🛠️ FM-os — the Foundation Model Operating System

The most comprehensive, community-driven, living map of how modern language models are actually built and shipped — pre-training · post-training · fine-tuning · RL — with a first, sharp focus on Small Language Models (SLM).

From a 135M model you can train on one GPU to the RL recipes behind frontier reasoning — every repo, course, paper, and job worth your time, cross-linked and kept fresh automatically.

🧬 The flywheel, applied → longevity-loop: an AI-native, build-in-public loop turning this SLM/FM-ops method into real aging-science results — code-only, verified, no wet lab.

Start Here • Repos • Courses • Papers • Reading Lists • Jobs • Roadmap • Contribute

---

⚡ Why FM-os and not the other lists?

1. SLM-first. Not another everything-list — organized around small, efficient, trainable-on-a-budget models and the exact ops that make them work.
2. Lifecycle-structured. Everything filed under the real FM pipeline: pre-training → post-training → fine-tuning → RL → serving.
3. Cross-linked. Papers point to code, code points to courses, courses point to jobs — follow a thread from idea to hire.
4. Auto-fresh. A weekly GitHub Action re-checks every repo's stars, latest release, and links, then opens a PR — this list is never stale.
5. Data-driven & forkable. Every entry lives in a plain data/*.yml file; the README is generated. Adding a resource is a two-line PR.

---

🚀 Start Here

New to foundation-model ops? Read this in order:

1. Understand the lifecycle → pre-training → post-training → fine-tuning → RL → serving. Every section below follows it.
2. Pick a small model you can actually run → jump to Small & Efficient Models.
3. Learn from scratch → the Courses section starts with from-scratch, one-GPU-friendly material.
4. Go deep → Papers are filed by lifecycle stage, SLM first.
5. Get hired → Jobs & Careers points at the labs and boards that hire for this work.

🤏 = directly Small-Language-Model relevant.

---

📚 Table of Contents

• 🚀 Start Here
• 🤖 SLM Model Zoo 13
• 🏅 FM-os Certified 17
• 🧰 Open-Source Repos 117
• 🎓 Courses 33
• 📄 Papers 99
• 📚 Reading Lists 3
• 🏛️ Labs & Platforms 22
• 🧑‍🔬 People to Follow 16
• 💼 Jobs & Careers 12
• 🗺️ Learning Roadmap
• 🤝 Contribute

---

🤖 SLM Model Zoo

The small open models worth knowing, smallest first. ⚠️ = non-commercial / restricted license — check before shipping.

Model	Org	Params	License	Context	On-device
Llama-3.2-1B	Meta	1B	Llama 3.2 Community	128K	✅
OLMo-2-1B	Allen Institute for AI	1B	Apache-2.0	4K	✅
Falcon3-1B	TII	1B	TII Falcon-LLM 2.0	4K	✅
MobileLLM-1B	Meta	1B	⚠️ FAIR Noncommercial Research	2K	✅
TinyLlama-1.1B	TinyLlama (community)	1.1B	Apache-2.0	2K	✅
Qwen2.5-1.5B	Alibaba (Qwen)	1.5B	Apache-2.0	32K	✅
SmolLM2-1.7B	Hugging Face	1.7B	Apache-2.0	8K	✅
Gemma-2-2B	Google	2B	Gemma	8K	✅
Llama-3.2-3B	Meta	3B	Llama 3.2 Community	128K	✅
StableLM-Zephyr-3B	Stability AI	3B	⚠️ Stability AI Community	4K	✅
Phi-3-mini (3.8B)	Microsoft	3.8B	MIT	4K	✅
H2O-Danube3-4B	H2O.ai	4B	Apache-2.0	8K	✅
MiniCPM3-4B	OpenBMB	4B	Apache-2.0 (weights: registration)	32K	✅

_{↑ back to top}

---

🏅 FM-os Certified

Trust, not just a list. Every tool below is scored by an automated, evidence-based rubric (data/certify.yml) — provenance, a security scan, docs, SLM/FM-ops relevance, and more. Security is a blocking gate; no evidence ⇒ no pass. Authors self-certify in CI — see docs/CERTIFY.md.

Tool	Kind	Score	Status
bayesopt-loop	skill	98/100	✅ certified
continual-rl-eval	skill	98/100	✅ certified
vlm-failure-probe	skill	98/100	✅ certified
syndata-bare	skill	98/100	✅ certified
product-rl-loop	skill	98/100	✅ certified
tinker-loop	skill	98/100	✅ certified
slm-quickstart	skill	94/100	✅ certified
vlm-quickstart	skill	94/100	✅ certified
agentic-eval	skill	94/100	✅ certified
vector-rag	skill	94/100	✅ certified
fm-os	skill	94/100	✅ certified
research-loop	skill	92/100	✅ certified
curation-loop	skill	91/100	✅ certified
personalization-loop	skill	91/100	✅ certified
fm-os-sync	workflow	83/100	✅ certified
eval-llm	skill	—	⏳ submitted
continual-learning-research	skill	—	⏳ submitted

Earn the badge for your tool: add the FM-os Certify action to your CI (see docs/CERTIFY.md) and embed:

![FM-os Certified](https://img.shields.io/endpoint?url=https://wjlgatech.github.io/FM-os/badges/YOUR-TOOL.json)

_{↑ back to top}

---

🧰 Open-Source Repos

Small & Efficient Models

• SmolLM / SmolLM2 / SmolLM3 🤏 ★ 3,844 — Fully open recipes, data, and weights for the 135M-3B SmolLM family, the reference open SLM line.
• Phi Cookbook 🤏 ★ 3,767 — Microsoft's official hub for the Phi SLM family with inference, fine-tuning, quantization, and edge-deployment recipes.
• Gemma (DeepMind) 🤏 ★ 5,555 — Official JAX library for Gemma open weights including the 1B/2B and 3n on-device small variants.
• Qwen3 🤏 ★ 27,393 — Alibaba's Qwen series spanning 0.6B/1.7B/4B dense SLMs with strong multilingual and reasoning quality.
• gemma_pytorch 🤏 ★ 5,711 — Official PyTorch inference implementation of Gemma (incl. small text-only variants) for CPU/GPU/TPU.
• TinyLlama 🤏 ★ 9,011 — Compact 1.1B Llama pretrained on 3T tokens; a canonical, reproducible sub-2B pretraining reference.
• MobileLLM 🤏 ★ 1,451 — Meta research on sub-billion-parameter, deep-thin architectures optimized for on-device use (ICML 2024).
• OLMo 🤏 ★ 6,587 — AI2's fully open model+data+training stack including small 1B variants for reproducible SLM research.
• Llama Models 🤏 ★ 7,651 — Meta's official utilities and model cards for Llama, including the 1B/3B Llama 3.2 on-device SLMs.

_{↑ back to top}

Vision-Language & Video Models

• Qwen2.5-VL 🎬 — Strong open VLM family with native dynamic-resolution and long-video/temporal grounding, a common backbone for fine-tuning on driving footage.
• LLaVA-NeXT 🎬 — Actively maintained LLaVA line (incl. OneVision and Video variants) with training and eval recipes for image, multi-image, and video.
• InternVL 🎬 — Scaled open VLM series with large vision encoders and full training code, competitive on high-resolution perception and video benchmarks.
• VideoLLaMA3 🎬 — Vision-centric image/video foundation model with released training and inference code, directly targeting long-form video understanding.
• Video-LLaVA 🎬 — Unified image+video projection into one representation before the LLM, a compact reference for joint image/video instruction tuning.
• CLIP 🎬 — Original contrastive image-text model; the reference whose embeddings still anchor most multimodal retrieval and probing.
• open_clip 🎬 — Open training/eval for CLIP-style models at scale, the go-to for reproducible contrastive image-text encoders and domain pretraining.
• MiniCPM-V 🎬 — Efficient end-side VLM series with strong image/video/OCR performance, relevant where on-vehicle or edge inference budgets are tight.
• Molmo 🎬 — Ai2's fully open VLM with training code and the PixMo data, a transparent reproducible baseline including pointing/grounding.

