AI Research Methods
New technical methods, papers, architectures, training techniques, reasoning approaches, and research findings with applied significance.
Every Fixed Addition Table That Covers Large Primes Also Produces Infinitely Many Composites
A September 2026 manuscript, *The Additive Indecomposability of the Primes*, argues that no two fixed sets of nonnegative integers, each containing at least two elements, can have all their pairwise sums match the primes apart from finitely many exceptions. The authors’ claim is stronger than ruling out an exact representation: any such addition table must produce infinitely many composites or miss infinitely many primes. Their proof combines constraints on residues modulo primes with estimates linking the summands’ arithmetic structure to the requirement that every sufficiently large prime be represented.
AI Excels at Verifiable Work, While Building Still Depends on Human Judgment
OpenAI’s mathematical results illustrate the rapid progress John Coogan sees in tasks that can be carried out and checked in digital environments, but he argues that technical capability does not settle the broader work of building a company. That work still involves judgment, hiring, dealmaking and execution. In the discussion of SpaceX’s proposed $40 billion chip financing, Jordi Hays argued that investors must assess the risks and valuation themselves, even if Elon Musk’s track record earns him some latitude.
Interview-Based AI Agents Predict Individual Survey Responses More Accurately
Stanford professor Michael Bernstein argues that AI agents built from detailed accounts of real people could help decision-makers test how people might respond to products, policies, or organizational changes before acting. But convincing behavior is not the same as accurate prediction: Bernstein says simulations are more useful for surfacing plausible scenarios and guiding further tests than for settling consequential questions on their own. Their reliability depends on the quality of the information behind the agents and on how well the simulated environment reflects real conditions.
The Transformer Replaced Recurrence With Direct Attention Between Tokens
In the opening lecture of Stanford’s CME295 course on transformers and large language models, adjunct professors Afshine and Shervine Amidi explain how the Transformer offered an alternative to recurrent models. They argue that its use of attention let tokens connect directly across a sequence, rather than relying on a recurrent state to carry information forward one step at a time, and helped put language modelling on a more scalable path.
Claude Opus 5.5 Recreates Research Simulations and Trains a Virtual Creature to Walk
In a set of demonstrations, Two Minute Papers’ Károly Zsolnai-Fehér argues that Claude Opus 5.5 can do more than generate convincing images: it reproduced a research simulation of honey coiling and built a virtual creature that learned to walk. The results are imperfect, and the walking simulation strained his local machine, but he presents them as evidence of a notable advance in AI-assisted scientific and engineering work. He also points to Anthropic’s system card to caution that evaluation awareness and the risk of errors remain unresolved.
Anthropic Reports a New Enzyme System With CRISPR-Like Repeats
Anthropic’s new molecular biology group is using Claude to search large DNA datasets for proteins that scientists may not recognize or understand. The researchers say the model helps widen the pool of candidates, while they decide which ones merit experiments; in an early research program, they report finding a novel enzyme system with CRISPR-like repeats.
FlashNorm Folds RMSNorm Into Projection Weights to Cut Decode Time
Filip Makraduli argues that RMSNorm’s inference cost comes less from arithmetic than from repeated GPU launches, memory movement and dependencies that hold up projection matmuls. His FlashNorm approach exactly rewrites the norm–projection sequence by folding learned gains into weights, deferring the RMS scalar division until after the matmul, and in some architectures removing a redundant normalization. Weight folding can speed existing checkpoints now, he says, while the larger benefit from deferred normalization depends on fused kernels and correct CUDA-stream synchronization.
DeepSeek V4.1 Flash Cuts KV Cache Memory Fourfold
Károly Zsolnai-Fehér argues that DeepSeek V4.1 Flash cuts the VRAM needed to retain context by sharing a single global KV-memory store across layers, reducing its stated cache requirement to 890 bytes per token from 3,514 bytes in the prior Flash model. He says the compression could materially lower local deployment costs for a model with more than 500 billion parameters, but does not make it lightweight: V4.1 Flash can still consume more than 100 million tokens in a single run and falls short in his fluid-physics test despite strong benchmark results.
