamond-based quantum processors. They split repeated measurements and separate fragments of the calculation between the devices, testing how much useful parallelism survives coordination overhead. It is an early demonstration of making several small processors work together, not a single larger quantum computer. The circuits use just one or two qubits, and the authors explicitly claim no advantage over classical chemistry methods. Larger, noisier workloads remain a different test.
6 October 2026
A coding study trains agents to reject bad fixes.
A newly announced preprint trains coding agents to write tests that distinguish working repairs from deliberately faulty alternatives. On 270 SWE-bench Verified issues excluded from training, the authors report first-attempt resolution rising from 31.9% to 43.0%. The point is not simply to make more attempts, but to improve how an agent checks its work. Those issues had already informed method development, however, so this is not a clean test on wholly unseen problems or proof of dependable production software.
A robot dog adjusts its gait to make less noise.
Researchers describe TACET, a system that changes a quadruped robot's gait and route according to the people and activity around it. Their preprint reports quieter walking on a Unitree Go2 and tests behavior near people who are working, talking, or walking. It treats noise as part of navigating shared space, not just avoiding collisions. The largest reported reduction was measured by a microphone mounted on the robot, not at a listener's ear; selected test scenarios do not establish everyday performance.
PyTorch adds an overlapping compute path for AMD GPUs.
PyTorch's development branch gained a native ROCm implementation that overlaps gathering data from other GPUs with matrix multiplication, rather than waiting for the whole transfer first. The change supports AMD's MI300X and MI355X architectures and remains opt-in. It extends a technique already available on NVIDIA's CUDA path to another hardware stack. This is a code integration, not a stable PyTorch release or a universal speed boost: the author's measurements show that some workload shapes remain slower than the fallback.
Researchers target bad robot habits with local retraining.
A new preprint introduces REDIRECT. Using demonstrations labeled successful or problematic, it locates a troublesome part of a robot task, retrieves a better continuation, and retrains the policy while rehearsing behavior worth keeping. The authors report improvements in simulated manipulation and small physical tests involving cups and towels. The useful idea is selective correction rather than relearning the whole task. The headline '1%' refers to a defined optimization budget in simulation, not a measured 99% saving in total cost or a guarantee that other skills stay intact.
5 October 2026
Dubai launches an AI-agent program for public services.
Dubai Future Foundation launched an accelerator to develop government services using AI agents. Its first phase will assess more than 300 public-facing use cases across 33 government entities. The program moves the question beyond chatbots toward systems that can carry out administrative tasks. Those numbers describe the scope of an assessment, not 300 deployed services or demonstrated improvements in speed and reliability.
Panasonic plans to start making humanoid robots by 2029.
Panasonic Holdings plans to develop its own humanoid robots, chief AI officer Akira Sakakibara told Nikkei in an interview published today. The company aims to begin making the machines by 2029, targeting factories and warehouses. That adds an established electronics and battery manufacturer to the field, but the statement is a development plan, not an available robot or proof of reliable factory work.
GPT-Rosalind reaches its API billing start date.
OpenAI lists October 5 as the billing start date for GPT-Rosalind's research API. Published rates per million tokens are $5 for input, $0.50 for cached input and $25 for output. Access remains limited to approved internal research through its trusted-access program. This is a scheduled cost change for eligible research teams, not a new model launch; the API's availability was announced in September.
Axiom orders solar wings for its second station module.
Redwire announced a follow-on contract to develop and deliver two roll-out solar-array wings for Axiom Space's second station module. The company is already building arrays for the first module. Power hardware is a concrete step toward the planned commercial station, but the launch schedule remains ahead: Axiom targets 2028 for its first module and less than a year later for the second.
5 October 2026 · Earlier edition
EV sales fell. Charging demand still grew.
U.S. light-duty electric vehicles used an estimated 8% more electricity in the first half of 2026 than in the preceding six months, even as new EV sales fell 19%. The EIA figures include battery-electric cars and plug-in hybrids. The distinction matters: fewer new sales do not mean the fleet already on the road needs less power. These are modeled national estimates, not direct measurements of every charger or local peak demand.
AI proposes the note. Human ratings determine its reach.