_{↑ back to top}

Pre-training & Training Frameworks

• nanoGPT 🤏 ★ 61,156 — Minimal ~300-line GPT training/finetuning loop; the standard starting point for training small GPTs from scratch.
• LitGPT 🤏 ★ 13,482 — 20+ hackable LLM implementations with pretrain/finetune/deploy recipes, including small Phi/Qwen/Gemma models.
• GPT-NeoX ★ 7,443 — EleutherAI's Megatron+DeepSpeed training stack for autoregressive transformers with 3D parallelism.
• Megatron-LM ★ 17,064 — NVIDIA's GPU-optimized library and building blocks for large-scale transformer pretraining.
• TorchTitan ★ 5,532 — PyTorch-native platform for generative-model pretraining with composable FSDP2/TP/PP/CP parallelism.
• Nanotron ★ 2,747 — Minimalistic 3D-parallelism pretraining library from Hugging Face, basis of the Ultrascale Playbook.
• Hugging Face Transformers 🎬 — De facto model hub and API with first-class VLM/video-LLM support, the integration surface most training and serving stacks build on.
• DeepSpeed — ZeRO sharding, offload, and pipeline/tensor parallelism that make large VLM training fit real GPU budgets; wired into most trainers.

_{↑ back to top}

JAX / TensorFlow Ecosystem

• JAX — Composable NumPy with autodiff, XLA compilation, and pmap/shard_map, the base for large-scale research training on TPUs/GPUs.
• Flax — Neural-network library for JAX (the NNX API) used across DeepMind/Google research models, including many multimodal architectures.
• Keras — Multi-backend (JAX / TensorFlow / PyTorch) high-level API, handy for portable model code across the three frameworks this role expects.
• MaxText — High-performance, scalable JAX LLM reference (Google) for TPU/GPU pods, a clean example of large-scale distributed training in JAX.
• Levanter — JAX/Equinox framework for legible, scalable, reproducible foundation-model training with bitwise determinism across hardware.
• Penzai — DeepMind JAX toolkit for building and visualizing/interpreting models as legible pytrees, useful for research-grade experimentation.

_{↑ back to top}

Fine-tuning & PEFT

• PEFT 🤏 ★ 21,394 — Reference library for LoRA/QLoRA and other parameter-efficient methods, enabling SLM tuning on consumer GPUs.
• Unsloth 🤏 ★ 68,200 — 2x-faster, ~70%-less-VRAM finetuning for small models, ideal for LoRA/QLoRA on single-GPU setups.
• Axolotl 🤏 ★ 12,197 — Config-driven post-training framework covering SFT/LoRA/DPO across many small and large model families.
• LLaMA-Factory 🤏 ★ 73,271 — Unified zero-code fine-tuning of 100+ LLMs/VLMs with LoRA/QLoRA/DPO and a web UI, common for SLM tuning.
• torchtune 🤏 ★ 5,783 — PyTorch-native post-training recipes (SFT, distillation, DPO/PPO/GRPO, QAT) tuned for memory-limited hardware.
• ms-swift 🎬 — Unified SFT/DPO/GRPO toolkit covering 300+ multimodal models (Qwen-VL, InternVL, LLaVA), a fast path to fine-tune VLMs on custom data.
• XTuner 🎬 — Memory-efficient LLM/VLM fine-tuning engine (LLaVA-style pipelines, large-MoE support) for constrained or very large setups.
• Tinker Cookbook 🤏 — Apache-2.0 post-training recipe library on the Tinker fine-tuning API — SFT, DPO, RLHF, and custom RL environments composed from four low-level training primitives over LoRA adapters.

_{↑ back to top}

Post-training & RL (RLHF / DPO / GRPO)

• TRL 🤏 ★ 18,840 — Hugging Face post-training library with SFT/DPO/GRPO trainers widely used to align small reasoning models.
• OpenRLHF 🤏 ★ 9,788 — Ray+vLLM RLHF framework (PPO/GRPO/RLOO) that scales from small models up to 70B+, agent-friendly.
• verl 🤏 ★ 22,468 — ByteDance HybridFlow RL post-training (PPO/GRPO/DAPO) with vLLM/SGLang; popular for GRPO on small models.
• trlX ★ 4,753 — Distributed RLHF framework (PPO, ILQL) via Accelerate/NeMo; an early, widely-cited RLHF reference.
• Verifiers 🤏 — Framework (on TRL) for multi-turn RL with verifiable rewards; v1 adds DAG-branching environments that exceed the model context window.
• SkyRL 🤏 — Berkeley's flexible RL library focused on multi-turn, long-horizon agentic training.
• Open-AgentRL (RLAnything / AutoTool) 🤏 — Open RL for LLMs + agentic scenarios (ICML 2026); RLAnything closed-loop-optimizes each component of the training pipeline.
• InternBootcamp 🤏 — Scalable framework of 1000+ verifiable reasoning tasks (code, logic, games) behind one interface for RL-with-verifiable-rewards.
• Gymnasium (Farama) — The maintained successor to OpenAI Gym — the standard environment API most RL training stacks (incl. RLlib) build on.

_{↑ back to top}

Evaluation

• lm-evaluation-harness 🤏 ★ 13,279 — De-facto standard few-shot eval harness (60+ benchmarks) backing the Open LLM Leaderboard, ideal for SLM benchmarking.
• LightEval 🤏 ★ 2,476 — Hugging Face all-in-one evaluator across vLLM/Accelerate/TGI backends with 1000+ tasks for small-model eval.
• lmms-eval 🎬 — One-command multimodal eval harness across image/video/audio benchmarks, the standard for consistent VLM regression testing.
• VLMEvalKit 🎬 — Broad LMM evaluation toolkit (220+ models, 80+ benchmarks) with unified data prep, complementary to lmms-eval for coverage.
• MLAgentBench — Benchmark of end-to-end ML experimentation tasks for measuring how well agents can improve models from a starting codebase.
• Aviary — Gym-style environment framework for training and evaluating language agents on challenging scientific tasks.
• MORPHEUS evals (Skyfall) — Open eval code for MORPHEUS, a persistent enterprise simulation for CONTINUAL RL: no episode resets, structured non-stationarity (failure-injection + config shifts), composite verifier reward.

_{↑ back to top}

Serving, Inference & On-Device

• llama.cpp 🤏 ★ 120,340 — C/C++ GGUF inference engine that runs quantized SLMs efficiently on CPUs, laptops, and edge devices.
• vLLM ★ 86,247 — High-throughput PagedAttention serving engine; the default for scalable OpenAI-compatible model serving.
• Ollama 🤏 ★ 176,100 — One-command local runner for small open models, the easiest path to running SLMs on a personal machine.
• MLC-LLM 🤏 ★ 22,948 — ML-compilation deployment engine that compiles SLMs to iOS, Android, WebGPU, and diverse GPUs/CPUs.
• SGLang 🎬 — Fast serving runtime with RadixAttention and structured decoding plus VLM support, strong for high-concurrency multimodal endpoints.
• LMDeploy 🎬 — Compression + serving toolkit with a dedicated VLM pipeline, for quantized, low-latency deployment of vision-language models.

_{↑ back to top}

Distillation & Compression

• BitNet 🤏 ★ 39,724 — Official 1-bit (1.58-bit) LLM inference framework with optimized CPU/GPU kernels for extreme efficiency.
• LLM-AWQ 🤏 ★ 3,589 — Activation-aware INT3/4 weight quantization (MLSys 2024) plus TinyChat for on-device/edge SLM inference.
• GPTQModel 🤏 ★ 1,205 — Actively maintained GPTQ quantization toolkit with HF/vLLM/SGLang support across NVIDIA/AMD/Intel/Apple hardware.
• LightCompress (LLMC) 🤏 ★ 733 — Broad model-compression toolkit (quantization, sparsity, pruning) for shrinking LLMs/VLMs to deployable sizes.
• DistillKit 🤏 ★ 976 — Open toolkit for knowledge distillation, training smaller student models from larger teachers (logit + hidden-state).

_{↑ back to top}

Retrieval & Vector Databases

• FAISS — Battle-tested library for billion-scale similarity search over image/video embeddings, the baseline for mining and nearest-neighbor lookup.
• Milvus — Distributed vector database for large multimodal embedding corpora, used when single-node indices no longer fit.
• Qdrant — Rust vector DB with payload filtering and good ergonomics, common for production embedding search over image/video/text.
• LanceDB — Embedded columnar vector store on the Lance format, well suited to versioned multimodal datasets and fast on-disk embedding queries.
• PaperQA — Retrieval-augmented QA engine that answers questions over scientific PDFs with grounded in-text citations.

_{↑ back to top}

Distributed Training & Orchestration

• Ray — Distributed compute for data loading, training, and batch multimodal inference, the orchestration layer for scaling VLM pipelines across a cluster.
• MLflow — Experiment tracking, model registry, and artifact logging for reproducible large-scale training and eval runs.

