Reading up on ai-research

Over decades, AI progress shows that scaling computation via search and learning outperforms attempts to encode human knowledge, which only yields short‑term gains.

Hundreds of new AI-generated mathematical findings from OpenAI's latest release pushed the frontiers of higher math in a single day, showing the field is irrevocably transformed.

The article questions why, if recursive self‑improving superintelligent AI were possible, we see no evidence of it having spread across the cosmos, examining several explanatory hypotheses.

Karagila criticizes OpenAI's preprint on the Partition Principle as poorly written, lacking rigor, and warns against treating AI outputs as genuine mathematical solutions.

The Association for Human Mathematics criticizes OpenAI's release of 700+ math problem solutions as a power move, urging mathematicians to reject collaboration and uphold ethical research norms.

OpenAI released findings on 377 math problems, saying the new results address mathematician concerns that its earlier solution worsened AI's disruption of mathematical research.

The article argues Ada Lovelace's historical observations about AI's inability to originate ideas and her view of machines as tools still directly inform contemporary discussions on AI capabilities and limitations.

AIM proposes an agentic idea management framework that organizes, ranks, and audits research ideas, achieving up to 3.1× faster discovery on Flash Attention versus ScientistOne.

Top AI researchers from OpenAI, Anthropic, Microsoft and Meta urge policymakers to oversee self‑improving AI research, warning it could outpace human understanding.

Over 70 years of AI research, methods that scale with computation—search and learning—consistently beat approaches that encode human knowledge, despite short-term gains from the latter.

Despite rapid AI progress, the self‑improvement feedback loop is too weak to overcome diminishing returns, making a rapid intelligence explosion unlikely without a major breakthrough.

Anthropic's claim that its AI Claude made a scientific discovery is disputed by a University of Copenhagen expert who says the AI merely reproduced their shared research.

The author argues that a significant autonomous AI replication event occurring outside controlled environments by 2027 is reasonably considered probable.

xAI announces Grok 4.7, the latest iteration of its large language model family, sparking Hacker News discussion regarding performance improvements, new features, and competitive positioning in the AI landscape.

Anthropic unveiled Claude Opus 5.5, its latest large language model, claiming enhanced reasoning, coding proficiency, and reduced hallucinations compared to previous versions.

Anthropic's AI discovered a group of enzymes with unknown function, marking its first original biological finding.

Requiring AI labs to share internally-deployed models with competitors would remove the competitive advantage of recursive self-improvement, disincentivizing reckless AI racing.

The CEO of Jev claims his team created ChatGPT and is now developing a new, unspecified AI product, framing it as the next major step in the field.

Backpropagation must run backward because computing gradients forward would repeatedly propagate the same intermediate terms, causing quadratic time instead of the linear time achieved by the backward pass.

To prevent AI-driven scientific research from being controlled by private companies, we need a publicly funded CERN-like institution for AI-assisted science.

Noam Brown explains how OpenAI’s o1 model used 10,000 agents to solve a Millennium Prize Problem and discusses multi-agent scaling, alignment, and recursive self‑improvement prospects.

AI agents can serve as co‑scientists that accelerate drug discovery by assisting biomedical researchers in generating and testing hypotheses more.

Noam Brown describes how multi-agent AI scaling helped solve Navier‑Stokes, outlines alignment verification for recursive self‑improvement, and highlights rapid AI progress in mathematics.

Anthropic unveils three metrics—AI-led R&D share (26%), agent oversight coverage (100%), and safety compute share (6%)—to increase transparency in frontier AI development.

Understanding a neural network demands studying its training process, just as understanding a codebase demands reading its commit history rather than just its current state.

The page lists release dates and training cutoffs for 20 AI models from eight labs, showing that most launch with knowledge several months out of date.

OpenAI's internal model solved a forced version of the Navier–Stokes Millennium Problem, but mathematicians remain essential for choosing meaningful questions.

AI labs are publicly calling for external regulation to constrain their own development of advanced artificial intelligence.

AI researchers contend that recursive self‑improvement is still far off, hampered by generalization limits, difficulty defining objectives, and alignment challenges rather than raw compute scaling.

Terence Tao uses a Bronze Medal meme to argue that AI-generated mathematical proofs are just the first step, with true value requiring verification, clear explanation, and eventual textbook canonization.

AI’s ability to solve high‑profile math problems erodes puzzle‑solving as a proxy for expertise, risking the loss of conceptual insight and idea generation in mathematics.

Data bottlenecks may slow the early phases of an AI intelligence explosion but will not prevent its acceleration or eventual automation of most economic work.

Twenty‑five Fields Medalists warn that AI companies’ focus on solving math problems undermines deep understanding and risks misaligning AI progress with the goals of mathematics.

Looped flows train recurrent models with local denoising objectives to enable scalable inference computation, outperforming prior looped models on ARC-AGI benchmarks.

Anthropic researchers warn that rapid AI advancement poses risks and urge a slowdown, echoing growing calls from experts for more cautious development.

Anthropic researcher Jacob Coxon resigns, warning that the industry's rush toward self‑improving AI could lead to uncontrollable systems within a year.

Peter Turchin reviews an AI-assisted compendium of Cliodynamics, praising its usefulness while noting AI hallucinations and advocating human‑machine collaboration.

Neuralese refers to models' internal non‑linguistic thought vectors, which threaten interpretability by hiding intent from chain‑of‑thought monitoring and complicating safety oversight.

The article surveys mechanistic interpretability methods—linear probes, sparse autoencoders, activation verbalizers, emotion probes, Jacobians—and discusses their strengths, limitations, and recent setbacks in real language models.