A new paper from the team behind X’s Community Notes describes how AI writes proposed notes while contributor ratings feed the algorithm that determines which ones appear broadly. The team says its Community Writer produced 52% of notes rated Helpful during its June measurement window. That is a reported share of selected notes, not an accuracy score. The operator-authored preprint does not independently establish that AI fact-checking is better than human work.
3 October 2026
Kolibri Opens a German-English AI Model to Local Deployment
Aleph Alpha released Kolibri with downloadable model weights under Apache 2.0. Organizations can run the German-English model on their own infrastructure, but its roughly 78-gigabyte memory footprint still requires substantial computing hardware.
RNA Therapy Raises a Muscle Protein in Three Boys
Ractigen reported that its experimental RNA therapy raised the muscle-protecting protein utrophin 3.5–5.3-fold in three boys with Duchenne muscular dystrophy. The small, uncontrolled study supports the mechanism, but does not establish clinical benefit.
A Blood-Cancer Drug Moves Into First-Line Treatment
The FDA expanded Jaypirca’s approval to adults with previously untreated chronic lymphocytic leukemia or small lymphocytic lymphoma who lack a known chromosome 17p deletion. The decision adds an initial treatment option for this defined group.
An AI-Designed Hair-Loss Drug Advances Toward Human Testing
XtalPi says its hair-loss drug program has entered safety and manufacturing studies needed before seeking permission for human trials. Encouraging mouse results support further development, but do not show that the treatment works in people.
2 October 2026
One Charging Trust System Reaches Six Car Brands
AUTOCRYPT says it built a shared authentication system for EV charging in South Korea and validated it with vehicles from six brands. Registered cars can identify themselves at compatible chargers; payment and settlement services remain a future possibility.
A Heart-Procedure Debris Catcher Enters U.S. Use
Emboline says doctors used its FDA-cleared Emboliner in two U.S. valve-replacement procedures, beginning a controlled commercial launch. The device is designed to catch debris released during these procedures; the announcement provides no new comparative outcome data.
Oracle Offers to Absorb Rising Nuclear Power Costs
Oracle plans to take on part of Wisconsin’s Point Beach nuclear electricity costs, projecting about $300 million in savings for utility customers. The arrangement links its data-center expansion to household bills, but still needs state regulatory approval.
Hospital Coding Gets an Autonomous AI Pitch
AKASA launched AI designed to code complex inpatient cases without human intervention, claiming blinded tests matched or beat human coders. The announcement does not publish detailed accuracy results, so the performance evidence remains vendor-reported.
30 September 2026
A robot learns targeted fixes from human corrections.
Microsoft researchers introduced Rho, open robot-control models tuned for three dual-arm platforms. In their preprint, 15 human-corrected practice runs improved two already-trained tasks, suggesting robots can learn targeted fixes without retraining their full control model.
AI agents benefit from thinking about what to try next.
Meta researchers report that giving AI agents a separate controller to plan their next steps improved coding and reasoning benchmarks. Their preprint finds a trade-off: the extra deliberation can hurt performance when computing budgets are small.
Cloudflare plans a new issuer for quantum-resistant web certificates.
Cloudflare plans a public certificate authority to verify website identities, with quantum-resistant certificates targeted for early 2027. Browser approvals are still pending; the announcement lays out a migration route rather than a service websites can use today.
Battery backup expands to emergency dispatch centers.
Evoltix says two new battery-backup units are now available for emergency dispatch centers and small data centers. They support three-phase power; outage coverage depends on the battery capacity and critical load configured for each site.
21 September 2026
Grok 4.7 is built to work longer before it stops.
SpaceXAI released Grok 4.7 through its API, Cursor and Grok Build, using a larger base model and longer reinforcement-learning run. Its benchmark and safety gains are company-reported; independent testing must show how reliably they transfer to real work.
Google backs 96 more megawatts from Georgia’s nuclear fleet.
Georgia Power and Google proposed funding uprates at two nuclear plants to add about 96 megawatts. The utility projects $900 million in customer benefits, but regulators must approve the tariff, Hatch upgrades and agreement before those gains exist.
