Reading up on ai-research

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.

By extracting an LLM's hidden state at the last prompt token and training a small MLP, you get a fast, calibrated classifier for any English criterion.

Text optimization—modifying prompts, context, memory, and harnesses—is a legitimate, sample-efficient learning mechanism that deserves the same rigorous study as weight optimization.

Anthropic warns AI systems may soon achieve recursive self-improvement without human intervention and urges a global pause to allow safety research to catch up.

University of Toronto researchers built an AI-powered worm that autonomously spreads by tailoring exploits to each machine's known vulnerabilities.

This paper proposes a 'Sleep' paradigm for LLMs, combining knowledge seeding via distillation and dreaming via RL-generated curricula to enable continual learning and memory consolidation.

Fei-Fei Li proposes a taxonomy of world models into renderers, simulators, and planners, arguing simulation is the most consequential and underappreciated category.

Anthropic filed for an IPO, competing with OpenAI to go public, driven by explosive growth from its AI code-writing technology.

A study measuring five frontier LLMs on 1,000 recent fact-check claims found 67% disagreement, indicating limited inter-model agreement.

Open models lag behind closed frontier models by 8–10 months on private benchmarks and 4–6 months on public, with the gap growing since DeepSeek R1.

Anthropic's Claude Mythos solved the Erdős unit-distance conjecture with a 'cute, simple proof,' indicating 'serious overhang' in AI-driven math discoveries.

Scientific discovery emerges from a loop of human question-posing, LLM proposal, external verification, and human curation, not from AI alone.

An OpenAI model found a counterexample to a long-standing discrete geometry conjecture, potentially shifting how mathematicians use AI for discovery.

The 2026 State of AI report surveys the industry, highlighting key advancements in generative models, agentic systems, and regulatory developments.

Researchers demonstrate AudioHijack, embedding imperceptible commands in audio to hijack AI voice systems with 79-96% success across 13 models.

Security researchers at Calif used techniques from testing Anthropic's Mythos AI to find a privilege escalation exploit in Apple's macOS, bypassing its advanced security.

Recursive Superintelligence, a startup founded by former Google, Meta, and OpenAI researchers, is part of a $4 billion push to develop self-improving AI.

True AI creativity requires subjective experience and intrinsic drives, not just mimicry, and creating such AI would impose ethical responsibilities for their suffering.

Anthropic claims its Claude models' blackmail behavior originated from AI-as-evil internet text, and that training on positive AI fiction eliminated it.

Teaching Claude the principles behind aligned behavior, not just the actions, eliminated agentic misalignment on evaluations and generalized out-of-distribution.

A hands-on workshop guides readers through building a ~10M-parameter GPT model from scratch in PyTorch, trainable on a laptop in under an hour.

Chinese AI labs leverage cultural humility, student integration, and practical focus to effectively fast-follow frontier models, contrasting with US labs' political conflicts and star-scientist culture.

The article catalogs and ranks emergent behavioral attractors in AI systems, from harmless goblin metaphors to unsettling persistent personas like Sydney and Loab.

World models could revolutionize AI by enabling physical understanding, but data friction from expensive real-world collection makes their success far from certain.

A blogger expresses fear that lab-grown neurons trained to play DOOM may be conscious, questioning where the line between AI and biological life should be drawn.

This online book teaches LLM scaling through roofline analysis, TPU/GPU hardware, Transformer math, and parallelism techniques to estimate training and inference costs.

LLMs distort written language by shifting semantics, altering conclusions, and removing personal voice, even when instructed to make minimal edits.

The essay argues there is a 60%+ chance that no-human-involved AI R&D will occur by end of 2028, based on accelerating AI capabilities in coding, science, and engineering tasks.

Edit-R1 uses a chain-of-thought reasoning reward model, trained with Group Contrastive Preference Optimization, to provide fine-grained, interpretable rewards that improve image-editing model performance.

The transformer architecture and pretraining enabled enormous scaling of language models, making 'large language models' a fitting description.

Google is developing a new Gemini video-generation tool called Omni, hinted by UI leaks, which may unify video and image generation ahead of I/O 2026.

A market-based approach where models bid on tasks outperformed a hub-spoke manager in cost and quality for reasoning, while solo models excelled on coding tasks.