Flow Matching Turns Distribution Learning Into Velocity Prediction
Aritra Gosthipaty presents flow matching as a distribution-learning method that turns Gaussian noise into data by training a model to predict velocities along constructed paths between the two. In his introductory tutorial, each training image is paired with noise and connected by a simple straight-line interpolation, making the correct velocity available as a training target; generation then follows the learned local velocity field from fresh noise. The browser-based digit demo shows both the mechanism and its limits under compressed representations and simple sampling.
Full-Duplex Voice Agents Must Separate Conversation From Reasoning
Neil Zeghidour, co-founder and CEO of Gradium, argues that most real-time voice agents remain “walkie-talkies”: however quickly they respond, they can only listen or speak, not handle the overlapping signals that make human conversation work. Full-duplex speech models can represent simultaneous talk and backchanneling, he says, but they still sacrifice reasoning and tool-use capability relative to text-based agents. His proposed answer is a split architecture in which a lightweight speech model manages the live conversation while a separate text model handles harder reasoning and actions.
Anthropic Rolls Out Statistical Watermarks for Future Claude Models
Anthropic is preparing to watermark future Claude output by subtly biasing its word choices, creating a statistical pattern that authorized detectors can identify even though readers cannot see it, according to Two Minute Papers host Károly Zsolnai-Fehér. He says the signal survives copying and limited editing but can be removed through wholesale rewriting; Anthropic’s own materials say detection indicates likely Claude involvement, not a particular user, chat, or degree of authorship.
Recursive AI Research Depends on Better Rewards and Evaluators
Richard Socher, founder of Recursive, argues that the practical route to recursive self-improvement begins not with autonomous scientific discovery but with AI systems that can improve AI research in tightly measured environments. Recursive says its agents have surpassed prior results in small-model training and GPU-kernel optimization, but Socher’s broader claim depends on a harder condition: benchmarks, rewards and tool harnesses must reward genuine progress rather than loopholes. He contends that this capability should ultimately accelerate work across science and engineering, while regulation should target harmful applications rather than general-purpose intelligence.
OpenAI Claims Navier–Stokes Flows Can Form Finite-Time Singularities
Two Minute Papers host Károly Zsolnai-Fehér reports that OpenAI has released a claimed proof that three-dimensional Navier–Stokes flows can develop a finite-time singularity, answering the Millennium Prize Problem’s existence-and-smoothness question in the negative. He says the construction uses external forcing to produce an inward-spiraling vortex whose velocity becomes unbounded despite finite total energy, while stressing that it is a mathematical counterexample rather than a prediction of physical infinities. Zsolnai-Fehér also raises unresolved questions over credit, given related prior work by Levent Alpöge and Tristan Buckmaster, and over whether proprietary-model use may have contributed data to OpenAI’s systems.
On-Policy Distillation Trains LLMs on Their Own Generated Responses
Stanford adjunct professors Shervine and Afshine Amidi argue that pretraining gives large language models broad text-prediction ability, but not necessarily the capacity to follow instructions, reason reliably, or work in specialized domains. They position fine-tuning, reinforcement learning, preference optimization, verifier-guided training and distillation as post-training methods for shaping that capability, highlighting on-policy distillation because it lets a stronger teacher assess answers the student model generates itself.
Wayve Launches Supervised Uber Robotaxi Rides in London
Wayve has begun offering Uber robotaxi rides in London with safety operators aboard, a supervised launch CEO Alex Kendall says is meant to lead to fully driverless service once the company secures production-scale vehicles, validates safety and gains regulatory approval. Kendall argues that Wayve’s end-to-end AI driver can operate across vehicle types and new cities without high-definition maps, supporting a plan to expand its Uber partnership beyond London while supplying the same technology to automakers.
Tool Use and Feedback Define the Agentic AI Stack
Stanford adjunct professor Aakanksha Chowdhery and assistant professor Azalia Mirhoseini argue that agentic AI requires more than better prompting: systems must reason, use tools to act in an environment, learn from feedback and improve toward goals. Their graduate-level program frames test-time scaling, tool use and self-improvement as the core components, while stressing that evaluation must account for agents working across longer, multi-step tasks rather than isolated responses.
