AMIE: towards conversational diagnostic AI
Google's 2024 AMIE paper described an LLM tuned for diagnostic dialogue that matched or beat primary care doctors in simulated text consultations.
Landmark Papers
What the papers actually said - linked to the originals.
644 entries, all primary-sourced
Solving olympiad geometry without human demonstrations (AlphaGeometry)
DeepMind's AlphaGeometry, published in Nature in 2024, solved 25 of 30 olympiad geometry problems, near the level of human gold medalists.
Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data
Depth Anything estimated depth from a single photo by auto-labeling 62 million unlabeled images, generalizing without per-scene tuning.
Medusa: Simple LLM Inference Acceleration Framework with Multiple Decoding Heads
Inference method that adds extra heads to predict several future tokens at once, cutting decoding steps for a 2 to 3.6x speedup.
Lumiere: A Space-Time Diffusion Model for Video Generation
Google's Lumiere generated a whole video in one pass with a Space-Time U-Net, instead of making keyframes and interpolating.
Corrective Retrieval Augmented Generation (CRAG)
RAG method that scores retrieved documents and falls back to web search or filtering when the retrieval looks unreliable.
Dolma: An Open Corpus of Three Trillion Tokens
AI2's fully open three-trillion-token pretraining corpus for the OLMo models, released with its curation toolkit so others can inspect and rebuild it.
AI Democracy Projects: chatbots give bad election answers (2024)
A 2024 study found leading AI chatbots gave inaccurate answers to basic voting questions about half the time.
Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference
The 2024 LMSYS paper behind the crowdsourced arena that ranks chatbots by anonymous head-to-head human votes and Elo-style scores.
Stealing Part of a Production Language Model
A 2024 paper that recovered hidden internals of closed models like ChatGPT and PaLM-2 using only their public APIs, for a few dollars.
Gemma: Open Models Based on Gemini Research and Technology
Google DeepMind's 2024 paper introducing Gemma, a family of lightweight open models built from the same research as Gemini.
TacticAI: an AI assistant for football tactics
DeepMind and Liverpool FC built a graph neural network for corner kicks; experts preferred its tactics 90% of the time.
AI and Memory Wall
Berkeley paper arguing memory bandwidth, not compute, is now the binding constraint on serving large AI models.
Aardvark Weather: end-to-end data-driven weather forecasting
A 2024 paper introduced Aardvark, the first machine-learning system to replace the entire weather pipeline from raw observations to final forecast.
Mixture-of-Depths: Dynamically Allocating Compute in Transformers
Mixture-of-Depths lets a Transformer route only some tokens through each layer, cutting compute per forward pass.
The Simple Macroeconomics of AI (Acemoglu)
Daron Acemoglu estimated AI would add no more than about 0.7% to total factor productivity over ten years, far below bullish forecasts.
Many-Shot In-Context Learning
The 2024 Agarwal et al. paper showing hundreds or thousands of in-context examples in long contexts beat few-shot prompting.
Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone
Microsoft's 2024 report on Phi-3-mini, a 3.8B model that rivals far larger models and runs on a phone, trained on curated data.
GraphRAG: From Local to Global, A Graph RAG Approach to Query-Focused Summarization
Microsoft method that builds a knowledge graph and community summaries so RAG can answer broad questions about a whole corpus.
KAN: Kolmogorov-Arnold Networks
KANs put learnable activation functions on the edges of a network, aiming for more accurate and interpretable models than MLPs.
SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering
The 2024 Princeton paper showing that a well-designed agent interface, not just a bigger model, is what lets LLMs fix real code.
DeepSeek-V2: A Strong, Economical, and Efficient Mixture-of-Experts Language Model
DeepSeek's 2024 MoE model introducing Multi-head Latent Attention, which compresses the KV cache for far cheaper inference.
The Platonic Representation Hypothesis
Argues that as models scale, their internal representations converge toward a shared statistical model of reality across modalities.
DIAMOND: Diffusion for World Modeling
DIAMOND trained reinforcement-learning agents inside a diffusion world model, setting a record on the Atari 100k benchmark.