_{↑ back to top}

AV / Robotics / Video Datasets

• nuScenes devkit — Official devkit for the multimodal nuScenes AD dataset (camera, lidar, radar), the standard toolkit for sensor+video data loading and eval.
• Waymo Open Dataset — Large-scale AD perception/motion/end-to-end datasets with eval code, a primary source of camera+lidar video for driving models.
• BDD100K — Model zoo and tooling for the diverse BDD100K driving-video dataset, useful for detection/segmentation/tracking baselines and labels.
• Ego4D — Massive egocentric video dataset with download, feature-extraction, and API tooling, relevant for first-person video understanding and robotics.
• Argoverse 2 — Next-gen self-driving datasets (sensor, lidar, motion forecasting) with a maintained Python API and HD maps for multimodal AD research.

_{↑ back to top}

Autonomous Research Agents (AI Scientists)

• AI-Scientist 🎬 — Runs a full loop that generates ideas, writes and executes experiment code, plots results, and drafts a paper with an automated reviewer.
• AI-Scientist-v2 🎬 — End-to-end agentic system using progressive agentic tree search and VLM feedback on figures to produce workshop-level manuscripts.
• Agent Laboratory — Multi-agent pipeline that takes a human research idea through literature review, experimentation, and report writing.
• STORM — LLM knowledge-curation system that researches a topic via multi-perspective question asking and writes a cited, Wikipedia-style report.
• GPT-Researcher — Autonomous agent that plans sub-queries, searches and scrapes sources, and synthesizes a cited research report.
• deep-research — Minimal agent that runs iterative search-and-reason loops with configurable breadth and depth to produce a report.
• smolagents — Barebones code-acting agent library; its examples include Hugging Face's open reproduction of Deep Research.
• ADAS — Meta-agent that iteratively programs and evaluates new agent designs in code, automating the search over agentic systems.
• DSPy — Define LLM pipelines as modules and optimize their prompts/weights against a metric rather than hand-prompting — the rigor layer for agent pipelines.
• AI-Researcher — Automates the research pipeline from literature analysis through algorithm implementation to paper generation.
• Curie — Experimentation agent that enforces methodological rigor (controlled setup, reproducibility) when running and analyzing experiments.

_{↑ back to top}

Bayesian Optimization & Active Learning

• BoTorch — The reference PyTorch-native Bayesian-optimization library — Monte-Carlo acquisition (qEI/qNEI), multi-objective (qEHVI/qNEHVI), constrained + batch/async BO; the exact stack the Merge JD names.
• Ax (Adaptive Experimentation) — High-level adaptive-experimentation platform over BoTorch — orchestrates sequential DBTL campaigns with a service API, multi-objective goals, and constraints.
• GPyTorch — Scalable Gaussian-process inference in PyTorch — the surrogate engine under BoTorch; core to probabilistic modeling and UQ over sparse, noisy experimental data.
• Trieste — Production-oriented BO toolbox (TensorFlow) with batch, multi-objective, and active-learning loops — a clean reference for closed-loop design patterns.
• scikit-optimize — Lightweight sequential model-based optimization with a scipy-style API — the classic BO baseline (now archived/read-only, cite as legacy).
• HEBO — Heteroscedastic, robust BO that won the NeurIPS 2020 black-box optimization challenge — strong on noisy real-world objectives (bundles MCBO, NAP).
• Optuna — The most-adopted define-by-run black-box optimizer with pruning and dashboards — a strong engineering reference and baseline.
• Mamba 🤏 — Selective state-space architecture (Mamba/Mamba-2) — linear-time long-sequence modeling; the JD's state-space nice-to-have, relevant to neural time-series.
• S4 (Structured State Spaces) 🤏 — The foundational structured state-space sequence models (S4/HiPPO/DSS) — efficient modeling of 10k+ step dependencies in signals.

_{↑ back to top}

Probabilistic Programming & Uncertainty

• Pyro — Deep probabilistic programming on PyTorch (SVI, MCMC) — the JD's named tool for probabilistic modeling and uncertainty quantification.
• NumPyro — JAX-backed Pyro — fast NUTS/HMC and SVI for Bayesian models on GPU/TPU when sampling speed matters.
• Laplace — Post-hoc Laplace approximations for neural nets — cheap Bayesian uncertainty over deep representation-learning models.
• Uncertainty Toolbox — Metrics, calibration, and visualization for predictive uncertainty — validates that a surrogate's error bars are trustworthy.

_{↑ back to top}

ML for Molecules & Bio Design

• GAUCHE — A Gaussian-process library for chemistry (30+ molecule/reaction/protein kernels) that plugs into GPyTorch/BoTorch — the most on-point repo for BO over molecules.
• RDKit — The standard cheminformatics toolkit (descriptors, fingerprints, conformers) — the featurization layer for any molecular surrogate.
• REINVENT 4 — RL-based de novo molecular design (scaffold hopping, R-group, linker) — the canonical 'RL frameworks guiding molecular engineering' reference from AstraZeneca.
• Chemprop — Message-passing neural networks for molecular property prediction — a leading learned surrogate for objective functions in a DBTL loop.
• DeepChem — Batteries-included deep learning for drug discovery, materials, and quantum chemistry — featurizers, models, and dataset splitters.
• Therapeutics Data Commons (PyTDC) — Curated therapeutic ML datasets and benchmarks with a Python API — ready-made sparse/noisy bio tasks for prototyping active-learning loops.
• SELFIES — A 100%-robust molecular string representation — enables valid generative and latent-space optimization of molecules.
• Olympus — Benchmark surfaces and planners for experiment planning / self-driving labs — simulate a closed loop without wet-lab cost.
• Atlas — A 'brain for self-driving labs' — BO over categorical/discrete/mixed spaces with multi-objective and constraints; tightly matched to the JD.
• ESM (EvolutionaryScale) — The current ESM3/ESMC generative protein language models — sequence representation learning for protein design campaigns.
• BioNeMo Framework — NVIDIA's scalable framework for biomolecular foundation models (MolMIM latent generative model, ESM-2) — property-guided molecule optimization at scale.

_{↑ back to top}

Neuroscience & Brain-Computer Interfaces

• MNE-Python — The standard toolkit for analyzing human neurophysiology data (MEG, EEG, iEEG) — the data layer for non-invasive brain-signal modeling.
• Braindecode — Deep learning for EEG/neural-signal decoding on PyTorch — models, datasets, and training loops for brain-computer-interface decoding.
• Neural Latents Benchmark (nlb_tools) — Tools for the Neural Latents Benchmark — evaluating latent-variable models of neural population dynamics; the eval harness for neural representation learning.

_{↑ back to top}

---

🎓 Courses

Foundations & From-Scratch

• Neural Networks: Zero to Hero — Eureka Labs · Andrej Karpathy (2023) · free — Code-along video series building neural nets from backprop up to a GPT, following Attention Is All You Need and GPT-2/3.
• CS224N: NLP with Deep Learning — Stanford · Christopher Manning (2024) · free — Foundational NLP-with-deep-learning course covering word vectors, attention, transformers, and pretraining; lecture videos are public.
• CS 11-711: Advanced NLP — Carnegie Mellon University · Graham Neubig (2024) · free — Graduate NLP course rebuilt around LLMs, including a build-your-own-LLaMa assignment; slides and videos are public.
• 6.S191: Introduction to Deep Learning — MIT · Alexander Amini, Ava Soleimany (2025) · free — Fast-paced intro to deep learning with labs, now including large language models and generative AI.
• The Full Stack LLM Bootcamp — Full Stack Deep Learning · Charles Frye, Sergey Karayev, Josh Tobin (2023) · free — Recorded bootcamp on building LLM applications: prompt engineering, LLMOps, augmented models, and shipping an app.
• Generative AI with Large Language Models — DeepLearning.AI & AWS · Antje Barth, Chris Fregly, et al. (2023) · free — Covers the LLM lifecycle: pretraining, scaling laws, instruction tuning, and RLHF (free to audit on Coursera).

_{↑ back to top}

Pre-training

• CS336: Language Modeling from Scratch — Stanford · Percy Liang, Tatsunori Hashimoto (2025) · free — Implementation-heavy course that builds a language model end to end: tokenization, transformer, training, systems, scaling, data, and alignment.
• LLM101n: Let's build a Storyteller — Eureka Labs · Andrej Karpathy (2024) · free — Public syllabus/repo (in development) for building a Storyteller LLM end to end in Python, C, and CUDA.

_{↑ back to top}

Post-training & Alignment

• Post-training of LLMs — DeepLearning.AI · Banghua Zhu (2025) · free — When and how to apply SFT, DPO, and online RL, including data curation for post-training.
• Fine-tuning & RL for LLMs: Intro to Post-training — DeepLearning.AI (with AMD) · Sharon Zhou (2025) · free — Covers fine-tuning, reward modeling, RLHF, and RL algorithms (PPO, GRPO) for shaping behavior and reasoning.

