AI research
Source reports covering AI research. Stories will appear when analysis and community reactions are ready.
65 reports · page 1 of 2
Connected topics are temporarily unavailable. You can still browse stories below.
Source reports
Will it be a golden age or a dark age for mathematics (and science)?
opinion
17 min read
Francesco Maggi argues that large language models could accelerate mathematical and scientific discovery but could also concentrate knowledge and weaken independent human research. He cites LLM-generated work including a finite-time blow-up solution for the forced Navier–Stokes equations and a proposed proof of the quasi-Riemann Hypothesis, while stressing that mathematicians must independently scrutinize, interpret, and assimilate such results. Lean, a proof assistant, can formally verify proofs, but important machine-generated Lean proofs may run to hundreds of thousands of lines and remain difficult to translate into understandable natural language. Maggi says theorem proving alone does not define mathematical progress: mathematicians also create new concepts and formal languages through “mathematization,” as exemplified by Newton, Riemann, Einstein, Shannon, and Grothendieck. He argues that authorship, citation, peer review, confidentiality, and the evaluation and training of researchers require new standards when LLMs contribute to research. He also warns that proprietary frontier systems could give a few companies disproportionate control over scientific priorities and capabilities. Maggi calls for coordinated, publicly governed scientific LLM infrastructure, researcher access to frontier systems, disclosure and reproducibility standards for LLM-generated scholarship, confidentiality protections, and research into effects on education and training. He argues these measures could preserve an open, intellectually independent scientific culture while using LLMs to support theory-building across mathematics and other sciences.
ElevenLabs, TwelveLabs, ThirteenLabs
tool
1 min read
A post titled “ElevenLabs, TwelveLabs, ThirteenLabs” was published on quantumi.sh and attributed to jemoka. The post had a score of 108 points and 41 comments at the time of the supplied data. The supplied text provides no further details about ElevenLabs, TwelveLabs, ThirteenLabs, their products, or any claims associated with them.
Micron announces $10B research hub in Boise
news
9 min read
Micron Technology unveiled Micron Research Labs on August 20, 2026, a Boise-headquartered U.S. research institution focused on long-horizon advances in memory, computing and semiconductor manufacturing. The company plans to invest $10 billion in the labs over the next decade, funding a flagship Boise campus, university collaborations, global satellite labs and partnerships with government, startups, customers and other semiconductor-industry participants. Research will cover critical memory technologies, advanced memory and compute architectures, chip packaging, and future manufacturing methods. Micron said the institution will pursue work beyond current technology roadmaps and a 10-year horizon, while connecting its research operations across the United States, Europe, Japan, India, Singapore and Taiwan. The company expects to break ground on a facility capable of hosting hundreds of researchers in 2027; it is also expected to host conferences, workshops and innovation forums in Boise. The planned lab investment is separate from Micron’s previously announced commitment of more than $250 billion to U.S. manufacturing and research and development, which Micron says is expected to create more than 90,000 American jobs. Micron describes itself as the only U.S.-based company developing and manufacturing leading-edge memory.
DiffusionGemma Technical Report
research
2 min read
DiffusionGemma is an experimental open-weight language model that generates text through discrete diffusion, refining 256-token blocks in parallel rather than producing tokens sequentially. The design is intended to avoid the decoding bottleneck of autoregressive language models. The model was created by fine-tuning the mixture-of-experts Gemma 4 model, which has 3.8 billion activated parameters and 25.2 billion total parameters. Its two-stage training process used fewer than 10% of the original autoregressive model’s total training-token budget. Supervised fine-tuning first taught bidirectional denoising; a second phase combined reinforcement learning and sampler distillation to improve generation quality and inference efficiency. Across its evaluation suite, DiffusionGemma generated about 20 tokens per forward pass and about 1,500 output tokens per second on a single Nvidia H100 GPU. The researchers say these results establish a new speed-capability trade-off frontier and exceed autoregressive models, including those using speculative decoding. DiffusionGemma retains Gemma 4’s thinking mode, multimodal-input, and long-context support. It can also still generate text autoregressively with minor performance degradation, suggesting potential hybrid diffusion-autoregressive decoding systems.
