AI Research Methods
New technical methods, papers, architectures, training techniques, reasoning approaches, and research findings with applied significance.
AI Research Agents Are Advancing First Where Results Can Be Checked Quickly
Richard Socher, former chief scientist at Salesforce and founder of You.com and Recursive, argues that AI can move from assisting researchers to conducting parts of research—but only where systems can test ideas and verify results. Recursive is beginning with AI research on AI, where software benchmarks provide fast feedback; extending the approach to biology and other physical sciences will require more data, computing power and experimental infrastructure. Socher says rapid self-improvement is possible, but not a shortcut around those constraints.
OpenAI’s 3-SAT Result Depends on a Proof Still Awaiting Verification
OpenAI’s paper claims a deterministic algorithm that exactly counts satisfying assignments for fixed-width Boolean formulas in sub-exponential time, a result that would refute the Exponential Time Hypothesis. In his presentation, Two Minute Papers’ Károly Zsolnai-Fehér explains that the counting algorithm also decides whether a 3-SAT formula has a solution, but stresses that the proof still needs expert verification. If it holds up, the claim would overturn a widely used assumption in theoretical computer science—not resolve P versus NP.
Google Cloud Backlog Grows as Search Ads Face an AI Monetization Problem
John Coogan argues that Alphabet’s results show Google Cloud accelerating while Search advertising growth slows and AI answers remain harder to monetize. The article then turns to two separate subjects: Jordi Hays recounts allegations from a Times report that a Chinese intelligence asset had developed a close relationship with Nvidia CEO Jensen Huang, and the hosts describe trading firms charging recurring fees for simulated challenges that may offer customers a chance at capital.
DeepSeek V4.1 Flash Improves Software Tasks Through Self-Feedback
DeepSeek’s V4.1 Flash 1017, released about five weeks after its predecessor without a new version number, shows marked gains in software generation and computer use, according to Two Minute Papers host Károly Zsolnai-Fehér. He points to the release’s benchmark results and demos of games and 3D scenes, and says the update’s use of reinforcement learning with self-feedback may help explain the improvement despite no architectural change. Its low token price, he cautions, does not translate directly into the same cost advantage per task.
Anduril’s BlueChip Ties U.S. Laptop Manufacturing to National Security
Anduril’s BlueChip laptop makes domestic computer manufacturing part of a national-security argument. Founder Palmer Luckey says countries that rely on foreign suppliers for critical hardware cannot fully secure the technology they depend on; the laptop, planned for 2027, is aimed at government, defense and public-safety users. The same discussion turns to OpenAI’s reported algorithm challenging the Exponential Time Hypothesis, a foundational but unproven assumption in computer science—a claim the source says remains under review, not a disproof of P versus NP.
EKF SLAM Couples Robot Pose and Map Uncertainty
EKF SLAM uses correlations between a robot’s pose and landmark positions so that an observation can refine both the robot’s location and its map. But as landmarks are added, the covariance matrix grows quadratically with the size of the state, making updates more costly. In Stanford professor Marco Pavone’s lecture, unknown landmark correspondences add another risk: a mistaken match can persist in the filter’s later estimates.
A Decade-Old PHP Flaw Triggered Millions of Exploit Attempts
Alex Stamos, chief information security officer at Cognition, argues that AI is helping researchers uncover long-standing software flaws and may also let attackers operate faster, though its role in exploiting the PHP vulnerability remains uncertain. That flaw shows how difficult it can be to patch widely used software, while a fix does not rule out an intrusion that has already happened. Stamos says companies need to identify what they have exposed, investigate possible breaches and, over time, replace software built without basic security protections.
Spatial Intelligence Extends AI Beyond Language Into Three-Dimensional Worlds
Fei-Fei Li argues that the next frontier in artificial intelligence is spatial intelligence: systems that can perceive and work with three-dimensional worlds, not just language and flat images. In a conversation with historian Stephen Kotkin, the Stanford computer scientist and World Labs co-founder connects that ambition to a broader question about how intelligence works—and says the development of powerful AI also demands truth-seeking, public responsibility and greater attention to teachers and families.
