← Back to archive中文
NeuronX AI Daily

PSI says AI proved the Aaronson–Ambainis conjecture, with the final version verified in LeanGR00T 2 improves robot task success rates with just a few demonstrations, while companies redraw the boundaries for coding agents.

November 9, 2026 Monday Sources · Blog
About this issue: This brief is automatically compiled, grouped and rewritten from public sources (X / podcasts / blogs and newsletters). Every item links to its original source — please defer to the original; AI rewriting may contain errors, and corrections against the source are welcome.

In one paragraph

PSI says its research system, using a non-public Anthropic model, found a complete proof of the Aaronson–Ambainis conjecture, proposed in 2008, on its third day of operation. Both the initial proof and the final version after two refinements passed Lean verification. Across 50 tasks unseen by the target robot, NVIDIA's GR00T 2 increased its success rate from 40% to 95% after adding five demonstrations per task and fine-tuning for 30 minutes. Goldman Sachs changed access permissions in October, leaving most bankers and traders without access to Claude Code and GitHub. The restrictions focus on third-party agents that can write and run code autonomously.

🔬AI mathematical proofs move into formal verification

PSI's reported result sets limits on quantum query speedups for yes-or-no problems with random inputs.

BlogPSI says AI proved a longstanding quantum computing conjectureResearch

PSI says its research system, using a non-public Anthropic model, found a complete proof of the Aaronson–Ambainis conjecture, proposed in 2008, on its third day of operation. Both the initial proof and the final version after two subsequent refinements passed verification in Lean. The result limits quantum speedups: for yes-or-no problems with random inputs, quantum computers can achieve at most a polynomial, rather than exponential, advantage in the number of times they must query the input.

Read original →

🤖Robot skill transfer, data collection, and simulation tools

GR00T 2 reduces the need for demonstrations on the target robot, while Tactus and Isaac Lab add manipulation data and physics simulation tools, respectively.

BlogNVIDIA launches the GR00T 2 robotics modelLaunch

NVIDIA has launched GR00T 2, a world-action model that jointly predicts future video and robot actions, supporting grippers, dexterous hands, and humanoid robots. It can transfer skills demonstrated by humans or other robots to a target robot. Across 50 tasks unseen by the target robot, its zero-shot success rate was 40%. Adding five demonstrations per task, amounting to less than three minutes of robot data, and fine-tuning on a single RTX 5090 for 30 minutes raised the success rate to 95%.

Read original →
BlogIsaac Lab 3.0 integrates the Newton physics engineLaunch

Isaac Lab 3.0 has officially launched. Its unified task API supports switching between PhysX and Newton without rewriting the logic of supported tasks. Newton supports rigid bodies, deformable objects, and particles, and most workflows no longer require installing or launching Isaac Sim. Physics, rendering, and visualization backends can be selected independently, though the available combinations still depend on the specific task.

Read original →
BlogTactus glove captures touch and hand pose simultaneouslyResearch

Researchers have introduced Tactus, a sensing glove with a fingertip layer stack 1.5 millimeters thick. It combines capacitive tactile sensing with IMU-based sensor-fusion tracking to capture contact forces and hand poses during natural manipulation. Designed to address occlusion between hands and objects, it records training data for robotic manipulation while minimizing interference with grasping. In tests, its minimum detectable force was 0.59 newtons, and its fingertip keypoint error during manipulation was 11 millimeters.

Read original →

🔍 Analysis: GR00T 2's increase in success rate from 40% to 95% shows that, for this set of tasks, skill transfer still needs a small number of demonstrations from the target robot to close the gap. Tactus captures contact forces and hand poses despite occlusion, addressing information that is difficult to record in manipulation demonstrations. Isaac Lab 3.0, meanwhile, uses a unified task API to reduce the amount of rewriting needed to switch physics engines for supported tasks. For robotics developers, model transfer capabilities, demonstration data quality, and the ability to swap simulation tools need to be assessed together—not just zero-shot performance.

🛠️Companies set operating boundaries for coding agents

Goldman Sachs tightens employee access, while Monaco uses separate environments to host agents running in parallel.