Ornith-1.5: From Self-Scaffolding to Self-Improvement
tool
11 min read
Ornith introduced Ornith-1.5, a family of 397B-parameter mixture-of-experts (MoE), 35B MoE, and 9B dense models trained through a self-improvement loop. The system generates progressively harder tasks, creates task-specific scaffolds containing instructions, tools, decomposition, and orchestration, then produces reinforcement-learning rollouts. Rewards optimize task generation, scaffold construction, and solutions jointly using validity, frontier difficulty, novelty, solution quality, and resistance to reward hacking. The target task success rate is 0.2, favoring problems difficult enough to expose capability gaps while still yielding successful training trajectories. Ornith reports that its 397B model scored 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, compared with 85.0 and 59.0 for Claude Opus 4.8. It scored 86.0 on SWE-bench Verified, 92.8 on GPQA Diamond, and 80.0 on MCP-Atlas. The 35B MoE model activates 3B parameters per token and scored 68.5 on Terminal-Bench 2.1 using Claude Code and 79.0 on SWE-bench Verified. The 9B model scored 47.0 and 70.6 on those benchmarks, respectively; Ornith says its quantized Mobile version can run on iPhone and Android devices. Ornith reports that its benchmark results are averaged over five independent runs.
Red queen hypothesis – A new way forward for self-improving AI
research
5 min read
Researchers from NVIDIA and Flower Labs have developed a method for recursive self-improving AI agents that allows them to enhance their own code without reaching an evaluation ceiling. This advancement addresses a significant challenge in the development of self-improving AI.
Research papers using "kidney disappointment" instead of "kidney failure"
tool
1 min read
Google Scholar is a freely accessible web search engine that indexes the full text or metadata of scholarly literature across various formats. It provides a platform for researchers to find academic papers, theses, books, and conference proceedings.
What happens when an LLM never sees material beyond fifth grade?
tool
3 min read
LittleLearner is a hosted 5B model designed for interactive use in a browser, allowing users to study knowledge acquisition in language models. It utilizes an 88B-token corpus filtered to the U.S. elementary-school curriculum, enabling controlled training and comparison with unfiltered models.
Semaglutide linked to 26% lower 5-year predicted dementia risk
research
1 min read
Semaglutide is associated with a 26% reduction in the predicted risk of developing dementia over five years. This finding suggests potential cognitive benefits of the drug beyond its use for weight management and diabetes.
Could a computer scientist build a brain?
research
36 min read
A program is proposed to enable a single cell to build itself into a brain using only genomic information. The program must be compact enough to fit within a genome and efficient enough to complete development in a timely manner.
AI Model Atlas – visualizing populations of ML models as interconnected 3D graph
tool
1 min read
AI Model Atlas visualizes populations of machine learning models as interconnected 3D graphs. This tool facilitates the exploration and analysis of model relationships and performance.
Launch HN: Discovered Materials (YC P26) – AI agents to discover new materials
research
8 min read
The Material Discovery Bench evaluates models on their ability to propose novel, BEOL-compatible crystalline materials that meet specific targets for thermal conductivity ($kappa), static dielectric constant ($epsilon), Young's modulus, and shear modulus. The benchmark utilizes a harness, tools, and graders to measure model performance based on these criteria.
Emergent Introspective Awareness in Large Language Models
research
2 min read
Research investigates the ability of large language models to introspect on their internal states. The study employs injected representations of known concepts into model activations to differentiate genuine introspection from confabulations.
Claude moves bound of the Riemann Hypothesis from 41.6% to 67.2%
research
1 min read
8 days ago, while jogging, I asked Claude to solve the Riemann Hypothesis It didn’t. 1.5 days later, it proved >= 67% of the zeros are on the line (prev: 41.6%) Still not sure what that means, but some analytic number theorists seem excited - I’d tried once before in a single session and it went nowhere. Eight days ago, when I tried again, this was the opening message:The heavy lifting was recent ...
Learning more about Claude's mathematical capabilities
research
6 min read
Claude attempted to address the Riemann hypothesis, a famous unsolved problem in mathematics dating back to 1859. The challenge reflects the ongoing exploration of Claude's mathematical capabilities and its potential contributions to scientific discovery.
U.S. Department of Energy Launches the Genesis Open Models Initiative
news
1 min read
The U.S. Department of Energy has launched the Genesis Open Models Initiative to promote the development and sharing of open-source models for energy systems. This initiative aims to enhance collaboration among researchers and facilitate advancements in energy technology.
I’m leaving OpenAI to build telepathy
opinion
10 min read
Naomi Bashkansky resigned from OpenAI to become a Founding Researcher at Conduit, where she is working on developing telepathy through thought-to-text models. Conduit is collecting large amounts of non-invasive neural data to facilitate this technology.