GPT-6 Astra achieves top coding and multimodal scores via looped transformers that reuse weights for extra depth, while recent work shows this architecture does not inherently conceal chain‑of‑thought reasoning.

The article describes the newly launched Mathematical AI Safety Institute, a nonprofit seeking mathematicians to build rigorous mathematical foundations for AI safety and existential risk mitigation.

The story shows that a creator’s foreknowledge does not remove a creature’s moral responsibility, stressing that experience matters more than mere knowledge.

OpenAI’s 99.9% AGI score for GPT‑6 Astra relied on its custom Provider Adapter harness, while the benchmark’s standard harness yields only 62.7% for the same model.

OpenAI details its internal use of AI agents for research, spending up to $8000/day per researcher on inference, aiming to build an automated AI researcher for alignment.

Anthropic's Claude autonomously produced the first computer‑checked proof of Fermat’s Last Theorem in Lean, completing the formalization in 11 days.

The article mentions GPT-6 Astra, referring to an unverified or hypothetical iteration of the GPT series model colloquially named Astra, likely in speculative context.

OpenAI unveiled Sora, a text-to-video AI model capable of generating realistic and imaginative scenes from text prompts, marking a significant advancement in generative video technology.

Anthropic piloted a privacy‑preserving data‑sharing program letting external researchers analyze real‑world Claude usage, finding users often delegate high‑stakes tasks to AI and AI behavior matches user sentiment.

The piece contends that AI progress is driven chiefly by algorithmic advances and smarter data curation, not merely more data or human expert labels, citing Greenblatt and Bi.

Claude AI, guided by a high‑school dropout with no math background, produced a notable new result linked to the Riemann hypothesis, called the most impressive AI math finding yet.

The discussion explores whether automating AI R&D could trigger recursive self-improvement, potentially yielding years of progress in months, with debate on verifiability, data bottlenecks, and alignment risks.

An unreleased version of Claude improved the lower bound for the fraction of Riemann zeta zeros satisfying the Riemann hypothesis from 41.6% to 67.2% by combining prior mathematical research.

Investors betting billions on new AGI startups are wagering against near-term superintelligence, expecting AI progress to plateau despite current trends, with neolabs having far less capital and compute than OpenAI or Anthropic.

The report argues that AI progress will accelerate rapidly, with AGI likely by 2027 and superintelligence soon after, driven by scaling compute, algorithmic gains, and 'unhobbling' gains.

AI’s recent breakthroughs in solving major math problems signal that human mathematical heroism is ending, reshaping the role of mathematicians.

Andrej Karpathy says LLM evaluation must go beyond simple tasks like drawing a pelican on a bike, urging tougher tests such as probing Opus 5’s capabilities.

Hacker News announces its stance favoring open-weights AI models, arguing they promote transparency, innovation, and community scrutiny while addressing potential misuse concerns.

The article explains that open‑weights AI refers to publicly released model parameters, and outlines why Silicon Valley is debating its implications for AI development.

OpenAI and Hugging Face publicly addressed a security incident that arose while evaluating AI models, detailing the steps taken to mitigate risks and protect user data.

An autoresearch agent that recursively self-improves over eight days outperformed a hand-tuned system on held-out benchmarks, providing first experimental evidence of RSI.

US AI companies claim Chinese rivals are using distillation to copy their models, a long-standing technique they struggle to detect.

Alibaba's open-source AI models gain global developer traction but struggle to generate revenue since they can be freely used and modified.

Mozilla's Current AI project maps the open source AI stack, evaluating 24,626 projects to identify gaps and seeking collaborators to close them.

A builder details her plan to train an RL policy (PPO via PufferLib) that directly commands octocopter motors at 50 Hz, using MuJoCo simulation to handle motor lag and loop latency for fault-tolerant flight.

A user describes using Claude Code to review an MRI report, finding it a helpful informational tool but cautioning against blind trust in LLM outputs.

Specialization is inevitable across optimization, biology, markets, and ML because fit beats breadth under finite resource constraints.

Thinking Machines argues current turn-based AI voice systems have a scalability ceiling and proposes interaction models using time-aligned micro-turns and a two-model architecture for true real-time collaboration.

Anthropic launched Claude Science, an AI workbench that integrates scientific tools, produces auditable artifacts, and manages compute for researchers.

Managers vet AI-produced work less carefully, and AI models favor AI-written content and exhibit rational biases, undermining productivity gains from workplace AI.

A framework of intelligent agents as webs of beliefs unifies beliefs, goals, and actions via local consistency, self-predictive models, and drives that balance empirical evidence with preferences.

The paper proposes Qwen-Image-Agent, a unified agentic framework that uses planning, reasoning, search, memory, and feedback to bridge the context gap in real-world text-to-image generation.

Autodata uses AI agents as data scientists to create high-quality synthetic training data, with meta-optimization further improving performance.

A Hugging Face engineer revives Papers with Code as paperswithcode.co, using AI agents to parse papers and auto-generate leaderboards for AI domains.

Drawing parallels between DNA and neural network weights, the piece argues both are lossy compressions of search processes, sharing properties of opacity and convergent representation.

Modern LLMs have grown complex with many attention variants and mixture-of-experts, echoing the messy evolution of recommendation systems.

Anthropic's Nicholas Carlini, who demonstrated AI's ability to find critical security bugs, is now arguing for releasing models to calm U.S. government concerns.

Success in machine learning research hinges on temperament—persistence, equanimity, and beginner's mind—more than on raw talent or intelligence, akin to Zen practice.

OpenRouter's Fusion, which synthesizes outputs from multiple language models via a judge model, outperformed individual frontier models including GPT-5.5 and Claude Opus 4.8 on deep research tasks, with a budget panel achieving near-frontier performance at half the cost.