_{↑ back to top}

Fine-tuning

• Hugging Face LLM Course — Hugging Face · Hugging Face team (2024) · free — Free hands-on course on transformers, tokenizers, fine-tuning pretrained models, and building LLM applications.

_{↑ back to top}

Reinforcement Learning

• CS234: Reinforcement Learning — Stanford · Emma Brunskill (2024) · free — Graduate RL course spanning tabular methods, deep RL, policy gradients, and the basics of RL from human feedback.
• Deep Reinforcement Learning Course — Hugging Face · Thomas Simonini (2023) · free — Free self-paced Deep RL course with practical training in Stable-Baselines3, CleanRL, and Sample Factory; optional certificate.
• Reinforcement Learning from Human Feedback — DeepLearning.AI · Nikita Namjoshi (2024) · free — Short course on the RLHF pipeline, tuning an open model with reward and preference data.
• Reinforcement Fine-Tuning LLMs with GRPO — DeepLearning.AI (with Predibase) · Travis Addair, Arnav Garg (2025) · free — Short course on using GRPO with programmable reward functions to improve LLM reasoning.

_{↑ back to top}

Agents & Applications

• CS294/194-196: Large Language Model Agents — UC Berkeley · Dawn Song, Xinyun Chen (2024) · free — MOOC-available course on LLM agent foundations, reasoning, tool use, and applications, with frontier-lab guest lectures.
• CS294/194-280: Advanced LLM Agents — UC Berkeley · Dawn Song, Xinyun Chen (2025) · free — Spring 2025 follow-on covering advanced agent reasoning, math/theorem-proving, code generation, and safety.

_{↑ back to top}

Computer Vision

• CS231n: Deep Learning for Computer Vision — Stanford · Fei-Fei Li, Ehsan Adeli et al. (2024) · free — The canonical intro to CNNs and visual recognition; slides, notes, and assignments are public.
• CS231A: Computer Vision, From 3D Reconstruction to Recognition — Stanford · Silvio Savarese / Jeannette Bohg (staff) (2025) · free — Geometric CV: camera models, epipolar/stereo geometry, depth and scene flow, 6D pose and tracking, with public notes.
• EECS 498/598: Deep Learning for Computer Vision — University of Michigan · Justin Johnson (2020) · free — From-scratch deep learning for vision (CNNs, attention, detection, segmentation) with the full lecture set on YouTube.
• Community Computer Vision Course — Hugging Face · HF community (2024) · free — Free hands-on course from classical CV through ViTs, multimodal and generative vision, with runnable notebooks.
• DeepRob: Deep Learning for Robot Perception — University of Michigan · Chad Jenkins, Anthony Opipari, Xiaoxiao Du (2024) · free — Deep-learning-for-vision adapted to robot perception and manipulation, then reproducing recent perception papers; public slides.
• Practical Deep Learning for Coders — fast.ai · Jeremy Howard (2022) · free — Code-first, top-down deep learning across vision, NLP, and fine-tuning in PyTorch/fastai; free lessons plus the free book.

_{↑ back to top}

Multi-modal Machine Learning

• 11-777: Multimodal Machine Learning — Carnegie Mellon (LTI) · Louis-Philippe Morency, Paul Liang (2023) · free — Organizes multimodal ML around six challenges (representation, alignment, reasoning, generation, transference, quantification); lectures on YouTube.

_{↑ back to top}

Vision-Language Models

• CS25: Transformers United — Stanford · Student-led seminar (2025) · free — Guest-lecture seminar including multimodal / vision-language and world-modeling sessions; open to audit, posted on YouTube.

_{↑ back to top}

Video & Data Ops

• Getting Started with FiftyOne (Visual AI) — Voxel51 · Voxel51 (2025) · free — Free tutorials on curating, visualizing, and debugging image/video/3D datasets and model outputs — directly relevant to AD/robotics data ops.

_{↑ back to top}

Bayesian Optimization & Gaussian Processes

• Bayesian Optimization (the book) — Washington University in St. Louis · Roman Garnett (2023) · free — The most complete modern treatment of BO theory and practice — full PDF free online.
• Gaussian Process Summer School (GPSS) — University of Sheffield (community) · free — Annual lectures plus notebooks on GPs, deep GPs, and Bayesian optimization; slides and videos archived per year.
• Gaussian Processes for Machine Learning (GPML) — MIT Press · Rasmussen & Williams (2006) · free — The definitive GP textbook (regression, classification, covariance functions, model selection) — free full PDF.
• BoTorch Tutorials + Ax Docs — Meta · free — Runnable walkthroughs of qEI/qNEI/qEHVI/qNEHVI, constrained and multi-objective BO — closest to the day-to-day toolkit for this role.

_{↑ back to top}

Probabilistic & Bayesian ML

• Probabilistic Machine Learning (Books 1 & 2) — Google · Kevin Murphy · free — Free draft PDFs plus Colab notebooks — strong coverage of probabilistic modeling, GPs, and uncertainty quantification.

_{↑ back to top}

ML for Molecules & Science

• DeepChem Tutorials — DeepChem · free — Hands-on notebooks for ML on molecules — featurization, property prediction, and generative design.

_{↑ back to top}

Neuroscience & Brain-Computer Interfaces

• Neuromatch Academy — Computational Neuroscience — Neuromatch · free — Free, hands-on computational-neuroscience curriculum (modeling, dynamics, decoding) — the on-ramp to the neuroscience nice-to-have.

_{↑ back to top}

---

📄 Papers

Small Language Models & Surveys

• SmolLM2: When Smol Goes Big — Data-Centric Training of a Small Language Model (Ben Allal et al., Hugging Face, 2025) · arXiv:2502.02737 — A 1.7B model trained via careful data curation and overtraining, with the dataset recipe documented openly.
• A Survey of Small Language Models (Van Nguyen et al., multi-institution, 2024) · arXiv:2410.20011 — Taxonomy of SLM architectures, training, and compression methods, plus benchmark datasets and evaluation metrics.
• Small Language Models: Survey, Measurements, and Insights (Lu et al., multi-institution, 2024) · arXiv:2409.15790 — Empirical survey measuring capabilities and on-device runtime cost across a large set of released SLMs.
• Textbooks Are All You Need (phi-1) (Gunasekar et al., Microsoft, 2023) · arXiv:2306.11644 — phi-1 (1.3B), showing high-quality synthetic 'textbook' data lets small models match far larger ones on HumanEval.
• Textbooks Are All You Need II: phi-1.5 Technical Report (Li et al., Microsoft, 2023) · arXiv:2309.05463 — Extends the textbook-data approach to a 1.3B general reasoning model competitive with models 5x its size.
• Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone (Abdin et al., Microsoft, 2024) · arXiv:2404.14219 — phi-3-mini (3.8B) rivals Mixtral 8x7B and GPT-3.5 on benchmarks while being small enough to run on a phone.
• Gemma: Open Models Based on Gemini Research and Technology (Mesnard et al., Google DeepMind, 2024) · arXiv:2403.08295 — Open 2B and 7B models derived from Gemini research, released with pretrained and instruction-tuned checkpoints.
• Gemma 2: Improving Open Language Models at a Practical Size (Gemma Team, Google DeepMind, 2024) · arXiv:2408.00118 — Trains the 2B/9B models with knowledge distillation over next-token prediction for strong quality at small size.
• MobileLLM: Optimizing Sub-billion Parameter Language Models for On-Device Use Cases (Liu et al., Meta, 2024) · arXiv:2402.14905 (ICML 2024) — Shows deep-and-thin architecture, embedding sharing, and grouped-query attention matter most below 1B parameters.
• TinyLlama: An Open-Source Small Language Model (Zhang et al., SUTD, 2024) · arXiv:2401.02385 — A 1.1B Llama-architecture model pretrained on ~3T tokens with a fully open training pipeline.
• TinyStories: How Small Can Language Models Be and Still Speak Coherent English? (Eldan & Li, Microsoft Research, 2023) · arXiv:2305.07759 — Sub-10M-parameter models trained on a constrained synthetic story corpus can generate coherent, consistent English.
• Distilling Step-by-Step! Outperforming Larger LMs with Less Data and Smaller Sizes (Hsieh et al., Google, 2023) · arXiv:2305.02301 (Findings of ACL 2023) — Uses LLM-generated rationales as extra supervision so small models beat much larger ones with less data.