BlogGoldman Sachs restricts bankers' and traders' access to Claude CodeGovernance

Goldman Sachs updated its internal guidance in October to limit Claude Code and GitHub to engineers, leaving most bankers and traders without access. They had previously used these tools to monitor client trading flows and build market models, practices the chief information officer had praised in July. The change followed a productivity and risk review. Strategists whose work spans trading, quantitative research, and software engineering retain access, and the bank-wide internal AI assistant is unaffected. The restrictions target third-party agents that can write and run code autonomously.

Read original →
BlogMonaco builds and launches an agent workspace in two weeksDeployment

Monaco built its own agent workspace, Monacoder, on the open-source Coder platform, with one cloud platform engineer completing it in about two weeks. Its previous setup, combining tmux and Git worktree, had not gained company-wide adoption. The new setup gives each agent a separate application environment and sanitized test data. Since launch, about half of individual contributors have used it weekly, with many running three to six workspaces simultaneously. The workspaces do not yet cover all of Monaco's functionality.

Read original →

🔍 Analysis: Goldman Sachs and Monaco are addressing the same question: once coding agents can execute tasks, how should companies limit their scope of operation? Goldman Sachs restricts access by role, while Monaco places agents in separate application environments with sanitized test data, allowing many users to run three to six workspaces in parallel. For companies, adoption rates should not be equated with the breadth of access permissions. Environment isolation can support more parallel use, but it cannot replace reviews of user identity and business permissions.

⚙️AI accelerates chip simulation, with speedups coming from combined changes

Samsung combines AI-based material property prediction, coarse meshes, and GPU acceleration in its analysis workflow.

BlogSamsung cuts AI-assisted simulation time from 18 hours to 41 secondsDeployment

Samsung is using NVIDIA PhysicsNeMo to develop a chip-level thermal and mechanical analysis workflow. In one AI-based homogenization case, simulation time fell from about 18 hours to 41 seconds, a speedup of roughly 1,600 times. The result combines AI predictions of equivalent material properties, a coarse mesh, and GPU acceleration. The baseline was a fine-mesh solver running on a 32-core CPU, while the comparison approach used a GPU-accelerated coarse-mesh solver.

Read original →

🔍 Analysis: Samsung's roughly 1,600-fold speedup compares a fine-mesh solver running on a 32-core CPU with a GPU-accelerated coarse-mesh approach. The changes go beyond adding AI. For engineering teams, the reduction from 18 hours to 41 seconds can be treated as the result of an end-to-end workflow overhaul, but not as the speedup contributed by the AI prediction module alone.

💰Company news: OpenAI funding and NVIDIA's independent investigation

OpenAI's new funding is arriving in tranches, while NVIDIA's board has launched a review overseen by independent directors.

BlogOpenAI receives $11 billion as funding arrives in tranchesFunding

As of November 6, 2026, OpenAI had received $11 billion in new funding in a round with a $1.25 trillion pre-money valuation, with SoftBank and MGX participating. The amount received represents about 37% of the original $30 billion fundraising target, with funds arriving in tranches. Subsequent portions have not yet closed; if completed, they could bring the round's total to $20 billion.

Read original →
BlogNVIDIA's board launches an independent investigation into reports concerning Jensen HuangGovernance

NVIDIA's board has announced an independent investigation in response to recent reports that founder and CEO Jensen Huang was targeted by a foreign intelligence operation. A committee of independent directors is overseeing the investigation, assisted by independent outside legal counsel, to review all available information. The board said it currently has no reason to believe national security or NVIDIA's confidential information has been compromised, and reaffirmed its confidence in Huang and his leadership team. Huang supports the review and has pledged to cooperate.

Read original →

🔑Key terms this issue

KEYWORD 01
Lean verification
Having a formal verification tool check a proof. PSI reports that both the initial proof and the final version after two refinements passed verification.
KEYWORD 02
Few-shot skill transfer
Transferring skills from existing demonstrations, then using a small number of target-robot demonstrations to improve task performance. GR00T 2 showed the effect of five demonstrations per task.
KEYWORD 03
Agent operating boundaries
Defining not only who can use a coding agent, but also which environment it runs in and what data it can access.
Worth watching (reference points, not predictions or advice)
NeuronX AI Daily
AI moves fast; you don't have to chase all of it. We read the primary sources and keep the few things that matter.
📮 Subscribe freeRSS