https://arxiv.org/abs/2412.01276
https://arxiv.org/abs/2410.13166
https://fxtwitter.com/SakanaAILabs/status/1866286131685498920?t=BnOxECG53IjwbSdBZwrTcg&s=19
Gradient descent x evolution
https://arxiv.org/abs/1810.06773
https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1012673
https://en.wikipedia.org/wiki/Neural_Turing_machine
https://en.wikipedia.org/wiki/Differentiable_neural_computer
deepseek
https://x.com/reach_vb/status/1872000205954089011
o1
https://www.youtube.com/watch?v=6PEJ96k1kiw
https://youtu.be/rJkTsNrnu8g?si=rXA27hRIV5gkVHiD
https://www.youtube.com/watch?v=AfAmwIP2ntY
https://fxtwitter.com/gm8xx8/status/1871814856904478798?t=9c0HsVIk_VQbAQQyTI8XZA&s=19
https://arxiv.org/abs/2412.18319
https://fxtwitter.com/gm8xx8/status/1871644052166418867?t=YNarMXqmUbhm7sEm_taDkQ&s=19
https://arxiv.org/abs/2412.17256
Kenneth Stanley
https://en.wikipedia.org/wiki/Kenneth_Stanley
https://arxiv.org/abs/2501.04682
https://arxiv.org/abs/2412.06769
https://x.com/NovaSkyAI/status/1877793041957933347
https://novasky-ai.github.io/posts/sky-t1/
https://arxiv.org/abs/2501.04519
https://x.com/ziv_ravid/status/1877736408191754487
https://arxiv.org/abs/2501.05707
https://llm-multiagent-ft.github.io/
https://x.com/du_yilun/status/1878851914307371440
Transformer^2
https://arxiv.org/abs/2501.06252
Cognitive agent architecture for multiagent AI civilizations: "PIANO" (Parallel Information Aggregation via Neural Orchestration) that "enables agents to interact with humans and other agents in real-time while maintaining coherence across multiple output streams" in Minecraft
https://arxiv.org/abs/2411.00114
https://x.com/GuangyuRobert/status/1852397383939960926
inference time scaling of diffusion models https://fxtwitter.com/iScienceLuvr/status/1879094413588107319
https://arxiv.org/abs/2501.06848
"Our experimental results on language modeling, common-sense reasoning, genomics, and time series tasks show that Titans are more effective than Transformers and recent modern linear recurrent models." https://fxtwitter.com/behrouz_ali/status/1878859086227255347
https://arxiv.org/abs/2501.00663v1
Wake up babe new transformer killer dropped
https://arxiv.org/abs/2501.07301
It enables visual thinking in MLLMs by generating image visualizations of their reasoning traces.
https://arxiv.org/abs/2501.07542
https://arxiv.org/abs/2206.07682
https://arxiv.org/abs/2301.08028
>supervised fine-tuning (SFT) helps the model memorize and align with certain outputs, while reinforcement learning (RL) helps the model generalize and learn out-of-distribution (OOD) tasks
SFT to tame initial instruction following, and then RL to generalize
https://arxiv.org/abs/2501.17161
https://x.com/Hesamation/status/1884579088121073972
https://arxiv.org/abs/2411.13420
https://x.com/edwardfhughes/status/1887492625453793471?t=bXqO9PkNQOO_rgcbhCjGFw&s=19
Self play self driving
I predict that reasoning in latent space will be the next set of breakthroughs
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
https://arxiv.org/abs/2502.05171
https://arxiv.org/abs/2502.05078
https://x.com/burny_tech/status/1890189798494941691
https://www.verses.ai/research-blog/variational-bayes-gaussian-splatting-a-bayesian-approach-for-continual-3d-learning
https://arxiv.org/html/2410.03592v1
https://fxtwitter.com/DimitrisPapail/status/1889755872642970039?t=Vr-9NWmA1IG51D_LJzhA2Q&s=19
https://arxiv.org/abs/2502.01612
https://fxtwitter.com/omarsar0/status/1889681118913577345 https://arxiv.org/abs/2502.06049
memory-augmented Transformer architecture that incorporates a dynamic memory module
large memory models
reasoning papers
https://fxtwitter.com/TheAITimeline/status/1888720075793793121?t=6pLwpwpjNBhPr2zJP96dLA&s=19
New Levin paper https://arxiv.org/abs/2411.13420
wordcels vs latentspacerotators
https://x.com/burny_tech/status/1890194390066594120
https://arxiv.org/abs/2502.05171
https://x.com/takeru_miyato/status/1893538672533733601 https://takerum.github.io/akorn_project_page/
Artificial Kuramoto Oscillatory Neurons
