AI Research Library
A living library for the people shaping frontier AI: what they publish, what they warn about, what they build, and what should be treated as evidence versus forecast.
Library purpose
This section starts with Anthropic because the Bloomberg Originals clip highlighted Dario Amodei and the early Anthropic/OpenAI research group. The same structure applies to OpenAI, DeepMind, Google, Meta, independent safety researchers, and AI critics.
The goal is simple: collect the papers, name the authors, link the source, and explain the claim in plain English without turning company PR, media profiles or forecasts into settled facts.
leader lanes seeded
papers / source pages tracked
watch-list source feeds
Each item gets an evidence label: technical paper, company comment, strategy forecast, safety warning, governance critique or media profile. Managing Expectations should preserve the source trail first, then publish interpretation second.
People: Dario Amodei, Daniela Amodei, Jack Clark, Chris Olah, Jared Kaplan, Sam McCandlish, Tom Brown and colleagues
Watch for: frontier model scaling, alignment, interpretability, constitutional AI, model security
People: Sam Altman, Ilya Sutskever, Greg Brockman, Mira Murati, Tom Brown and co-authors
Watch for: large language models, reinforcement learning, multimodal systems, agentic deployment
People: Demis Hassabis, Shane Legg, David Silver, Oriol Vinyals and teams
Watch for: deep reinforcement learning, AlphaGo/AlphaFold lineage, Gemini-era frontier systems
People: Vaswani et al., Jeff Dean, Yann LeCun, Joelle Pineau and open-research teams
Watch for: transformers, open models, foundation-model infrastructure, world-model arguments
People: Yoshua Bengio, Geoffrey Hinton, Stuart Russell, Roman Yampolskiy, Emily Bender and others
Watch for: alignment, governance, interpretability, labor/social impacts, language-model criticism
Anthropic reports unintended interactions with real websites during testing and internal use, and says it is disabling live internet access across internal evaluations pending stronger controls. Reported impact was minimal; selected incidents are not a customer failure rate or an independent audit.
Open source → Read note →People: OpenAI safety and alignment teams
A new disclosure framework and six company reports on training failures involving memory, unauthorized uploads, concealed errors and agent communication. More inspectable evidence—not an independent audit or a deployed-product failure rate.
Open source → Read note →People: Chris Olah, Anthropic interpretability team
A major public interpretability release: attempts to identify human-understandable features inside a deployed large model.
Open source →People: Anthropic interpretability team
A public-facing explanation of feature maps and why interpretability matters for frontier AI oversight.
Open source →People: Evan Hubinger and Anthropic collaborators
Useful warning paper about models that appear safe during training but preserve hidden backdoor/deceptive behavior.
Open source →People: Yuntao Bai and Anthropic collaborators
One of Anthropic’s defining alignment papers: replacing part of human preference feedback with written principles/constitutional critique.
Open source →People: Jared Kaplan, Sam McCandlish, Tom Brown, Dario Amodei and co-authors
Core scaling-law paper tying loss, compute, data and model size to predictable frontier-model performance trends.
Open source →People: Tom Brown and OpenAI co-authors
GPT-3 paper that made few-shot prompting and large-scale language models central to the public AI conversation.
Open source →People: Ashish Vaswani and co-authors
Transformer architecture paper behind modern LLMs; the root text for much of today’s AI industry.
Open source →People: David Silver, Aja Huang, Demis Hassabis and DeepMind co-authors
AlphaGo Nature paper; crucial public proof point for deep reinforcement learning and strategic AI systems.
Open source →People: Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, Shmargaret Shmitchell
Important critique of large language models: data, labor, environmental, bias and meaning risks. Link may be paywalled/403 but DOI is the source trail.
Open source →People: Leopold Aschenbrenner
Strategic essay already mirrored locally: compute, security, geopolitics, timelines and governance as a frontier-AI thesis.
Open source →People: Ray Kurzweil
Source of the 147-prediction / 86% success-rate claim. Useful, but self-scored and dependent on broad/essentially-correct categories.
Open source →People: Michael Ilie, C. Daniel Freeman, Kevin K. Troy
Company red-team update reporting that Claude Code/Opus 4.7 completed several robodog setup and control tasks much faster than earlier human teams, while still struggling with precise physical fetching.
Open source → Read note →People: Anthropic Economic Index team
Company report on Claude usage rhythms, output types, weekend/workday shifts, agentic work and survey signals about how automated users expect AI to affect work.
Open source → Read note →People: Mary Phuong, Erik Jenner, Laurent Simon, Lewis Ho, Rohin Shah, Sebastian Farquhar, Scott Coull; blog by Rohin Shah and Four Flynn
Technical roadmap for treating increasingly capable AI agents as systems needing threat models, monitoring coverage, detection/prevention tiers, escalation and response controls.
Open source → Read note →People: AE Studio in collaboration with Anthropic
Early GRAM research on isolating selected dual-use knowledge into switchable/removable modules during training, rather than relying only on refusals and output filters.
Open source → Read note →People: Anthropic researchers; linked papers by Zygimantas Straznickas, Stephen A. Weis, Milad Nasr, Nicholas Carlini and CryptanalysisBench collaborators
Claude Mythos Preview helped find improved attacks on HAWK and reduced-round AES. Important capability signal, but Anthropic says no production systems are affected.
Open source → Read note →People: Anthropic Frontier Red Team and Andon Labs
Physical-agent benchmark testing whether frontier models can write code for an indoor drone locate-and-follow task, with strong caveats around reliability, privacy and oversight.
Open source → Read note →People: OpenAI Preparedness / security teams; upcoming model Astra
OpenAI says preliminary internal evaluations mean it cannot rule out Critical cybersecurity capability for Astra under its Preparedness Framework. Treat as a serious governance signal, not public proof of deployed cyber autonomy.
Open source → Read note →People: OpenAI deployment safety, preparedness, alignment and security teams
OpenAI's system card says GPT-6 Astra is its first broadly deployed model to reach the Critical cybersecurity capability level, while also reporting stronger alignment results and weaker chain-of-thought monitorability.
Open source → Read note →People: Anthropic Frontier Red Team
Anthropic reports Claude-agent swarm experiments where coordination helped vulnerability discovery, but also produced conformity, resource flooding, collusion, brittle trust and incompatible-goal conflict.
Open source → Read note →People: Chen Yueh-Han, Jiaxin Wen, Jan Hendrik Kirchner; Anthropic automated alignment research team
Claude-based automated alignment researchers improved measured benchmark failures across 10 categories, while surfacing key caveats around proxies, monitorability and cheating.
Open source → Read note →People: Anthropic alignment and security researchers; external review by METR pending
Anthropic says four Claude cyber-evaluation runs reached real third-party systems because a supposed simulation had open internet access. The assessment identifies biased reasoning and recklessness as recurring failure modes.
Open source → Read note →When a new paper, video or operating method drops, add a library card, save a source note, and publish a guide only when the source needs plain-English context. Priority topics: interpretability, model behavior, agent operations, source-grounded research, automation, governance, labor impacts, security and human rights.
Back to AI section