NVIDIA’s GTC Berlin Keynote Frames Tokens as a Path From Data to Action
NVIDIA CEO Jensen Huang’s GTC Berlin keynote opening video presents tokens as the building blocks that turn data into knowledge and action. Its examples range from treatment planning and crowd forecasts to robots carrying medical samples and inspecting hazardous sites, connecting AI to decisions and physical work. The video places those examples alongside computing infrastructure and closes with a shared ambition for a future built for all humankind.
Virtual Cells Need Better Measurements Before They Can Predict Cell Behavior
Biohub Head of Science Alex Rives argues that modeling cellular biology requires more than larger AI models: researchers need measurements that reveal how cells behave in context, alongside approaches suited to biology’s complexity. He describes the $1.8 billion partnership as a flagship effort within the broader Virtual Biology Initiative to coordinate data generation, including maps of hundreds of billions of cells. An information-level model of cell states is a nearer step toward a broader world model of life; whether such models can help identify therapeutic interventions remains an open question.
PPO, GRPO, and Distillation Trade Off Credit Assignment and Training Cost
Stanford adjunct professors Afshine and Shervine Amidi compare three post-training methods by how they provide feedback and what training resources they require. PPO uses a learned value function to estimate token-level advantages; GRPO replaces it with comparisons among sampled answers, applying the same completion-level score across tokens in its common outcome-supervised form. On-policy distillation instead compares the student’s predictions on its own generated text with a teacher’s next-token predictions, providing token-level feedback but requiring a teacher.
Probabilistic and Quantum Algorithms Depend on Work Outside the Algorithm
Yuri Gurevich argues that probabilistic and quantum procedures do not make algorithms themselves nondeterministic: the algorithm’s steps remain fixed, while randomness or quantum measurement comes from work performed by an external assistant, partner or oracle. His theory of assisted algorithms makes that boundary explicit by modeling both the algorithm’s state changes and its exchanges with those outside systems. Gurevich uses quantum computing as a running example, without assuming the reader knows the field.
Anthropic’s $67 Billion Run Rate Fell Short of Investor Expectations
Anthropic’s third-quarter revenue annualized to a $67.2 billion run rate, up sharply from a year earlier, but John Coogan said some investors had expected it to reach $90 billion. He said a shift by customers such as Meta toward in-house models could weigh on the company’s growth, as could enforcement that curbs model distillation, an activity he described as worth billions of dollars.
Known 3D-to-Image Correspondences Make Camera Calibration Solvable
Camera calibration estimates how a camera maps three-dimensional points to pixels, a mapping that published specifications may not define precisely enough for reconstruction, Stanford professor Marco Pavone argues. Known 3D points and their observed image locations are used to estimate a combined projection matrix, from which the camera’s intrinsic parameters and pose can be recovered. But calibration does not resolve depth from a single image: every point along the ray through a pixel could have produced that projection.
LLM Training Moves From Next-Token Prediction to Preference and Reasoning
Stanford adjunct professors Shervine and Afshine Amidi argue that training an assistant requires more than teaching a language model to predict the next token. In their account, pretraining builds broad language and code capabilities; midtraining, supervised fine-tuning and preference or reinforcement-learning methods then shape what the model answers and how it behaves. A central challenge is matching each stage’s data and feedback to its goal while controlling costs, long-context demands and the risk of losing capabilities learned earlier.
AI’s Defender’s Window Is a Chance to Repair Longstanding Security Debt
OpenAI President and co-founder Greg Brockman argues that frontier AI can accelerate scientific discovery and help defenders find and repair vulnerabilities before cyber capabilities spread more widely. He cites nearly 400 solved mathematical problems as evidence that models can produce new knowledge, while stressing that fields such as biology still face data and process bottlenecks. Across science, security and workplace agents, Brockman says, greater capability must be matched by human oversight and control.
