https://github.com/naklecha/llama3-from-scratch
https://github.com/trotsky1997/MathBlackBox?tab=readme-ov-file#news https://x.com/casper_hansen_/status/1852342730527023473?t=hwqUzQ5Gv00pqh4eQdB9lA&s=19
implementation details of o1
https://arxiv.org/abs/2501.09686
O1 replication https://arxiv.org/abs/2410.18982
https://github.com/hijkzzz/Awesome-LLM-Strawberry Democratization of o1 is happening
https://github.com/srush/awesome-o1
https://x.com/DrJimFan/status/1834279865933332752
Large Language Monkeys: Scaling Inference Compute with Repeated Sampling. Brown et al. https://arxiv.org/abs/2407.21787v1
Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters. https://arxiv.org/abs/2408.03314 https://x.com/rohanpaul_ai/status/1835443326205517910
https://x.com/terryyuezhuo/status/1834286548571095299
ReFT: Reasoning with Reinforced Fine-Tuning
https://arxiv.org/abs/2401.08967
Training Large Language Models for Reasoning through Reverse Curriculum Reinforcement Learning https://arxiv.org/abs/2402.05808
https://x.com/iamgingertrash/status/1834297595486675052
tree search distillation + RL post training! https://x.com/rm_rafailov/status/1834291016192360743
Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking https://arxiv.org/abs/2403.09629
https://x.com/laion_ai/status/1834564564601729421
Let's Verify Step by Step https://arxiv.org/abs/2305.20050 https://www.youtube.com/watch?v=hZTZYffRsKI
https://www.reddit.com/r/LocalLLaMA/comments/1fgr244/reverse_engineering_o1_architecture_with_a_little/
https://www.interconnects.ai/p/reverse-engineering-openai-o1
https://github.com/srush/awesome-o1/
https://huggingface.co/collections/philschmid/llm-reasoning-papers-66e6abbdf5579b829f214de8
https://x.com/srush_nlp/status/1846599704194302130?t=_vLPjFMc0JCaHYXz37IATQ&s=19
https://arxiv.org/abs/2411.16489
Elvis AI research news
https://youtu.be/vSm9wNw1Yek?si=ONbLSlGnpHIYE4mv
https://trendingpapers.com
https://mail.bycloud.ai/
llama from scratch
https://github.com/rasbt/LLMs-from-scratch?tab=readme-ov-file#bonus-material https://www.youtube.com/results?search_query=implementing+llama+from+scratch https://www.youtube.com/watch?v=lrWY4O5kUTY https://www.youtube.com/watch?v=oM4VmoabDAI
https://github.com/NirDiamant/RAG_Techniques
https://www.interconnects.ai/p/frontier-model-post-training
https://x.com/srchvrs/status/1830333083066814507?t=onr6xERkw5pUDj-qbcWUwQ&s=19
Agents survey
https://arxiv.org/abs/2404.11584v1
https://aiagentsdirectory.com/landscape
https://trigaten.github.io/Prompt_Survey_Site/
Computable Artificial General Intelligence https://arxiv.org/abs/2205.10513
Agents on software engineering
https://arxiv.org/abs/2409.09030
"90% of frontier ai research is already on arxiv, x, or company blog posts.
q* is just STaR
search is just GoT/MCTS
continuous learning is clever graph retrieval
+1 oom efficiency gains in deepseek-coder paper"
https://x.com/aidan_mclau/status/1811898849189069068?t=Sp5B-qFF_7xohpTwNvbnHg&s=19
Foundational models in music
https://arxiv.org/abs/2408.14340
Multimodal diffusion plus next token prediction https://www.arxiv.org/abs/2408.11039
Llama 3.1 details
https://x.com/danielhanchen/status/1815798460052038095
Distilling transformers to mamba
https://arxiv.org/abs/2408.10189
Agent 2.0
https://arxiv.org/abs/2406.18532
Constraining llms using automata
https://arxiv.org/abs/2407.08103v1
https://deepmind.google/discover/blog/ferminet-quantum-physics-and-chemistry-from-first-principles/
Lagrangian neural networks
https://youtu.be/HLUIx6FqAvg?si=irQp7qKm-7kW5Wf4
Qietstar
Star selfthaught reasoner
https://youtu.be/T9gAg_IXB5w?si=VzdBeoei0qLQsQ5Z
LSTM plus mamba
https://arxiv.org/abs/2403.16536v2
Foundation models course
https://x.com/rohanpaul_ai/status/1828577063378547017
https://cs.uwaterloo.ca/~wenhuche/teaching/cs886/
"Funny how secret sauce methods become openly published just as the frontier moves forward.