An ovarian-cancer blood test wins permission for its pivotal study.
The FDA approved Mercy BioAnalytics’ investigational-device exemption for a pivotal study of its blood-based ovarian-cancer screening test, pending ethics review. That permits research in average-risk women; it does not approve the test for screening.
AI data centers get a qualified battery building block.
LG Energy Solution says its 2.5-megawatt battery blocks meet NVIDIA’s DSX requirements for AI data centers, including load smoothing and voltage ride-through. Qualification identifies a compatible power component; it does not demonstrate performance in an operating facility.
20 September 2026 · Launch edition
Memory chips tackle a precision problem.
Researchers demonstrated a memory-chip method for simplifying complex datasets while limiting accumulating calculation errors. Small hardware tests matched software accuracy; larger applications relied on simulations, leaving real-world speed and energy savings to be established.
Rapid heart ablation gets a one-year follow-up.
Field Medical reported one-year results from its first human trial of rapid electrical-pulse heart ablation. With the optimized pulse setting, 80% remained free of atrial arrhythmias. The investigational device still needs broader clinical evaluation.
A wrist-worn heart treatment returns mixed early results.
A 77-patient pilot tested wrist-worn nerve stimulation after heart ablation. StimCardio reported fewer recurrent arrhythmia events, but the difference in patients remaining arrhythmia-free was not statistically significant; larger trials must clarify benefit.
An inflammation drug enters a larger colitis trial.
Accro Bioscience says the first patient has received AC-101 in a placebo-controlled ulcerative-colitis trial. The study will test whether blocking an inflammatory signaling pathway improves remission after 12 weeks; efficacy remains unproven.
16 September 2026 · Evening edition
OpenAI tests conversations with advertisers’ agents.
OpenAI announced a U.S. test of clearly labeled sponsored agents that users can choose after seeing an ad. The separate chats would let advertisers answer product questions, extending advertising beyond a static placement.
Europe’s joint AI project moves toward state-aid review.
The European Commission welcomed the design of an AI project involving 19 member states. Participating countries are preparing state-aid submissions for shared capabilities and services; this step is not final funding approval.
A $1 billion quantum-foundry award is finalized.
IBM’s Anderon says it finalized a $1 billion CHIPS award with the U.S. Commerce Department, following May’s preliminary agreement. The funding supports quantum-wafer manufacturing research, expanding infrastructure rather than demonstrating useful quantum-computing advantage.
Epson expands robotic printing beyond flat surfaces.
Epson launched CONTOURJET with upgraded robots and support for larger workpieces, combining six-axis motion with inkjet printing. The company targets complex three-dimensional surfaces and short production runs; efficiency claims remain manufacturer-reported.
16 September 2026 · Morning edition
A new coalition pushes for human control of AI.
More than 50 organizations launched the Pro-Human AI Coalition, co-founded by Project Liberty, the Future of Life Institute and Ergo Impact. The nonprofit will coordinate policy, research and technology efforts around human control of AI.
Georgia adds a four-hour battery to its grid.
Georgia Power says its 128-megawatt Robins battery facility has entered commercial operation. Paired with nearby solar generation, it can supply stored electricity for up to four hours, shifting energy toward periods of higher demand.
Hybrid wheat gets a late-2027 launch target.
Corteva confirmed a late-2027 launch for hybrid wheat made with Vylor’s Xpedite system. It reports higher yields in company trials, including under water stress; commercial performance across farms remains to be demonstrated.
$7.5 million backs the route from research to treatment.
Royalty Pharma announced a $7.5 million donation to Blood Cancer United’s Therapy Acceleration Program. The funding backs early-stage biotech investments and development support, with investment returns recycled into future blood-cancer therapies.
14 September 2026 · Evening edition
Perplexity’s local agent reaches Windows.
NVIDIA says Portable Computer now runs on compatible Windows RTX systems with at least 24GB of video memory. Local work avoids cloud credits; more demanding tasks can request permission to use cloud models.
NVIDIA ·
NIH funds models of sex-specific hormone biology.
NIH is awarding $21 million across six projects to model hormonal biology and its effects on drug response and toxicity. The initiative supports research into sex differences; its models still need development and validation.