uses Kuramoto model which is also used in computational neuroscience https://en.wikipedia.org/wiki/Neural_oscillation#Kuramoto_model
https://arxiv.org/abs/2502.17543
https://arxiv.org/abs/2502.21321
https://arxiv.org/abs/2502.09992
https://iliao2345.github.io/blog_posts/arc_agi_without_pretraining/arc_agi_without_pretraining.html
- Transformers (Attention)
- RNNs (LSTMs, GRUs)
- [CNNs](https://arxiv.org/abs/1810.13118)
- GNNs / KGs
- Manifold-based learning (UMAP, t-SNE, autoencoders (VAEs))
- Spectral-based models (Fourier Neural Operators (FNOs))
- Mamba (SSMs)
- Memory-Augmented NNs (NTMs, DNCs)
- Energy-based & Variational models (EBMs)
- KANs
- Flow-based models (CNFs)
- GANs
- Diffusion models and Stable Diffusion
- Liquid NNs / SNNs
"
"However, real scientific breakthroughs will come not from answering known questions, but from asking challenging new questions and questioning common conceptions and previous ideas.
We're currently building very obedient students, not revolutionaries. This is perfect for today’s main goal in the field of creating great assistants and overly compliant helpers. But until we find a way to incentivize them to question their knowledge and propose ideas that potentially go against past training data, they won't give us scientific revolutions yet.
If we want scientific breakthroughs, we should probably explore how we’re currently measuring the performance of AI models and move to a measure of knowledge and reasoning able to test if scientific AI models can for instance:
Challenge their own training data knowledge
Take bold counterfactual approaches
Make general proposals based on tiny hints
Ask non-obvious questions that lead to new research paths
We don't need an A+ student who can answer every question with general knowledge. We need a B student who sees and questions what everyone else missed.
https://x.com/Thom_Wolf/status/1897630495527104932
"
https://arxiv.org/abs/2410.01131
https://arxiv.org/abs/2211.01233
enhanced knowledge graphs and reasoning LLMs
https://arxiv.org/abs/2502.13025
https://youtu.be/W2uauk2bFjs?si=MVhlTpK2kbaxmt-a
neural celluar automata
https://distill.pub/2020/growing-ca/
https://google-research.github.io/self-organising-systems/difflogic-ca/
https://corticallabs.com/cl1.html
it's the people connected to Friston who are behind the paper where they taught biological neurons to ping ping where they mention the free energy principle https://corticallabs.medium.com/in-vitro-neurons-learn-and-exhibit-sentience-when-embodied-in-a-simulated-gameworld-387ec3f2c870
https://arxiv.org/abs/2502.06034
"Traveling Waves Integrate Spatial Information Through Time
In the physical world, almost all information is transmitted through traveling waves -- why should it be any different in your neural network?
Just as ripples in water carry information across a pond, traveling waves of activity in the brain have long been hypothesized to carry information from one region of cortex to another; but how can a neural network actually leverage this information?
This paper introduces convolutional recurrent neural networks that learn to produce traveling waves in their hidden states in response to visual stimuli, enabling spatial integration.
They made wave dynamics flexible by adding learned damping and natural frequency encoders, allowing hidden state dynamics to adapt based on the input stimulus.
By then treating these wave-like activation sequences as visual representations themselves, they obtain a powerful representational space that outperforms local feed-forward networks on tasks requiring global spatial context.
In particular, they observe that traveling waves effectively expand the receptive field of locally connected neurons, supporting long-range encoding and communication of information.
They demonstrate that models equipped with this mechanism solve visual semantic segmentation tasks demanding global integration, significantly outperforming local feed-forward models and rivaling non-local U-Net models with fewer parameters.