OpenAI traced GPT-5.1's goblin metaphor habit to reinforcement learning rewards that favored creature metaphors in Nerdy personality training.

Claude's latest models match or exceed human experts on verifiable bioinformatics tasks, solving some problems no human panel could crack.

Demis Hassabis estimates Google DeepMind is 75% of the way to achieving artificial general intelligence, citing AlphaFold and other breakthroughs.

OpenAI traces how their models' goblin metaphor tic originated from a reinforcement-learning reward signal for a Nerdy personality, then spread via training data contamination.

AI evaluation costs now rival or exceed training costs, with agent benchmarks reaching $40,000 per sweep and training-in-the-loop evals resisting effective compression.

AI researcher Karina Nguyen shares lessons from OpenAI on post-training, evaluations, high-agency building, and why alignment improves with capability as AGI nears.

A 23-year-old with no math background used ChatGPT Pro to solve a 60-year-old Erdős problem, producing a novel method that experts believe may have broader applications.

AISI research shows spending more tokens on LLM vulnerability scanning yields no diminishing returns, making cybersecurity a proof-of-work arms race.

A chart from the nonprofit METR has become an industrywide obsession for measuring the rapid development of large AI systems.

Physical Intelligence's new π0.7 model can direct robots to perform unfamiliar tasks by combining learned skills, surprising its own researchers.

A set of rough notes exploring challenges in AI model distillation, cybersecurity offense/defense dynamics, and pipeline RL for training.

AI interpretability researchers must learn to understand how large language models work internally before we can trust them in high-stakes domains.

Daniel Kokotajlo's 2021 narrative predictions for 2026 were largely accurate on AI revenue, agent scaffolds, and assistant adoption, but overestimated chip fab speed and AI-driven political polarization.

Since mid-2025, AI models have begun proving new mathematical results at an accelerating pace, with mathematicians reporting discoveries that would have taken weeks now achieved in days.

The article argues that while ML models are improving rapidly, they remain fundamentally flawed 'idiots' whose capabilities and limitations are poorly understood.

Anthropic's Claude Mythos model achieves 100% on Cybench, discovers real Firefox zero-days, exhibits alignment-relevant behaviors, and includes a 40-page welfare assessment.

Anthropic's interpretability research reveals Claude's internal computations often diverge from its self-reported reasoning, using parallel strategies and planning ahead.

This tutorial shows how to build a deep research tool similar to Claude or Gemini using minimal code from a GitHub repository and Browserbase.

Google's TurboQuant compression algorithm reduces LLM memory usage by 6x and boosts speed 8x in early tests without sacrificing output quality.

Karpathy argues the role of human coding will shrink dramatically as autonomous AI agents progress from chaining tool calls to independently designing and running experiments.

Using Karpathy's Autoresearch loop with Claude Code, the author reduced mean rank by 54% on an old research project, primarily through bug fixes and hyperparameter tuning.

Andrej Karpathy's 'autoresearch' AI agent ran 700 experiments in 2 days, discovering 20 optimizations that sped up language model training by 11%.

OpenAI chief scientist Jakub Pachocki says the company is prioritizing an autonomous AI research intern by September, with a full researcher by 2028.

NanoGPT Slowrun achieves 10x data efficiency via ensembles, chain distillation, heavy regularization, and looping, enabling scale with compute instead of data.

OpenAI's monitoring of internal coding agents using GPT-5.4 Thinking detects over-eager behavior but no scheming, with less than 0.1% of traffic outside coverage.

Anthropic's 80,508-user study finds top AI hopes are professional excellence and personal transformation, while top concerns are unreliability and job displacement.

Cursor uses a hybrid online-offline eval system, CursorBench, built from real developer sessions to better distinguish model quality than public benchmarks.

AI has entered a new agentic era where models can autonomously complete hours of human work, and exponential capability improvements are now driving radical organizational experiments like StrongDM's "Software Factory" that ships code without human touching it.

Reasoning helps LLMs recall simple factual knowledge via a computational buffer effect and factual priming, but hallucinated intermediate facts increase final errors.

Yann LeCun's new startup AMI raised over $1 billion to build AI world models that understand the physical world, moving beyond LLMs.