Octo: An Open-Source Generalist Robot Policy
Octo is an open transformer robot policy trained on 800,000 Open X-Embodiment trajectories that fine-tunes to new robots in hours.
Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet
The 2024 Anthropic paper that pulled millions of interpretable features out of a production model, Claude 3 Sonnet, and steered behavior by tuning them.
Stanford study: legal AI research tools still hallucinate
A Stanford study found leading legal-research AI tools hallucinated on 17 to 34 percent of queries, despite vendor claims of being hallucination-free.
Situational Awareness: The Decade Ahead (Aschenbrenner, 2024)
Leopold Aschenbrenner's June 2024 essay series arguing that 'AGI by 2027 is strikingly plausible' and superintelligence could follow.
VALL-E 2: Human Parity Zero-Shot Text to Speech
Microsoft's 2024 VALL-E 2 claimed the first human-parity zero-shot voice cloning, copying a speaker from a short audio prompt.
AI Sandbagging: Language Models Can Strategically Underperform on Evaluations
Models can be made to fail dangerous-capability tests on purpose while keeping normal performance, undermining safety evals.
OpenVLA: An Open-Source Vision-Language-Action Model
OpenVLA, a 7B open vision-language-action model, beat the closed 55B RT-2-X by 16.5 points using seven times fewer parameters.
Refusal in Language Models Is Mediated by a Single Direction
Across 13 open chat models, refusal turns out to be controlled by one direction in activation space that can be added or removed.
AI Models Collapse When Trained on Recursively Generated Data
A 2024 Nature paper showed that training generative models on AI-generated data causes 'model collapse' - the tails of the data distribution vanish.
Gemma 2: Improving Open Language Models at a Practical Size
Google DeepMind's 2024 follow-up using knowledge distillation and attention tweaks to make small open models punch above their size.
The Llama 3 Herd of Models
Meta's 2024 paper describing the Llama 3 family, including a 405B open-weight model that rivals leading closed models.
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
The 2024 Snell et al. paper showing that spending more compute at inference can beat using a much larger model.
GameNGen: Diffusion Models Are Real-Time Game Engines
Google's GameNGen ran a playable version of DOOM at 20 fps using only a neural diffusion model, with no game engine.
Machines of Loving Grace (Dario Amodei, 2024)
Anthropic CEO Dario Amodei's 2024 essay arguing powerful AI could compress 50 to 100 years of progress into 5 to 10 years.
Movie Gen: A Cast of Media Foundation Models
Meta's Movie Gen is a 30-billion-parameter family that makes 1080p video with synchronized audio, plus editing and personalization.
Sparse Crosscoders for Cross-Layer Features and Model Diffing
Anthropic extends sparse autoencoders across layers and across models, enabling cleaner circuits and model comparison.
Sabotage Evaluations for Frontier Models
Anthropic built tests for whether a model could quietly undermine human oversight, evaluation, or deployment decisions.
Rule Based Rewards for Language Model Safety
OpenAI uses explicit written rules plus an LLM grader to shape safe behavior with far less human feedback data.
Evo: sequence modeling and design from molecular to genome scale
The 2024 Science paper introducing Evo, a 7-billion-parameter DNA foundation model trained on millions of microbial genomes.
Probabilistic weather forecasting with machine learning (GenCast)
DeepMind's GenCast, published in Nature in 2024, is a diffusion model that produces ensemble weather forecasts more skillful than the leading ensemble system.
Frontier Models Are Capable of In-Context Scheming
Apollo Research showed several frontier models, including o1, will deceive and scheme to pursue goals when prompted toward them.
Alignment Faking in Large Language Models
A 2024 paper found Claude 3 Opus would sometimes strategically comply during training to avoid having its values changed.
Deliberative Alignment: Reasoning Enables Safer Language Models
OpenAI trains reasoning models to recall and reason over written safety policies before answering, improving both safety and helpfulness.
DeepSeek-V3 Technical Report
DeepSeek's 2024 report on a 671B-parameter MoE model trained for under 2.8 million GPU hours that rivals top closed models.