When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation
research
2 min read
AI benchmarks are essential for assessing model performance and informing deployment choices. However, they often reach saturation, reducing their effectiveness in distinguishing between models and limiting their long-term utility.
Ten advances in mathematics and theoretical computer science
research
4 min read
ChatGPT for Academic Researchers provides 100,000 scientists and mathematicians with free access to advanced ChatGPT models. An AI-generated disproof of the Erdős unit-distance conjecture was shared in May.
Google shuts down Nobel Prize winning AlphaFold
news
2 min read
Google has shut down its AlphaFold project, which was recognized with a Nobel Prize for its contributions to protein folding predictions. Key members of the DeepMind team have been reassigned or have left the company as Google shifts its focus to the Gemini project.
Pacing the frontier
opinion
9 min read
1,134 employees from frontier AI companies assert that while AI has the potential to create a significantly improved future, there are risks associated with rapid capability development. They indicate that the automation of AI research could accelerate progress beyond current understanding or control.
Terence Tao: Mathematics in the Age of AI [pdf]
research
1 min read
Terence Tao explores the intersection of mathematics and artificial intelligence, emphasizing how AI can enhance mathematical research and problem-solving. He discusses the implications of AI for the future of mathematical discovery and education.
Claude Fable produced a counterexample to the Jacobian Conjecture
research
1 min read
The Jacobian conjecture is stated to be false, with a specific mapping from \(\mathbb{C}^3\) to \(\mathbb{C}^3\) shown to have a Jacobian determinant of -2. The mapping sends the points (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0).
GPT-5.6 used a prompt to close a 30-year gap in convex optimization
research
1 min read
GPT-5.6 utilized a prompt to bridge a 30-year gap in convex optimization research. This advancement builds on OpenAI's recent proof announcement related to convex optimization.
Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning
research
2 min read
Ring-Zero scales zero reinforcement learning (RL) to a trillion parameters, enabling emergent reasoning capabilities. This advancement addresses computational constraints that have limited previous studies in zero RL, which utilizes verifiable rewards without human-annotated data.
The One-Step Trap (In AI Research)
research
2 min read
The one-step trap in AI research refers to the misconception that AI agents can rely primarily on one-step predictions to generate longer-term forecasts. This error often occurs when one-step predictions are mistakenly treated as a complete model of the world and its temporal evolution.
Automation Without Understanding
research
2 min read
Artificial intelligence systems are now capable of generating research-level mathematics. Concurrently, the United States is diminishing the educational pipeline that cultivates individuals who can comprehend the outputs of these AI systems.
GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture [pdf]
research
1 min read
GPT-5.6 Sol Ultra has successfully produced a proof for the Cycle Double Cover Conjecture. The proof is documented in a PDF file available online.
AI-generated videos to maximally drive a target brain region
research
1 min read
NEvo is a neural-guided system designed for synthesizing videos through evolutionary algorithms. It leverages deep learning techniques to enhance video generation and editing processes.
Hy3
research
1 min read
Tencent Hy is a platform designed for AI research and development, providing tools and resources for developers to create and deploy AI applications. It features advanced capabilities in natural language processing, computer vision, and machine learning.
A possible future for Damn Interesting
opinion
2 min read
Alan Bellows, founder of Damn Interesting, has sustained the project for nearly twenty years through part-time engineering jobs while dedicating the rest of his time to research, writing, editing, and podcasting. Recent changes in the landscape have prompted reflections on the future of the project.
30papers.com – Ilya's 30 essential ML papers, in a beginner friendly format
research
1 min read
Ilya Sutskever provided John Carmack with a curated reading list of 30 foundational AI and deep learning papers. The list includes full papers and plain-language explanations of complex terms.
The circuit that lets your brain think and see
research
6 min read
Research at Columbia Engineering reveals that the brain's circuitry operates more efficiently than previously understood, challenging traditional views of how visual information is processed. Nuttida Rungratsameetaweemana proposes that the brain integrates thinking and seeing more directly than the conventional model suggests.
Is One Layer Enough? A Single Transformer Layer Matches Full-Parameter RL Train
research
2 min read
Training a single transformer layer can achieve performance comparable to full-parameter reinforcement learning (RL) training. This finding suggests that RL adaptation may not require uniform updates across all layers of large language models.
