Sector 4 · 4 sub-themes
Reasoning & Test-Time Compute
OpenAI's unreleased model resolving ten long-standing open problems in mathematics — spanning high-dimensional geometry, coding theory, and arithmetic circuits — is the sharpest signal yet that AI is crossing from assistance into genuine mathematical discovery. This follows the earlier AI-generated disproof of the Erdős unit-distance conjecture, which has already seeded further work in theoretical computer science. Fields Medalist James Maynard publicly described "soul searching" about his field's future. Anthropic's parallel effort is equally striking in method if not outcome: an unreleased Claude research model attempted the Riemann Hypothesis across roughly 650 distinct approaches, then spawned ~60 parallel instances and reasoned for around a day and a half before failing — though reportedly making progress on a related problem. The image of Claude expressing self-doubt mid-run became a viral meme, unusually broad cultural attention for a mathematical non-result. Together these two efforts define the current frontier of extended test-time compute: brute reasoning at scale applied to problems that have resisted human mathematicians for decades. Against that backdrop, SSI's reported pivot to test-time training — where model parameters actually update during inference rather than just compute longer — represents a genuinely different architectural bet, one eyewitness called results "radically better than anything I've seen before." And Pathway's 150M-parameter BDH-CQ quietly redraws the cost frontier, hitting 29.5% on ARC-AGI-1 at $0.0007 per task, roughly 11× cheaper than GPT-5 on the same benchmark.
OpenAIAnthropicClaudeSSIPathwayJames MaynardBDH-CQARC-AGI-1Erdős unit-distance conjectureRiemann HypothesisAI mathematical discoveryextended test-time computetest-time trainingopen problems in mathematicslarge language model reasoningcost-efficient AI inferencebrute-force reasoning at scaleAI and theoretical mathematicsparallel inference instancesarchitectural innovation
4.1
Anthropic's unreleased Claude research model attempts Riemann Hypothesis
- Anthropic tasked an unreleased research version of Claude with attempting the Riemann Hypothesis; the model tried approximately 650 distinct approaches, failed all of them, but reportedly made some progress on a related problem rather than solving the hypothesis itself. [1][2]
- Sources [3][4][5] describe the model's extended multi-agent effort: after exhausting ~650 attempts, Claude reportedly spawned roughly 60 parallel instances and continued reasoning for approximately a day and a half, illustrating an extended test-time compute approach to an unsolved mathematical problem.
- The model's self-aware commentary during repeated failures — characterized in social media reaction as the model expressing doubt before being encouraged to continue — became widely circulated as a meme across English and Japanese Twitter, reflecting broad public attention to the Claude research experiment. [3][4][5]
AnthropicClaudeRiemann Hypothesisunsolved mathematical problemsextended test-time computemulti-agent parallelizationAI research experimentationautonomous reasoningviral AI mememathematical problem-solving
4.2
OpenAI Unreleased Model Resolves Ten Long-Standing Mathematics Problems
- OpenAI revealed that an unreleased model, discovered while being evaluated on open research problems during development, produced results resolving or making substantial progress on ten long-standing open problems spanning high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography, and extremal combinatorics. [2][6]
- Prior to the ten-problem announcement, OpenAI shared in May an AI-generated disproof of the Erdős unit-distance conjecture — also discovered while evaluating an unreleased model — which has already inspired further developments in mathematics and theoretical computer science. [2][6]
- OpenAI's approach to applying AI in mathematics involves training models to learn patterns and connections from vast amounts of material, enabling them to combine known results, methods, and tools in new ways to attack unsolved problems — some of which had confounded academics for decades. [1][3][4][5]
- Oxford mathematician and Fields Medal winner James Maynard told The Verge he has been engaged in "soul searching" about the future of his field, reflecting the expert community's mixture of excitement and concern about AI's growing role in mathematical research. [1][3][4][5]
- Separately, OpenAI announced ChatGPT for Academic Researchers, an initiative providing 100,000 scientists and mathematicians with free access to its best ChatGPT models. [2][6]
OpenAIJames MaynardChatGPTOxfordThe VergeErdős unit-distance conjectureAI mathematical research breakthroughsopen problems solved by AIunreleased OpenAI modelErdős conjecture disproofmathematics and AI collaborationChatGPT for Academic ResearchersFields Medal mathematician reactionextremal combinatorics and geometryAI training on mathematical literature
4.3
SSI reportedly pursues Test-Time Training as distinct reasoning paradigm
- Viral posts on X claim Ilya Sutskever's Safe Superintelligence (SSI) is building a small test-time training (TTT) reasoning engine — described as a "paradigm shift" diverging from mainstream inference-time scaling and standard pretraining — and that SSI is transitioning from pure research toward scaling. [1][2][4][7][8][10]
- Test-Time Training (TTT) is a technique where a model adapts itself during inference by using input data to update its own parameters, rather than relying solely on pre-trained states; this approach is designed to handle distribution shifts between training and testing data, and has been applied to transformers and RNNs, with researchers at MIT and Stanford demonstrating notable performance improvements. [9]
- An author writing about TTT reported witnessing a demonstration of TTT combined with state-space models (TTT + SSM) and described the results as "radically better than anything I've seen before," while also noting that TTT remains an overlooked area compared to the broader test-time compute trend that gained mainstream attention after DeepSeek. [9]
- One framing of TTT, offered by a researcher building TTT models, is that pre-training and post-training are ways to amortize model running costs, and that within a fixed compute budget, allocating part of that budget to test-time training may be more efficient than spending it entirely on pre-training and post-training. [9]
- The concept underlying TTT was historically coined "dynamic evaluation" by Mikolov, according to posts attributing early test-time compute ideas to that prior work. [3][5][6]
Ilya SutskeverSafe SuperintelligenceMITStanfordDeepSeekMikolovtest-time traininginference-time adaptationparameter updates at inferencestate-space modelsreasoning enginescompute budget allocationdistribution shiftsafe superintelligence scalingdynamic evaluationtest-time compute
4.4
Pathway's 150M-parameter BDH-CQ sets new ARC-AGI-1 cost-efficiency record
- Pathway introduced BDH-CQ, a 150M-parameter reasoning model that combines in-context learning with recurrent latent reasoning: inference-time inputs continuously update the model's recurrent memory, and the model solves queries through iterative computation in a high-dimensional latent space without verbalizing intermediate reasoning (i.e., no chain-of-thought token generation). [1][2]
- BDH-CQ's 150M-parameter configuration achieves 29.5% pass@2 on the public ARC-AGI-1 evaluation set at a computed inference cost of $0.0007 per task, which the paper states breaks through the previously reported ARC-AGI-1 cost-accuracy Pareto frontier and establishes a new state of the art in benchmark cost efficiency. [1][2]
- Community comparisons place BDH-CQ's per-task cost at approximately 11× cheaper than GPT-5 on the same benchmark, framing it as an extreme cost-efficiency outlier relative to billion-parameter competitors. [3]
- The paper also uses controlled ARC-like interventions to study what BDH-CQ learns from demonstrations, how consistently it applies an inferred transformation, and which concepts remain difficult — providing an analytical breakdown beyond raw benchmark scores. [1][2]
PathwayBDH-CQARC-AGI-1GPT-5small language modellatent space reasoningin-context learningrecurrent memoryinference cost efficiencyARC-AGI benchmarkno chain-of-thoughtcost-accuracy Pareto frontieriterative computationfew-shot reasoning