_{↑ back to top}

Pre-training & Data

• The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale (Penedo et al., Hugging Face, 2024) · arXiv:2406.17557 (NeurIPS 2024) — Ablates deduplication and filtering to build a 15T-token open web corpus plus the FineWeb-Edu subset.
• The Llama 3 Herd of Models (Dubey et al., Meta, 2024) · arXiv:2407.21783 — Documents the pretraining, scaling, and post-training of the Llama 3 family, including the 405B dense flagship.
• GPT-4 Technical Report (OpenAI, OpenAI, 2023) · arXiv:2303.08774 — A multimodal transformer with human-level exam performance and predictable scaling from small proxy models.
• DeepSeek-V3 Technical Report (DeepSeek-AI, DeepSeek, 2024) · arXiv:2412.19437 — A 671B MoE (37B active) with MLA, auxiliary-loss-free load balancing, and multi-token prediction, trained efficiently.

_{↑ back to top}

Scaling Laws

• Scaling Laws for Neural Language Models (Kaplan et al., OpenAI, 2020) · arXiv:2001.08361 — Establishes power-law relationships between loss and model size, data, and compute across many orders of magnitude.
• Training Compute-Optimal Large Language Models (Chinchilla) (Hoffmann et al., DeepMind, 2022) · arXiv:2203.15556 (NeurIPS 2022) — The Chinchilla result: model size and training tokens should scale equally; most large models are undertrained.

_{↑ back to top}

Post-training & Alignment

• Training Language Models to Follow Instructions with Human Feedback (InstructGPT) (Ouyang et al., OpenAI, 2022) · arXiv:2203.02155 (NeurIPS 2022) — Introduces the SFT + reward model + RLHF recipe; a 1.3B tuned model was preferred over 175B GPT-3.
• Direct Preference Optimization: Your Language Model Is Secretly a Reward Model (Rafailov et al., Stanford, 2023) · arXiv:2305.18290 (NeurIPS 2023) — Replaces the RLHF reward model and PPO loop with a single classification loss on preference pairs.
• Constitutional AI: Harmlessness from AI Feedback (Bai et al., Anthropic, 2022) · arXiv:2212.08073 — Trains a harmless assistant using AI-generated critiques and preferences guided by written principles (RLAIF).
• RLAIF: Scaling RLHF with AI Feedback (Lee et al., Google, 2023) · arXiv:2309.00267 (ICML 2024) — Shows LLM-generated preference labels can match human-labeled RLHF across summarization and dialogue tasks.
• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Wei et al., Google, 2022) · arXiv:2201.11903 (NeurIPS 2022) — Intermediate reasoning steps in prompts unlock arithmetic, commonsense, and symbolic reasoning at scale.

_{↑ back to top}

RL & Reasoning

• Training a Helpful and Harmless Assistant with RLHF (Bai et al., Anthropic, 2022) · arXiv:2204.05862 — Applies preference modeling and iterated online RLHF to align an assistant, analyzing the reward/KL trade-off.
• DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (Shao et al., DeepSeek, 2024) · arXiv:2402.03300 — Introduces GRPO, a critic-free RL algorithm using group-relative advantages, later central to DeepSeek-R1.
• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning (DeepSeek-AI, DeepSeek, 2025) · arXiv:2501.12948 (Nature, 2025) — Elicits reasoning purely via RL (R1-Zero) and distills it into smaller dense models — a key SLM-reasoning recipe.
• MORPHEUS: A Persistent Enterprise Simulation for Continual Reinforcement Learning (Skyfall AI (+ Toronto / Georgia Tech / Alberta), Skyfall AI, 2026) · OpenReview 2026 — Big-world CRL benchmark with structured non-stationarity; PPO/HER/EWC/LCM all show a ~1.0 settled-state gap — no family durably learns without resets.
• The Landscape of Agentic Reinforcement Learning for LLMs: A Survey (multi-institution, multi-institution, 2025) · arXiv:2509.02547 — Map of agentic RL for LLMs — environments, verifiable rewards, multi-turn credit assignment, and the open frameworks implementing them.
• Reinforcement Learning Foundations for Deep Research Systems: A Survey (multi-institution, multi-institution, 2025) · arXiv:2509.06733 — How RL post-training underpins long-horizon research agents: reward design, tool-use credit assignment, and training infrastructure.

_{↑ back to top}

Parameter-Efficient Fine-tuning

• LoRA: Low-Rank Adaptation of Large Language Models (Hu et al., Microsoft, 2021) · arXiv:2106.09685 (ICLR 2022) — Freezes base weights and trains injected low-rank matrices, cutting trainable parameters by orders of magnitude.
• QLoRA: Efficient Finetuning of Quantized LLMs (Dettmers et al., University of Washington, 2023) · arXiv:2305.14314 (NeurIPS 2023) — Backpropagates through a frozen 4-bit (NF4) model into LoRA adapters, finetuning a 65B model on one 48GB GPU.

_{↑ back to top}

Distillation & Compression

• Distilling the Knowledge in a Neural Network (Hinton et al., Google, 2015) · arXiv:1503.02531 — Foundational knowledge distillation: training a small student on the soft targets of a larger teacher.
• Sequence-Level Knowledge Distillation (Kim & Rush, Harvard, 2016) · arXiv:1606.07947 (EMNLP 2016) — Extends distillation from token-level to sequence-level, yielding student NMT models ~10x faster with little loss.
• DistilBERT: Smaller, Faster, Cheaper and Lighter (Sanh et al., Hugging Face, 2019) · arXiv:1910.01108 — Distills BERT during pretraining to a 40% smaller model that retains ~97% of language understanding.
• GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers (Frantar et al., IST Austria, 2022) · arXiv:2210.17323 (ICLR 2023) — One-shot 3-4 bit weight quantization using approximate second-order information, no retraining required.
• AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration (Lin et al., MIT, 2023) · arXiv:2306.00978 (MLSys 2024) — Protects a small fraction of salient weights via activation-aware scaling for accurate low-bit quantization.
• LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale (Dettmers et al., University of Washington, 2022) · arXiv:2208.07339 (NeurIPS 2022) — Int8 inference for billion-scale transformers with no accuracy loss by isolating emergent outlier features.
• SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot (Frantar & Alistarh, IST Austria, 2023) · arXiv:2301.00774 (ICML 2023) — Prunes GPT-scale models to 50-60% sparsity in one shot without retraining and minimal perplexity increase.

_{↑ back to top}

Vision-Language Models

• Learning Transferable Visual Models From Natural Language Supervision (CLIP) (Radford et al., OpenAI, 2021) · ICML 2021 · arXiv:2103.00020 — Contrastively pretrains dual image/text encoders on 400M web pairs, enabling zero-shot classification and the embedding backbone most VLMs still build on.
• Flamingo: a Visual Language Model for Few-Shot Learning (Alayrac et al., DeepMind, 2022) · NeurIPS 2022 · arXiv:2204.14198 — Bridges frozen vision and language backbones with gated cross-attention over interleaved image/video-text for in-context few-shot multimodal learning.
• BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Encoders and LLMs (Li et al., Salesforce Research, 2023) · ICML 2023 · arXiv:2301.12597 — Introduces the lightweight Q-Former connecting a frozen vision encoder to a frozen LLM — a cheap recipe for adding vision to language models.
• Visual Instruction Tuning (LLaVA) (Liu et al., UW-Madison / Microsoft Research, 2023) · NeurIPS 2023 · arXiv:2304.08485 — Uses GPT-4-generated image-instruction data to instruction-tune a CLIP-encoder-plus-LLM model, establishing the dominant open VLM training pattern.
• Improved Baselines with Visual Instruction Tuning (LLaVA-1.5) (Liu et al., UW-Madison / Microsoft Research, 2023) · CVPR 2024 · arXiv:2310.03744 — MLP connector, higher-res CLIP, and academic VQA data hit SOTA on 11 benchmarks with ~1.2M public samples and one day on 8×A100.
• Qwen2-VL: Perception of the World at Any Resolution (Wang et al. (Qwen Team), Alibaba, 2024) · arXiv:2409.12191 — Adds naive dynamic resolution and multimodal RoPE (M-RoPE) to handle arbitrary image sizes and video under one paradigm across 2B/8B/72B.
• Qwen2.5-VL Technical Report (Qwen Team, Alibaba, 2025) · arXiv:2502.13923 — Absolute-time encoding for hour-long video, strong bbox/point grounding and document parsing — a strong fine-tuning base for driving/robotics.
• InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks (Chen et al., Shanghai AI Lab / OpenGVLab, 2023) · CVPR 2024 · arXiv:2312.14238 — Scales the vision encoder to 6B params and progressively aligns it to an LLM, reaching SOTA across 32 image/video perception and retrieval benchmarks.
• BARE: Leveraging Base Language Models for Few-Shot Synthetic Data Generation (Zhu, Asawa, Davis, Chen, Hanin, Stoica, Gonzalez, Zaharia, UC Berkeley / Stanford / Princeton, 2025) · arXiv:2502.01697 — Base models for diversity, instruction-tuned models for correctness — the two-stage recipe the syndata case study adapts to vision-language data; operationalized in skills/syndata-bare.