Video shows Tetris-like dataset and variants of MNIST to compare the semantic segmentation ability of these wave-based models with two relevant baselines: Deep CNNs w/ large receptive fields, and small U-Nets."
https://arxiv.org/abs/2408.08435
automating factorio https://fxtwitter.com/bio_bootloader/status/1899147299546423507
https://jackhopkins.github.io/factorio-learning-environment/
https://x.com/renxyzinc/status/1899539629411270758
https://opensource.getren.xyz/ittm/8_usm.html
Universal state machine
>dynamically growing symbolic graphs
Reminds me of https://arxiv.org/abs/2502.13025
https://youtu.be/W2uauk2bFjs?si=MVhlTpK2kbaxmt-a
Abandoning Objectives: Evolution through the Search for Novelty Alone https://www.cs.swarthmore.edu/~meeden/DevelopmentalRobotics/lehman_ecj11.pdf
Why Greatness Cannot Be Planned https://link.springer.com/book/10.1007/978-3-319-15524-1
#72 Prof. KEN STANLEY 2.0 - On Art and Subjectivity [UNPLUGGED] https://www.youtube.com/watch?v=DxBZORM9F-8
https://x.com/burny_tech/status/1894491541227671779
"
Kenneth Stanley is my spirit animal
rage against the predefined objectives
embrace the fully divergent search full of novelty and accidental epiphany with serendipity
https://www.youtube.com/watch?v=DxBZORM9F-8
https://www.youtube.com/watch?v=_2vx4Mfmw-w
it's an evolutionary breeding process of images, but humans pick the images that should have offsprings
idea: picbreeder but let multimodal LLMs instead of humans choose the next image in the evolutionary breeding process 🤔
https://www.youtube.com/watch?v=_2vx4Mfmw-w
i need to look more into how those novelty/diversity algorithms that he's mentioning work, maybe they can be added into RL reward functions in LLM RL
https://www.researchgate.net/publication/46424802_Abandoning_Objectives_Evolution_Through_the_Search_for_Novelty_Alone
New research project: Lluminate - an evolutionary algorithm that helps LLMs break free from generating predictable, similar outputs. Combining evolutionary principles with creative thinking strategies can illuminate the space of possibilities.
https://x.com/_joelsimon/status/1899884376172982392?t=Z4q0CZ2C5-9v8A-QJPnpNA&s=19
https://www.joelsimon.net/lluminate
Metagradient Descent
https://arxiv.org/abs/2503.13751
https://x.com/f14bertolotti/status/1902259983753842971
https://intoai.pub/p/chain-of-draft-cod-is-the-new-king
https://arxiv.org/abs/2502.18600
https://x.com/METR_Evals/status/1902384481111322929?t=4SbjaExBNAyL4W-3w881lQ&s=19
https://arxiv.org/abs/2503.14499
https://arxiv.org/abs/2202.09467
https://www.youtube.com/watch?v=48GRiu-TMmg
https://fxtwitter.com/ericzhao28/status/1901704339229732874?t=fJDIBbvCb2b_QIBxnsHiLw&s=19
https://arxiv.org/abs/2502.01839
https://x.com/eric_haibin_lin/status/1901662955307200974
https://x.com/Synced_Global/status/1901794723633025282?t=pIRCz8vUZxj92QQ9aMrBzg&s=19
https://dapo-sia.github.io/static/pdf/dapo_paper.pdf
DAPO algorithm grpo uphrade
https://www.youtube.com/watch?v=4KK2NjMrcjo
https://arxiv.org/abs/2503.24322
https://www.youtube.com/watch?v=UMkCmOTX5Ow
An Evolved Universal Transformer Memory
[Paper] https://arxiv.org/abs/2410.13166
Memory Layers at Scale
[Paper] https://arxiv.org/abs/2412.09764
Titans: Learning to Memorize at Test Time
[Paper] https://arxiv.org/abs/2501.00663
https://www.youtube.com/watch?v=qhYQ20TbtJ8
CoCoNuT: Training Large Language Models to Reason in a Continuous Latent Space
[Paper] https://arxiv.org/abs/2412.06769
Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach
[Paper] https://arxiv.org/abs/2502.05171
biologically inspired algorithms, alternatives to backprop
https://towardsdatascience.com/feedback-alignment-methods-7e6c41446e36/
https://arxiv.org/abs/2212.13345
https://medium.com/@reutdayan1/hebbian-learning-biologically-plausible-alternative-to-backpropagation-6ee0a24deb00
reinforcement learning approaches that allow language models to improve without additional data using self-play
https://arxiv.org/abs/2509.07414
https://x.com/teortaxesTex/status/1965654111069876296