- Data composition solved
- LR schedules 80% solved
- efficient RLHF ≈solved
New frontier: actually strong comprehensive *pretraining* synthetic data recipes, distillation, LLM RL…"
https://x.com/teortaxesTex/status/1812282965935673545?t=iazC_r-JAdiGSbnXC8-vXQ&s=19
https://x.com/TheAITimeline/status/1812216803323461736 new papers
https://arxiv.org/abs/2407.12077
Prompt engineering survey
https://x.com/bindureddy/status/1814409737557160044?t=6PqfEy63nWOQJ5lIEKBZEg&s=19
"self-modeling to artificial networks causes a significant reduction in network complexity" https://x.com/juddrosenblatt/status/1814463465572184133
1. Interfacing an LLM with a reliable symbolic system (Prolog) raises math performance near ceiling: Tested on an *entirely new* collection of math word problems, the Non-Linear (NLR) reasoning dataset, to ensure all were outside the LLM training set. GPT fails completely. But GPT writing prolog code succeeds near ceiling. https://arxiv.org/abs/2407.11373
2. How can informal reasoning improve formal theorem proving? Lean-STaR: A framework for learning to interleave informal thoughts with steps of formal proving. Training language models to produce informal thoughts prior to each step of a proof, thereby improving the model’s formal theorem-proving capabilities. https://arxiv.org/abs/2407.10040
3. Adding self-modeling to artificial networks causes a significant reduction in network complexity. When artificial networks learn to predict their internal states as an auxiliary task, they change in a fundamental way. https://arxiv.org/abs/2407.10188
4. A system that incorporates both natural language pre-training and reinforcement learning from the start. https://arxiv.org/abs/2308.01399
5. Georgia Tech researchers have developed a neural network, RTNet, that mimics human decision-making processes, including confidence and variability, improving its reliability and accuracy in tasks like digit recognition. https://research.gatech.edu/new-neural-network-makes-decisions-human-would
9. Why Has Predicting Downstream Capabilities of Frontier AI Models with Scale Remained Elusive? https://arxiv.org/abs/2406.04391
https://arxiv.org/abs/2402.14735
https://x.com/EshaanNichani/status/1762122616255709665?t=rdJ34C0am2JCWkSFLREQYA&s=19
AI with universal physical understanding
https://x.com/AnimaAnandkumar/status/1815441489398546741?t=3VhKU4MHfe6v25BrxeBZug&s=19
Model distillation
https://x.com/rohanpaul_ai/status/1815464921032991003?t=yE3LN0PAdL9GzhKBk07Cwg&s=19
https://arxiv.org/abs/2402.13116
Jailbreaking techniques
https://arxiv.org/abs/2403.04786v1
Training works at the edge of stability, the edge of chaos is optimal
https://x.com/Ethan_smith_20/status/1815759691068350875
https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=
AI x math and reasoning
https://www.harmonic.fun/news
https://x.com/reach_vb/status/1811069858584363374
https://arxiv.org/abs/2404.06405
https://deepmind.google/discover/blog/funsearch-making-new-discoveries-in-mathematical-sciences-using-large-language-models/
https://arxiv.org/abs/2006.08381
https://arxiv.org/abs/2407.10040
https://arxiv.org/abs/2203.14465
https://github.com/teacherpeterpan/self-correction-llm-papers
https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/
Making LLMs do arithmetic
https://x.com/jxmnop/status/1816958426385383753
Synthetic data engineering
https://x.com/alexandr_wang/status/1816491442069782925?t=jDuJ7Ka8wLKVvJPXoXvrLA&s=19
OpenAI’s Strawberry hypothesis Active Inference
https://x.com/tedx_ai/status/1819761571952042448
https://x.com/tedx_ai/status/1820425220991393929
AI resources
https://x.com/sebkrier/status/1824849583362723975?t=4eznZWSih89sslTfnSLXOA&s=19
https://iai.tv/articles/the-future-of-ai-is-analogue-auid-2917
https://www.nature.com/articles/d41586-024-02704-y
https://gamengen.github.io/
https://www.nature.com/articles/s41467-024-50966-x
quantum neural networks https://arxiv.org/abs/2408.12739
https://arxiv.org/abs/2403.01946
Chomsky Impossible language models https://www.youtube.com/watch?v=8lU6dGqR26s&feature=youtu.be
https://www.livescience.com/technology/artificial-intelligence/novel-chinese-computing-architecture-inspired-by-human-brain-can-lead-to-agi-scientists-say
https://www.nature.com/articles/s43588-024-00674-9
https://www.quantamagazine.org/how-the-higgs-field-actually-gives-mass-to-elementary-particles-20240903/
Open source implementation of AlphaFold3
https://github.com/Ligo-Biosciences/AlphaFold3
https://honeycomb.sh/blog/swe-bench-technical-report
https://news.cs.washington.edu/2024/08/19/mind-over-model-allen-schools-rajesh-rao-proposes-brain-inspired-ai-architecture-to-make-complex-problems-simpler-to-solve/