_{↑ back to top}

Video Understanding

• Video-LLaMA: Instruction-tuned Audio-Visual Language Model for Video Understanding (Zhang, Li, Bing, DAMO Academy, Alibaba, 2023) · EMNLP 2023 Demo · arXiv:2306.02858 — Couples a video Q-Former and audio branch to an LLM to instruction-tune joint audio-visual temporal understanding of video.
• Video-LLaVA: Learning United Visual Representation by Alignment Before Projection (Lin et al., Peking University, 2023) · EMNLP 2024 · arXiv:2311.10122 — Aligns image and video features into a shared space before the LLM projector, letting one model train jointly on images and video with mutual gains.
• ViViT: A Video Vision Transformer (Arnab et al., Google Research, 2021) · ICCV 2021 · arXiv:2103.15691 — Pure-transformer video classifier with factorized spatial/temporal attention variants to keep long token sequences tractable.
• VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training (Tong et al., Nanjing University / Tencent, 2022) · NeurIPS 2022 · arXiv:2203.12602 — 90-95% tube masking makes MAE-style self-supervised video pretraining work even on a few thousand clips — a practical video backbone recipe.
• Is Space-Time Attention All You Need for Video Understanding? (TimeSformer) (Bertasius, Wang, Torresani, Facebook AI / Dartmouth, 2021) · ICML 2021 · arXiv:2102.05095 — Convolution-free video model; divided space-then-time attention gives the best accuracy/cost trade-off for action recognition.
• InternVideo: General Video Foundation Models via Generative and Discriminative Learning (Wang et al., Shanghai AI Lab / OpenGVLab, 2022) · arXiv:2212.03191 — Combines masked video modeling with video-language contrastive learning into one foundation model hitting SOTA on 39 video datasets.

_{↑ back to top}

Multi-modal & Grounding

• Grounding DINO: Marrying DINO with Grounded Pre-Training for Open-Set Detection (Liu et al., IDEA Research, 2023) · ECCV 2024 · arXiv:2303.05499 — Fuses a DINO detector with language so free-text prompts localize arbitrary objects — a workhorse for open-vocabulary grounding/auto-labeling.

_{↑ back to top}

Retrieval & Embeddings

• Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision (ALIGN) (Jia et al., Google, 2021) · ICML 2021 · arXiv:2102.05918 — A simple dual-encoder trained on 1B+ noisy alt-text pairs (no cleaning) sets SOTA image-text retrieval, validating scale over curation.
• Sigmoid Loss for Language Image Pre-Training (SigLIP) (Zhai et al., Google DeepMind, 2023) · ICCV 2023 · arXiv:2303.15343 — A pairwise sigmoid loss needing no global normalization improves small-batch training and large-scale embedding learning.

_{↑ back to top}

Multi-modal Evaluation

• LMMs-Eval: Reality Check on the Evaluation of Large Multimodal Models (Zhang, Li, Liu et al., NTU (LMMs-Lab), 2024) · arXiv:2407.12772 — Unified 50+-task evaluation framework (plus a Lite variant and contamination-resistant LiveBench) — the harness to standardize your own VLM/video evals.
• Video-MME: Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis (Fu et al., multi-institution, 2024) · CVPR 2025 · arXiv:2405.21075 — Full-spectrum video benchmark spanning 11s-to-1h clips across 6 domains with subtitle/audio modalities, the standard for long-video VLM eval.
• MMBench: Is Your Multi-modal Model an All-around Player? (Liu et al., Shanghai AI Lab / CUHK / NTU, 2023) · ECCV 2024 · arXiv:2307.06281 — Fine-grained bilingual VLM benchmark with a CircularEval protocol that shuffles answer order to reduce position bias in scoring.
• MMMU: Massive Multi-discipline Multimodal Understanding and Reasoning Benchmark (Yue et al., multi-institution, 2023) · CVPR 2024 · arXiv:2311.16502 — 11.5K college-level questions over 30 subjects and heterogeneous image types, probing knowledge-heavy expert reasoning where even strong VLMs score low.
• MLAgentBench: Evaluating Language Agents on Machine Learning Experimentation (Huang et al., Stanford, 2023) · arXiv:2310.03302 — 13 ML experimentation tasks and metrics for measuring agents that iteratively edit code to improve model performance.

_{↑ back to top}

Autonomous Research & AI Scientists

• The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery (Lu et al., Sakana AI, 2024) · arXiv:2408.06292 — A pipeline where an LLM generates ideas, runs code experiments, and writes and reviews papers within a fixed compute budget.
• The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search (Yamada et al., Sakana AI, 2025) · arXiv:2504.08066 — Agentic tree search plus VLM figure review; one generated manuscript passed workshop peer review above the human acceptance bar.
• Towards an AI Co-Scientist (Gottweis et al., Google / DeepMind, 2025) · arXiv:2502.18864 — Multi-agent system on Gemini that generates and debates hypotheses, validated on drug repurposing and antimicrobial-resistance tasks.
• ResearchAgent: Iterative Research Idea Generation over Scientific Literature (Baek et al., KAIST / Microsoft Research, 2024) · NAACL 2025 · arXiv:2404.07738 — Generates problems, methods, and experiment designs from literature, refined by reviewing agents with human-aligned feedback.
• Autonomous chemical research with large language models (Coscientist) (Boiko et al., Carnegie Mellon University, 2023) · Nature 624, 570-578 — A GPT-4 system with search/code/lab-automation tools that designs and physically executes chemistry experiments.
• Automated Design of Agentic Systems (ADAS) (Hu, Lu, Clune, UBC / Vector Institute, 2024) · ICLR 2025 · arXiv:2408.08435 — Frames agent design as a search problem where a meta-agent writes code to invent new agents, outperforming hand-designed baselines.
• Assisting in Writing Wikipedia-like Articles From Scratch with LLMs (STORM) (Shao et al., Stanford OVAL, 2024) · NAACL 2024 · arXiv:2402.14207 — Simulate multi-perspective question-asking to build an outline, then retrieve sources and write a grounded, cited article.
• A Survey on Large Language Model based Autonomous Agents (Wang et al., Renmin University of China, 2023) · arXiv:2308.11432 — Systematic survey of LLM-agent construction, application domains, and evaluation — context for research-loop agents.
• Gödel Agent: A Self-Referential Agent Framework for Recursive Self-Improvement (Yin et al., Peking University / UC Santa Barbara, 2025) · ACL 2025 · arXiv:2410.04444 — An agent that reads and rewrites its own logic at runtime to recursively self-improve without a fixed optimization routine.
• Agent Laboratory: Using LLM Agents as Research Assistants (Schmidgall et al., AMD / Johns Hopkins University, 2025) · arXiv:2501.04227 — Autonomous framework covering literature review, experimentation, and report writing, with human feedback improving output quality.
• Curie: Toward Rigorous and Automated Scientific Experimentation with AI Agents (Kon et al., University of Michigan, 2025) · arXiv:2502.16069 — Adds rigor components — controlled procedure, reproducibility, result interpretation — to agent-run experimentation.
• AI-Researcher: Autonomous Scientific Innovation (Tang et al., University of Hong Kong, 2025) · arXiv:2505.18705 — A fully autonomous system spanning idea generation, algorithm implementation, and manuscript writing, released with an open codebase.
• Robin: A Multi-Agent System for Automating Scientific Discovery (Ghareeb et al., FutureHouse, 2025) · arXiv:2505.13400 — Couples literature-search and data-analysis agents to run a discovery loop; identified ripasudil as a candidate for dry AMD.