NIH ·
Photonic proposes a shared Canadian chip facility.
Photonic proposed a multi-tenant semiconductor facility costing up to C$500 million for quantum, AI and other advanced technologies. Its inclusion in Canada’s investment-summit prospectus presents an investment opportunity; funding and construction are not established.
Cornelis puts computation inside the network.
Cornelis announced a programmable networking architecture intended to reduce time AI accelerators spend waiting for data, alongside $205 million in funding. Broader CN6000 availability is expected in the fourth quarter; performance claims remain vendor-reported.
14 September 2026 · Earlier edition
A bone marrow chip keeps immune cells in view.
Researchers combined a lymphoid organoid with a bone marrow chip to track how human antibody-producing cells mature and survive. The laboratory platform could improve studies of vaccines, allergies and autoimmunity; it is not a treatment.
NIH ·
NASA opens its lunar AI model to researchers.
NASA and IBM released an open AI model trained on lunar observations. Researchers can adapt it to map craters, identify volcanic features and estimate where polar ice may persist. Those estimates still require scientific validation.
NASA ·
A Swiss quantum hub opens before its local machine arrives.
IBM and Lockheed Martin launched a quantum research hub with ETH Zurich. A Nighthawk-powered Quantum System Two is planned for Lugano by year-end; cloud access starts sooner. The announcement expands research infrastructure, without establishing a new computational advantage.
IBM ·
SolarEdge puts 800-volt AI power to the test.
SolarEdge says its medium-voltage-to-800-volt DC converter is running under load in its labs. It also published a protection framework with NVIDIA. Full-system validation continues, and the AI data-center products are not yet generally available.
9 September 2026
Nine billion DNA changes get a searchable map.
DeepMind released AlphaGenome Atlas, a free academic portal predicting the molecular effects of nine billion possible DNA letter changes. It helps researchers prioritize experiments; the predictions are not approved for clinical use.
An Iowa reactor gets financing. Permission comes next.
The U.S. Energy Department closed a loan of up to $1.9 billion for Duane Arnold’s restart. The 615-megawatt Iowa reactor still needs licensing approvals before it can supply electricity again.
IonQ takes orders for its next quantum computer.
IonQ says it has fabricated Superion 256 chips and trapped the first ions in prototype systems. Orders are open, with customer deliveries planned for 2027; large-scale, fault-tolerant operation remains a roadmap goal.
A cell’s past may be hiding in its present.
Researchers trained IRIS on stem-cell experiments, then inferred signaling histories in data from mouse embryos. It helped improve a protocol for producing cells involved in lung development; broader transfer across cell types still needs testing.
8 September 2026
Intel’s new chipmaking tools cross a million wafers.
Intel and ASML say their next-generation chipmaking tools have processed over a million wafers. That includes testing and research as well as selected production layers—a manufacturing milestone, not a million finished chips.
An AI-designed lung drug moves six aging clocks.
A new analysis of 42 trial participants found lower predicted biological age with rentosertib. The researchers cannot yet separate effects on aging from effects on lung disease.
Figure’s robot brains get a $3.5bn compute commitment.
Figure and Nscale plan capacity of up to 100,000 Nvidia GPUs, starting in the second half of 2027. The commitment reveals the infrastructure bill behind humanoids; better robots still have to follow.
Figure ·
Jülich switches on JION. Useful quantum work comes next.
The trapped-ion computer is running, with access through JUNIQ planned for researchers and companies. The next test is combining it with supercomputers to tackle practical problems.
7 September 2026
Sanders draws a line around intelligence.
The announced proposal reaches human-level capabilities too. The definition—and how it would be measured—deserves close attention.
Charlotte puts the data-center argument in public.
Five more September feedback sessions are scheduled before the city’s November 5 deadline. What rules replace the pause is the next real test.
City of Charlotte · Checked
Ohio wants AI’s power risk to stay with AI.
The regulator wants data centers to fund a proposed capacity backstop. Its requested manufacturing exemption would draw another politically consequential line.
A striking methane map still needs a careful reading.
Google’s new announcement builds on older observations. Our explainer shows why a confidence label changes what you can conclude.