https://www.nature.com/articles/s41593-024-01673-9
https://virtual-protocol.github.io/mario-videogamegen/
https://arxiv.org/abs/2408.14837
https://arxiv.org/abs/2405.05254
https://github.com/verazuo/jailbreak_llms
Transformers solve an open problem in symbolic mathematics https://x.com/f_charton/status/1834725632728613094 https://www.youtube.com/watch?v=yCzV97QNG8w
Denny Zhou (Founded & lead reasoning team at Google DeepMind) - "We have mathematically proven that transformers can solve any problem, provided they are allowed to generate as many intermediate reasoning tokens as needed. Remarkably, constant depth is sufficient." https://fxtwitter.com/denny_zhou/status/1835761801453306089
diagram of thought for math https://arxiv.org/abs/2409.10038
Parables on the Power of Planning in AI: From Poker to Diplomacy: Noam Brown, researching reasoning in OpenAI https://www.youtube.com/watch?v=eaAonE58sLU https://x.com/burny_tech/status/1836991461692231707
deepmind not gradient based but evolutionary tennis robot https://x.com/davidad/status/1821622048596144490
https://x.com/GoogleDeepMind/status/1821562365931855970
https://sites.google.com/view/competitive-robot-table-tennis/home?utm_source&utm_medium&utm_campaign&utm_content&pli=1
https://www.nature.com/articles/s41467-024-51477-5
https://x.com/GoogleAI/status/1839730191347699820?t=C0zweN806imi8-S_gXD_4g&s=19
flow matching
https://arxiv.org/abs/2210.02747
https://www.youtube.com/watch?v=7NNxK3CqaDk
Liquid foundation models, transformer killers?
https://x.com/LiquidAI_/status/1840768728402993352?t=GhYuA85zs6FmKec0qNxgig&s=19
https://www.liquid.ai/blog/liquid-neural-networks-research
Automated Design of Agentic Systems
https://arxiv.org/abs/2408.08435
https://x.com/rohanpaul_ai/status/1841502578028503298
thermodynamic AI
https://twitter.com/MaxAifer/status/1841671025991229874
https://twitter.com/MaxAifer/status/1841902089015824393
https://arxiv.org/abs/2410.01793
https://www.youtube.com/watch?v=6DrCq8Ry2cw
Human-Timescale Adaptation in an Open-Ended Task Space
https://arxiv.org/abs/2301.07608
POET: Endlessly Generating Increasingly Complex and Diverse Learning Environments and their Solutions through the Paired Open-Ended Trailblazer https://arxiv.org/abs/1901.01753 https://www.uber.com/en-CZ/blog/enhanced-poet-machine-learning/
RAG over a lot of files in practice
https://x.com/shanselman/status/1842667956100296722?t=AeWTwMZPkya78xAOTllngA&s=19
automatic design of agentic systems, multiagent systems creating multiagent systems https://arxiv.org/abs/2408.08435
https://arxiv.org/abs/2410.08146
LeanAgent: Lifelong Learning for Formal Theorem Proving https://arxiv.org/abs/2410.06209
Recursively selfimproving swarms maybe soon mmmmm...
https://github.com/openai/swarm
AI agents https://github.com/hyp1231/awesome-llm-powered-agent
Autorotating neural networks https://openreview.net/forum?id=Gg3JlR9btC
cool playlist of machine learning resources https://www.youtube.com/playlist?list=PLg642XuzWS1Ts3ED6ym_HJ9hqq8QOdB9Z
Understanding LLMs: A Comprehensive Overview from Training to Inference https://arxiv.org/abs/2401.02038
automated agent workflow generation https://arxiv.org/abs/2410.10762
illya recommended AI papers https://x.com/Sumanth_077/status/1853078836239613970
Automated prompt engineering https://x.com/cwolferesearch/status/1853465302031556725?t=nMBZJzaz_xMWFklZCxcnAQ&s=19
Differential Transformer https://www.youtube.com/watch?v=6Xy0tCidPl0
Harvard Presents NEW Knowledge-Graph AGENT (MedAI)
https://www.youtube.com/watch?v=Fm68I-phaiY
Alternatives to LangChain https://www.reddit.com/r/LangChain/s/IdKXYMpiua
better optimizer than adam ADOPT https://arxiv.org/abs/2411.02853 https://x.com/ishohei220/status/1854051859385978979
jailbreaks https://rentry.org/jb-listing
local models https://rentry.org/meta_golocal_list
Best neurosymbolic AI architecture recommended by Chollet
https://x.com/fchollet/status/1855362153563463762?t=hAojgyzrh4468vLJH-MnQA&s=19
https://arxiv.org/abs/2410.23156
MIT researchers develop an efficient way to train more reliable AI agents
https://news.mit.edu/2024/mit-researchers-develop-efficiency-training-more-reliable-ai-agents-1122
Model-Based Transfer Learning for Contextual Reinforcement Learning
https://arxiv.org/abs/2408.04498
"Experimental results suggest that MBTL can achieve up to 50x improved sample efficiency compared with canonical independent training and multi-task training. This work lays the foundations for investigating explicit modeling of generalization, thereby enabling principled yet effective methods for contextual RL.