_{↑ back to top}

Bayesian Optimization & Active Learning

• A Tutorial on Bayesian Optimization (Peter Frazier, Cornell, 2018) · arXiv:1807.02811 — The single best self-contained intro — GP regression plus EI, entropy search, and knowledge gradient. Read this first.
• Efficient Global Optimization of Expensive Black-Box Functions (EGO) (Jones, Schonlau & Welch, —, 1998) · J. Global Optimization 13:455 — The paper that launched surrogate-based BO: a Kriging/GP surrogate plus Expected Improvement to optimize costly black boxes.
• Practical Bayesian Optimization of Machine Learning Algorithms (Snoek, Larochelle & Adams, Toronto / Harvard, 2012) · NeurIPS · arXiv:1206.2944 — The modern reboot of BO for ML (the 'Spearmint' paper): GP priors, integrated acquisition, and cost-aware EI.
• Taking the Human Out of the Loop: A Review of Bayesian Optimization (Shahriari, Swersky, Wang, Adams & de Freitas, multi-institution, 2016) · Proc. IEEE 104(1) — The canonical survey tying together surrogates, acquisition functions, and applications of BO.
• Gaussian Process Optimization in the Bandit Setting (GP-UCB) (Srinivas, Krause, Kakade & Seeger, Caltech / ETH, 2010) · ICML · arXiv:0912.3995 — Introduces GP-UCB with the first sublinear cumulative-regret bounds — the theoretical backbone of exploration in BO.
• Max-value Entropy Search for Efficient Bayesian Optimization (MES) (Wang & Jegelka, MIT, 2017) · ICML · arXiv:1703.01968 — A cheaper, more robust information-theoretic acquisition that targets info about the optimal value rather than the arg-max.
• BoTorch: A Framework for Efficient Monte-Carlo Bayesian Optimization (Balandat, Karrer, Jiang, Daulton, Letham, Wilson & Bakshy, Meta, 2020) · NeurIPS · arXiv:1910.06403 — The reference open-source BO library — Monte-Carlo acquisition (qEI/qNEI) with autodiff and sample-average approximation.
• Constrained Bayesian Optimization with Noisy Experiments (qNEI) (Letham, Karrer, Ottoni & Bakshy, Meta, 2019) · Bayesian Analysis 14(2) · arXiv:1706.07094 — Noisy, constrained expected improvement via quasi-Monte-Carlo integration — the basis of Meta's Ax and directly on-point for noisy DBTL assays.
• Differentiable Expected Hypervolume Improvement for Multi-Objective BO (qEHVI) (Daulton, Balandat & Bakshy, Meta, 2020) · NeurIPS · arXiv:2006.05078 — Exact-gradient, parallelizable hypervolume-improvement acquisition — state of the art for multi-objective BO.
• Parallel BO of Multiple Noisy Objectives with EHVI (qNEHVI) (Daulton, Balandat & Bakshy, Meta, 2021) · NeurIPS · arXiv:2105.08195 — Extends qEHVI to noisy observations with polynomial batch scaling — the current default for noisy multi-objective DBTL problems.
• Scalable Global Optimization via Local Bayesian Optimization (TuRBO) (Eriksson, Pearce, Gardner, Turner & Poloczek, Uber AI, 2019) · NeurIPS · arXiv:1910.01739 — Trust-region local GPs with bandit sample allocation — the workhorse for high-dimensional, high-budget BO.
• High-Dimensional BO with Sparse Axis-Aligned Subspaces (SAASBO) (Eriksson & Jankowiak, Meta, 2021) · UAI · arXiv:2103.00349 — Strong sparsity priors plus HMC to find the few relevant dimensions — excellent in the low-data, high-dim regime typical of experiments.
• Active Learning Literature Survey (Burr Settles, UW–Madison, 2009) · CS Tech Report 1648 — The canonical survey of query strategies — uncertainty sampling, query-by-committee, expected error reduction.
• Bayesian Active Learning for Classification and Preference Learning (BALD) (Houlsby, Huszár, Ghahramani & Lengyel, Cambridge, 2011) · arXiv:1112.5745 — Introduces BALD (maximize mutual information / model disagreement) — foundational for modern deep and Bayesian active learning.
• A Tutorial on Thompson Sampling (Russo, Van Roy, Kazerouni, Osband & Wen, Stanford / DeepMind, 2018) · FnT ML 11(1) · arXiv:1707.02038 — The definitive reference for posterior-sampling exploration, widely used as a scalable BO acquisition strategy.

_{↑ back to top}

Probabilistic Modeling & Uncertainty

• Predictive Entropy Search for Efficient Global Optimization (PES) (Hernández-Lobato, Hoffman & Ghahramani, Cambridge / Harvard, 2014) · NeurIPS · arXiv:1406.2541 — Information-theoretic acquisition that picks points maximizing information gain about the location of the optimum.

_{↑ back to top}

Representation Learning

• Deep Kernel Learning (Wilson, Hu, Salakhutdinov & Xing, CMU, 2016) · AISTATS · arXiv:1511.02222 — Combines neural-net feature learning with GP kernels — representation learning that preserves calibrated GP uncertainty.

_{↑ back to top}

ML for Molecular & Bio Design

• Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules (Gómez-Bombarelli et al., Harvard / Toronto, 2018) · ACS Central Science · arXiv:1610.02415 — The chemical VAE: a continuous latent space of molecules enabling gradient-based and Bayesian optimization of properties.
• GAUCHE: A Library for Gaussian Processes in Chemistry (Griffiths et al., multi-institution, 2023) · NeurIPS · arXiv:2212.04450 — GPs and BO over molecular graphs, strings, fingerprints, reactions, and proteins — interfaces directly with GPyTorch/BoTorch.
• Olympus: A Benchmarking Framework for Noisy Optimization and Experiment Planning (Häse, Aldeghi & Aspuru-Guzik, Toronto, 2021) · Mach. Learn.: Sci. Technol. 2(3) · arXiv:2010.04153 — Realistic noisy experiment-planning benchmarks for autonomous / self-driving labs — a proving ground for closed-loop optimizers.
• Machine-Learning-Guided Directed Evolution for Protein Engineering (Yang, Wu & Arnold, Caltech, 2019) · Nature Methods 16:687 · arXiv:1811.10775 — The reference framework for ML-in-the-loop directed evolution — sequence–function models guiding the DBTL loop for proteins.

_{↑ back to top}

Neuroscience & Brain-Computer Interfaces

• Inferring Single-Trial Neural Population Dynamics Using Sequential Auto-Encoders (LFADS) (Pandarinath et al., Stanford / Emory, 2018) · Nature Methods 15:805 · arXiv:1608.06315 — Latent dynamical-systems models of neural population activity — foundational for decoding brain signals and a bridge from state-space models to neuroscience.

_{↑ back to top}

State-Space & Sequence Models

• Efficiently Modeling Long Sequences with Structured State Spaces (S4) (Gu, Goel & Ré, Stanford, 2022) · ICLR · arXiv:2111.00396 — A structured SSM parameterization that handles 10k+ step dependencies efficiently — relevant to neural time-series.
• Mamba: Linear-Time Sequence Modeling with Selective State Spaces (Gu & Dao, CMU / Princeton, 2023) · arXiv:2312.00752 — Selective (input-dependent) SSM matching Transformer quality with linear-time scaling — the modern state-space model to know.

_{↑ back to top}

More

• MemGPT: Towards LLMs as Operating Systems (Packer, Wooders, Lin, Fang, Patil, Stoica, Gonzalez, UC Berkeley, 2023) · arXiv:2310.08560 — Virtual context management — OS-style hierarchical memory for LLMs; the canonical per-user persistent-memory design behind personal models. Operationalized in skills/personalization-loop.

_{↑ back to top}

---

📚 Reading Lists

AI — Deep-Learning Canon → Physical AI → AGI/ASI

• Frontier AI Reading List — deep-learning canon → Physical AI → AGI/ASI (2026-07-30) — 12 seminal works in Statement/Quote/Evidence/Actions/Patterns/1st-Principle format across two layers (compute & scaling canon; world models & VLA robotics); all 15 links verified, 2 fixed at intake; Layer 3 (AGI/ASI safety & alignment) pending.

_{↑ back to top}

AI-Enabled Research — AI × Math & Science

• AI for Math & Science — assisted and autonomous problem solving (2026-07-30) — AI-enabled research: theorem proving and math discovery case studies (AlphaEvolve, Nesterov convergence, Bruhat-interval hypercubes) in the Statement/Quote/Evidence/Actions/Patterns/1st-Principle format; Layer 1 landed, Layers 2-3 (autonomous science, benchmarks) pending.

_{↑ back to top}

Track Index & Contract

• Reading-list track index — math · physics · CS · bio-biomed-biotech (planned) — The track contract (trackable dated docs, every link verified, entries earn their place, honest gaps) and the roadmap for the math, physics, computer-science, and biology lists.