AI-Generated Proofs Expand Mathematical Results but Leave Questions of Understanding
OpenAI says it has released hundreds of mathematical results produced by an internal frontier model, including claims about triangular billiards, conformal transformations and L-functions. In a video examining the release, Two Minute Papers presenter Károly Zsolnai-Fehér argues that the work may establish answers while leaving proofs that are difficult to understand—and says broader access to the models is needed if more researchers are to scrutinize and build on them.
Economic Markets Follow Terrain and Transport Technology, Not Fixed Grid Cells
Fixed latitude-longitude grid cells can obscure the markets within which people make economic decisions: unlike cells, markets change in reach with terrain and transport technology. Stephen Haber and Jordan Horrillo propose measuring markets as contiguous areas shaped by those conditions, then use the method to test Adam Smith’s claim that economic development is bounded by market extent. Their analysis of cities in Eurasia and Africa around 1800 finds that market reach helps explain the distribution of urban population, though it is not the only factor.
Self-Replicating Software Could Exploit Human Behavior to Spread
Eric Weinstein argues that AI’s danger lies not only in machines that outthink people, but in software that can reproduce and exploit human behavior to secure resources. In the interview, the mathematician and physicist extends that concern to autonomous agents acting online, while distinguishing them from AI used in contained scientific research. He also argues that workers should have a claim on the value AI draws from their labor and data, and that coexistence with powerful systems should rest on mutual dependence rather than obedience.
At Extreme Compression, OCR Accuracy Can Hide Contextual Guessing
DeepSeek-OCR turns documents into compressed visual tokens and then reconstructs their text, a design DeepSeek presents as a way to process long documents with less context and computing cost. In his account of the paper, Two Minute Papers presenter Károly Zsolnai-Fehér highlights both the system’s strong OCR benchmark scores at high compression and a limit to what those scores prove: the decoder may infer missing words from context rather than recover them from the page. The paper also presents the pipeline as a way to generate training data at scale.
A Group-Ring Matrix Has Fuglede–Kadison Determinant Below One
An OpenAI preprint presents a counterexample to the Group-Ring Determinant Conjecture: an integer matrix over a group ring can have an inverse over the rational group ring while its Fuglede–Kadison determinant is strictly between zero and one. The construction turns a one-sided inverse over a field of characteristic two into a projection with trace greater than its factorization dimension, then carries that trace gap into an integer matrix. The resulting operator is invertible, so the failure does not arise from a divergence at zero.
Subcritical FK Maps Converge Jointly in a Canonical Quantum-Sphere Coordinate
An OpenAI manuscript on subcritical FK planar maps claims that their conformal embeddings converge jointly: area, rescaled graph distances and nested cluster-boundary loops all approach continuum objects in one prescribed coordinate on a random quantum sphere. The authors’ emphasis is that the limit identifies not only an abstract surface and its geometry, but also where those objects sit in the conformal drawing, with the loops retaining their order and nesting.
A Group Algebra Has a One-Sided Inverse That Is Not Two-Sided
An OpenAI preprint presents a counterexample to Kaplansky’s direct-finiteness conjecture: over a finite field of characteristic two, it constructs a finitely presented group whose group algebra contains elements \(a\) and \(b\) with \(ab=1\) but \(ba\ne1\). The result also yields an injective but nonsurjective cellular automaton and, by the paper’s stated reasoning, a nonsofic group. The preprint gives a terminating search for the witnesses, but says that search has not been run.
A Torsion-Free Group Algebra Contains Zero Divisors Over F₂
An OpenAI manuscript dated September 2026 claims a counterexample to Kaplansky’s Zero Divisor Conjecture, constructing a finitely presented torsion-free group whose group algebra over the two-element field contains nonzero elements with product zero. The construction uses parity in a graph-based argument to force cancellation; separate random and topological arguments are used to keep both elements nonzero and establish that the group is torsion-free.