We introduce Model-Based Transfer Learning (MBTL), which layers on top of existing RL methods to effectively solve contextual RL problems. MBTL models the generalization performance in two parts: 1) the performance set point, modeled using Gaussian processes, and 2) performance loss (generalization gap), modeled as a linear function of contextual similarity. MBTL combines these two pieces of information within a Bayesian optimization (BO) framework to strategically select training tasks."
https://github.com/NirDiamant/Prompt_Engineering?tab=readme-ov-file#-advanced-strategies
https://youtu.be/68_BKkymYs4?si=HGGP6UIJHCcvHP1e llama 3.1 405b
Brain learning algorithm https://www.nature.com/articles/s41593-023-01514-1
Database of How Companies Actually Deploy LLMs in Production (300+ Technical Case Studies, Including Self-Hosted
https://x.com/rohanpaul_ai/status/1863606556228788354?t=6ei5ThiAa7oGY_lVZx0SGw&s=19
https://www.amazon.com/AI-Engineering-Building-Applications-Foundation/dp/1098166302
AlphaZero from scratch https://www.youtube.com/watch?v=wuSQpLinRB4
https://en.wikipedia.org/wiki/MuZero
AlphaFold from scratch
https://www.youtube.com/playlist?list=PLJ0WcPQS7xJVJr6ceIPFSkAGAgrkmw1c9
Deepseek V3 technical report
https://api-docs.deepseek.com/news/news1226
https://x.com/nrehiew_/status/1872318161883959485?t=SB9rLEF4MoCy4Gt9FBVicQ&s=19
https://en.wikipedia.org/wiki/Physics-informed_neural_networks
sheaf neural networks
https://x.com/meowdib/status/1922315466401308965
flow matching visualization https://fxtwitter.com/alec_helbling/status/1924451851316932758
A Principled Bayesian Framework for Training Binary and Spiking Neural Networks https://x.com/adeelrazi/status/1926849339596439804 https://arxiv.org/abs/2505.17962 (edited)
https://arxiv.org/abs/2505.12540
It seems to be support for platonic representation hypothesis https://arxiv.org/abs/2405.07987
"Our conjecture is as follows: neural networks trained with the same objective and modality, but with different data and model architectures, converge to a universal latent space such that a translation between their respective representations can be learned without any pairwise correspondence."
and it surprisingly works
They had to design pretty sophisticated loss functions in that paper to get those translation mappings, I think it wasn't obvious that it will even work
Also I don't think that it was obvious that the semantic shape of the different embeddings would be this similar such that it's this translatable
AI x physics intersection is an endless rabbithole. You can try to:
- make AI systems model physics better than other AI systems (ferminet, lagrangian neural networks https://arxiv.org/abs/2003.04630 latent space squeezing https://www.youtube.com/watch?v=XRL56YCfKtA ,...),
- study existing AI systems using methods from physics (the principles of deep learning theory book https://arxiv.org/abs/2106.10165 statistical physics https://arxiv.org/abs/2506.04374 ,...),
- design better AI architectures using insights and methods from physics (liquid neural networks https://arxiv.org/abs/2006.04439 ,...),
- advance many AI architectures that are applied physics (diffusion models https://en.wikipedia.org/wiki/Diffusion_model#Non-equilibrium_thermodynamics ,...),
- make neurosymbolic systems to do physics algebraically (DreamCoder https://arxiv.org/abs/2006.08381 , formal theorem proving https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/ https://arxiv.org/abs/2405.14333 ,...),
- do machine designed physics experiments https://www.youtube.com/watch?v=T_2ZoMNzqHQ
etc.
https://github.com/bgavran/Category_Theory_Machine_Learning
Best AI papers of 2025
https://x.com/i/status/2007379211414319194
https://www.alphaxiv.org/shared/folder/019b804e-175c-7d6f-b198-c9e36afeb41b
Foundations of Schrödinger Bridges for Generative Modeling
https://x.com/i/status/2034969756902527321
https://arxiv.org/abs/2603.18992