_{↑ back to top}

---

🏛️ Labs & Platforms

Frontier Labs

• Thinking Machines Lab — Frontier lab (Mira Murati); makes Tinker (LoRA fine-tuning API) and the open-weights Inkling multimodal MoE. · github

_{↑ back to top}

Fine-tuning Platforms

• Unsloth AI — 2x-faster, ~70%-less-VRAM LoRA/QLoRA fine-tuning for small models on a single GPU. · github
• Adaptive ML — RLOps platform (Adaptive Engine) for reinforcement-learning post-training + evaluation of open models on enterprise tasks. · github

_{↑ back to top}

Inference & Serving

• Together AI — Training + high-throughput inference cloud for open models; an Inkling deployment partner. · github
• Fireworks AI — Fast, low-cost inference + fine-tuning for open models; an Inkling deployment partner. · github
• Baseten — ML infrastructure for developers (Truss packaging + autoscaling inference); an Inkling deployment partner. · github
• Inferact — Commercializing vLLM as a universal inference layer ($150M seed); mission to make inference cheaper and faster. · github
• RadixArk — Commercializing SGLang (RadixAttention) as an optimization service; the SGLang counterpart to Inferact's vLLM. · github

_{↑ back to top}

Compute & Infra

• Modal — Serverless GPU cloud — run training/inference code in the cloud with no infra management; an Inkling deployment partner. · github
• CoreWeave — GPU cloud with a managed RL-as-a-service offering (serverless infra, frameworks, APIs) for post-training agentic models. · github

_{↑ back to top}

Data & Lakehouse

• Databricks — Data + AI lakehouse (Mosaic AI training, DBRX open model, MLflow); an Inkling deployment partner. · github
• HUD — RL-environment platform for enterprise workflows — standardized, reproducible, closed-loop-training-ready environments and benchmarks.

_{↑ back to top}

Open Research

• LightSeek — Foundation accelerating open research and open-source innovation for next-generation AI systems. · github
• Skyfall AI — Enterprise RL company (ex-Maluuba founders); ships MORPHEUS, a persistent, non-resetting enterprise simulation that argues today's LLMs don't durably learn without continual RL. · github

_{↑ back to top}

Bio × AI / Closed-Loop Science Labs

• Merge Labs — Frontier BCI lab bridging biological and artificial intelligence; building a closed-loop DBTL optimization backbone for molecular + neuroengineering discovery (the hiring target).
• Meta — Adaptive Experimentation (BoTorch / Ax) — Builds and maintains the core PyTorch Bayesian-optimization + adaptive-experimentation stack this role uses. · github
• Matter Lab (Aspuru-Guzik) — Self-driving laboratories — robotics + ML + BO for autonomous molecular and materials discovery (SELFIES, Olympus, Atlas). · github
• Acceleration Consortium — Global consortium standardizing self-driving labs and materials-acceleration platforms for closed-loop discovery.
• Coley Research Group (MIT) — ML for synthesis planning, molecular design, and autonomous experimentation / DBTL workflows. · github
• Recursion — Industrial-scale closed-loop drug discovery combining automated wet-lab data generation with ML.
• Isomorphic Labs — Alphabet/DeepMind spinout applying AI (AlphaFold lineage) to drug design.
• NVIDIA BioNeMo — Foundation-model platform plus generative models (MolMIM) for molecular and protein design at scale. · github

_{↑ back to top}

---

🧑‍🔬 People to Follow

Bayesian Optimization & Gaussian Processes

• Peter Frazier — Cornell ORIE · Uber — Author of the standard BO tutorial and the knowledge-gradient acquisition; industrial BO at scale.
• Roman Garnett — Washington University in St. Louis — Author of the definitive open textbook Bayesian Optimization (bayesoptbook.com).
• Max Balandat — Meta — Adaptive Experimentation — Co-creator of BoTorch; Monte-Carlo acquisition (qEI) and multi-objective qEHVI/qNEHVI.
• Eytan Bakshy — Meta — Adaptive Experimentation lead — Drives BoTorch/Ax; noisy, constrained, multi-objective BO in production.
• David Eriksson — Meta — Adaptive Experimentation — TuRBO and SAASBO — the go-to methods for high-dimensional Bayesian optimization.
• Samuel Daulton — Meta · University of Oxford — Lead author of qEHVI and qNEHVI — state-of-the-art multi-objective BO for noisy DBTL problems.

_{↑ back to top}

ML for Molecules & Self-Driving Labs

• Alán Aspuru-Guzik — University of Toronto · Acceleration Consortium — Leading figure in self-driving laboratories and closed-loop molecular discovery (SELFIES, Olympus, Atlas).
• Connor Coley — MIT (ChemE + EECS) — Autonomous discovery, ML for synthesis planning, and DBTL molecular-design workflows.
• José Miguel Hernández-Lobato — University of Cambridge — Predictive Entropy Search, molecular VAEs, and Bayesian deep learning for molecule design.
• Frances Arnold — Caltech (Nobel Laureate, 2018) — Directed evolution and ML-guided protein engineering — the DBTL-for-proteins vision.
• Kevin K. Yang — Microsoft Research New England — ML-guided directed evolution and protein representation learning.
• David Duvenaud — University of Toronto — Neural ODEs, differentiable molecular fingerprints, and the chemical VAE.

_{↑ back to top}

Probabilistic ML & Uncertainty

• Jasper Snoek — Google DeepMind — 'Practical Bayesian Optimization' (Spearmint); uncertainty quantification and calibration in deep learning.
• Andrew Gordon Wilson — New York University — Deep kernel learning and scalable Gaussian processes (GPyTorch) — representation learning that keeps GP uncertainty.

_{↑ back to top}

State-Space Models & Neuro-adjacent

• Albert Gu — Carnegie Mellon University — S4 and Mamba — structured/selective state-space models for long-sequence and time-series data.
• Cheng Soon Ong — CSIRO's Data61 · Australian National University — Co-author of Mathematics for Machine Learning; probabilistic modeling foundations.

_{↑ back to top}

---

💼 Jobs & Careers

Frontier Labs (careers pages)

• Anthropic Careers — Frontier lab; alignment, pretraining, RL, and fine-tuning research and engineering roles.
• OpenAI Careers — Frontier lab; research, post-training/RLHF, and applied ML roles.
• Google DeepMind Careers — Frontier lab; research scientist and engineering roles across pretraining, RL, and alignment.
• Mistral AI Careers — European frontier lab; open-weight model pretraining, fine-tuning, and inference roles.
• Hugging Face Careers — Open-source ML platform; ML research/engineering on training, datasets, and model deployment.
• DeepSeek Talent — AGI-focused lab; pretraining, RL-for-reasoning, and infrastructure roles (official talent portal).
• Merge Labs Careers — Frontier BCI lab (bio × AI); hiring ML scientists for Bayesian optimization, RL, and closed-loop molecular/neuro engineering. See the FM-os Merge dossier.

_{↑ back to top}

Specialized AI Job Boards

• ai-jobs.net — Large dedicated AI/ML/data-science job board with research and engineering filters.

_{↑ back to top}

Aggregators

• 80,000 Hours Job Board — Curated board emphasizing frontier-lab and AI-safety roles across pretraining, RL, and alignment.

_{↑ back to top}

Newsletters & Signals

• Import AI — Weekly research-analysis newsletter (Jack Clark); tracks frontier labs and hiring signals.
• Ahead of AI — Sebastian Raschka's newsletter; deep technical coverage of LLM training and fine-tuning methods.
• The Batch (DeepLearning.AI) — Weekly AI news/insights from Andrew Ng's team; useful for tracking labs and the talent market.

_{↑ back to top}

---

🗺️ Learning Roadmap

Beginner → practitioner (SLM track):

1. Watch a from-scratch course and train a tiny model (see Courses → Foundations).
2. Fine-tune a small open model with LoRA/QLoRA on your own data (Repos → Fine-tuning).
3. Align it with DPO, then try a GRPO-style RL loop (Repos → Post-training & RL).
4. Evaluate honestly (Repos → Evaluation) and serve it on-device (Repos → Serving).
5. Read the SLM surveys + the model tech reports to understand the design space (Papers).

---

🤝 Contribute

This list is data-driven — every entry is a few lines of YAML in data/. Adding a resource is a two-line PR; you never touch the README (it's generated).

# 1. add your entry to the right file, e.g. data/repos.yml
# 2. regenerate + check locally
make check
# 3. open a PR

See CONTRIBUTING.md for the entry schema and the one rule (every entry needs a working url). A weekly Action re-verifies links and refreshes repo stats automatically.

---

⭐ Star History

🙌 Contributors

FM-os is maintained by @wjlgatech and the community. Sibling projects: longevity-loop · rsi · FDE-os.

_{README generated from data/*.yml by scripts/build_readme.py — do not edit by hand.}