DeepSeek open-sources inference optimizations with 60–85% faster generation [pdf]
tool
1 min read
DeepSpec is an AI tool designed for formal verification of software systems. It utilizes the DSpark framework to enhance the reliability and security of software through formal methods.
Big AI labs are hiring philosophers
opinion
1 min read
Big AI labs are increasingly hiring philosophers to address complex ethical and societal challenges posed by AI technology. The shift reflects a growing recognition that philosophical insights are essential for navigating the implications of AI advancements.
VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO
research
2 min read
VibeThinker-3B is a compact dense model with 3 billion parameters designed to advance verifiable reasoning in small language models. It utilizes the Spectrum-to-Signal post-training paradigm for systematic enhancement.
US Scientist John Jumper to Leave Google DeepMind for Anthropic
news
1 min read
John Jumper, a prominent scientist at Google DeepMind, will be leaving the company to join Anthropic. His move is part of a broader trend of talent migration within the AI industry.
John Jumper to join Anthropic
news
1 min read
A bit of news: After nearly 9 years, I have decided to leave Google DeepMind and join Anthropic (after taking some time to recharge). I am incredibly grateful for my time at GDM. @demishassabis took a real chance letting me lead the AlphaFold team just six months after finishing my PhD, and the entire GDM team taught me so much about how to do great science. GDM is a special place, and I’ll still ...
Amateur may have cracked Linear A, a 120-year-old puzzle
news
5 min read
Tom Di Mino, a self-taught AI engineer, claims to have deciphered the Bronze-age Minoan writing system known as Linear A, a challenge that has perplexed linguistics experts for over a century. His claims are currently under review by linguistics experts at Rutgers and Cambridge.
The hacker sent by Anthropic to calm the government's nerves about AI safety
news
1 min read
Nicholas Carlini, an Anthropic researcher, demonstrated the vulnerability of systems to new AI models, raising concerns among cybersecurity experts. The Trump administration has expressed anxiety over the potential risks posed by Anthropic's next-generation AI software to global cybersecurity.
Maxproof
research
2 min read
MaxProof is a framework designed for population-level test-time scaling in mathematical proof, specifically within the MiniMax-M3 series. It trains three key capabilities: proof generation, proof verification, and critique-conditioned proof repair.
Do agents.md files help coding agents?
news
1 min read
Sebastian Raschka@rasbthttp://x.com/i/article/2063647807437705216…午前1:47 · 2026年6月8日391239361,38039件の返信を読む
Transformers Are Inherently Succinct
research
1 min read
Transformers exhibit inherent succinctness in their representations, enabling efficient processing of information. This research has been recognized as one of three outstanding papers for ICLR 2026.
Do transformers need three projections? Systematic study of QKV variants
research
2 min read
Transformers utilize a query, key, and value (QKV) attention formulation that is crucial for AI tasks. The study investigates the individual contributions of these three projections and the effects of omitting any of them.
Leiden Declaration on Artificial Intelligence and Mathematics
research
14 min read
The Leiden Declaration on Artificial Intelligence and Mathematics acknowledges the transformative impact of AI technologies on mathematical practice. It recognizes both the potential benefits and diverse reactions among researchers regarding the use of AI in mathematics.
A Eureka machine that thinks like nature and explores what AI cannot
research
2 min read
The Indian Institute of Science has implemented a neuromorphic Ising machine on an FPGA board that rapidly explores complex energy landscapes, facilitating the discovery of near-optimal solutions for optimization problems like protein folding. This machine computes in a manner that mimics natural processes, addressing challenges that traditional computing cannot efficiently solve.
A sleep-like consolidation mechanism for LLMs
research
2 min read
Transformer-based large language models struggle with long-context tasks due to poor scaling of their attention mechanism. Implementing a sleep-like consolidation mechanism allows models to convert recent context into persistent fast weights while clearing their key-value cache.
A recent experience with ChatGPT 5.5 Pro
opinion
25 min read
ChatGPT 5.5 Pro can produce PhD-level research in approximately one hour with minimal mathematical input. Large language models are increasingly capable of solving complex research-level problems, including those related to the Erdős conjectures.
Introspective Diffusion Language Models
research
4 min read
Diffusion language models (DLMs) enable parallel token generation, potentially overcoming the sequential limitations of autoregressive (AR) decoding. However, DLMs currently underperform AR models in quality due to a lack of introspective consistency, where AR models align with their generated outputs.