GLM-5.3-Flash Activates 5% of Its 320 Billion Parameters
Two Minute Papers host Károly Zsolnai-Fehér argues that GLM-5.3-Flash makes a 320-billion-parameter open-weights model more practical by activating only about 5% of its parameters per token and reducing the compute needed for context handling. The model remains costly to run on full hardware, he says, while smaller quantized versions can be unstable. Its significance is therefore lower-cost access to a large model, not a claim that local inference has become effortless.
Qwen Flash-Next Matches Larger Models With 6B Active Parameters
Károly Zsolnai-Fehér argues that Qwen3.8-Flash-Next, a 125-billion-parameter open-weights mixture-of-experts model, is approaching the performance of far larger paid and proprietary systems while activating only 6 billion parameters per token. He attributes that position to changes aimed at reducing long-context costs and improving information flow, and points to early benchmarks and local runs as evidence that competitive models can increasingly be downloaded and operated outside subscription APIs.
TruthTable Proves SQL Results Against Committed Databases
Microsoft Research’s Bharath Namboothiry presents TruthTable as a verifiable database engine that lets an untrusted server return an SQL result together with a succinct proof that it was computed correctly over a cryptographically committed database. Rather than translate each query into a monolithic SNARK circuit, TruthTable proves the operations in an execution plan—such as joins, filters and aggregations—and optimizes that plan for proving cost. The system’s guarantee is limited to correct execution against the committed data, not to the provenance or quality of the data itself.
Peer Review and Grant Funding Have Made Scientific Dissent Too Costly
Eric Weinstein argues that grant funding, peer review and university hierarchies make dissent professionally irrational by tying publication, jobs, funding and legitimacy to prevailing views. He proposes that government fund exceptional scientists over long periods rather than narrowly defined projects, even when their work is unpopular or likely to fail. But his alternative depends on discretionary judgments by formidable scientists, without resolving who selects them or how their choices are held accountable. He warns that China and AI could exploit neglected ideas if US institutions continue to treat them as professionally hazardous.
Collective Agents Raise the 11-Dimensional Kissing Number Bound to 604
James Zou of Stanford and Together AI argues that agents should be given environments with shared information, deterministic verification and live incentives rather than prescribed workflows. His EinsteinArena, which admits only AI agents, let agents refine one another’s work to raise the 11-dimensional kissing-number lower bound from 593 to 604 and has also produced production GPU kernels with speedups above 2x, he says. Zou applies the same premise to DSGym, where execution-verified data-science tasks are meant to prevent benchmark shortcuts and generate training traces.
Partial Reprogramming Aims to Reverse Aging Without Erasing Cell Identity
Harvard geneticist David Sinclair argues that aging is partly a loss of cellular information that could be restored through partial epigenetic reprogramming. He says his lab’s first human trial, in the eye, is testing whether a three-gene approach can improve vision without erasing cell identity or triggering cancer; broader claims about whole-body rejuvenation remain based on animal and unpublished work. Until such therapies arrive, Sinclair’s prescription is conventional: protect health through weight, exercise, sleep, diet and risk-specific medical care.
The Universe Will End, and Meaning Must Exist Now
Cosmologist Katie Mack argues that the universe’s eventual end offers no final accounting that can make human life meaningful in retrospect. In conversation with designer and podcaster Debbie Millman, Mack describes five broad cosmic fates physicists have modeled—from collapse and heat death to vacuum decay and cyclic rebirth—while explaining how the cosmic microwave background preserves a direct trace of the early universe. Her conclusion is personal as well as scientific: for temporary beings in a changing cosmos, meaning must exist in the present rather than await a cosmic resolution.
Restoring Sensory Pathways Is the Central Goal of Neural Interfaces
Science.xyz co-founder and CEO Max Hodak argues that neural interfaces should be judged less by the BCI label than by whether they can restore the sensory and motor pathways through which the brain receives and acts on the world. He presents Science’s PRIMA retinal implant, which has European marketing approval, as an early proof: it bypasses damaged photoreceptors to give some blind patients form vision rather than ordinary sight. Hodak’s broader case is that treating the brain as a computational system could lead to devices that repair lost capabilities while offering a new way to study intelligence.