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](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
zatím mě v GRPO napadá akorát nechat vyřešit nějakej matematickej problém víc způsobama, a víc upweightovat ty víc rare correct solutions, místo rewarding všech correct solutions uniformě
v accuracy rewards
to by mohlo incentivizovat směrem k víc rare ale zároveň correct postupům při řešení problémů
pokusím dát tu moji ideu do Huggingfacový open source replikace DeepSeek R1 (nebo jiný replikace podle toho jak je dobrá a simple) 😄
ale nejdřív to dám do rovnic
Options:
Most likely tokens
Scan for correctness in less likely tokens too
Upweight less likely tokens only in correct solutions
Keep track of correct solutions and do cosine similarity and upweight according to that
Measure similarity přes:
Cosine similarity nad openai embeddings, llm classification, nearest neighbours
Distance k ground truth solutions v datasetu
https://x.com/natolambert/status/1901758394446528586?t=KKJ_9AVn-OxFQLNR_JtlKA&s=19
Automatically check which diversity hyperparameter decreases loss best in as long term as possible
extension that hides nonSTEM tweets/blueskies etc. (Create tapermonkey script that scans tweets and hides those that are not STEM related)
Fluid dynamics
Ball simulation
Physics simulator
minecraft clone
Omnimodal waifu
instead of rogue like game, make interactive the sims like llm game instead
Flying in universe being with planets attracting me with gravity
Chaos coding specialized for game dev
Racing game with hyper movement
Create mathematics, science, philosophy knowledge graph with as many nodes as possible
Neural network model trained on snake
Train physics based AI on physics dataset
3blue1brown videos for physics and AI
incentivize generalization in the loss function
roguelike:
fix error with broken character caching
add direct claude API support
more logical level generation
enemies moving,, strategically blocking entrance to next level
enemies dropping AI generated loot
AI generated visual images of enemies, maybe later videos, objects in game engines etc.
"
"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
"
I think this could as a naive starter be put into current LLMs by putting it into RL scoring function using LLMs as judges
Quantum gravity beyond anti De Sitter space
[https://youtu.be/2p_Hlm6aCok?si=UpxtcYWtmMcHwjeG](https://youtu.be/2p_Hlm6aCok?si=UpxtcYWtmMcHwjeG)
GTA clone
Planes
alien oscillating physics to my bhop game
trackmania
neural network for snake
ant simulator
General togglable movements game
There are countless different definitions of intelligence motivated by different goals that yield different general equations and mathematical frameworks of intelligence compatible with different types of systems that yield different concrete equations of intelligence that can be concretely by different methods empirically localized in a system or implemented in code, and all of them were created by human intelligences, so wait for what kinds of models will all sorts of alien artificial intelligences running all sorts of algorithms on all sorts of substrates come up with that will be incomprehensible for human intelligences. All kinds of intelligences live in a high dimensional space where each dimension correspond to some degree of capability measured by some methodology, and some of these dimensions are interconnected with each other.
Write a paper on this taxonomy, survey, and meta-analysis.
Ask if I can help pony diffusion or nous research or primeintellect
Create multiagent LLM system corresponing to a whole company
openrouter collection of llms under collections of personas under collection of voices under collection of agentic frameworks under collection of tools
writings, kniha, hra, komix, animak:
přemýšlím že vezmu moje oblíbený matiky a každou část zasadím do magickýho příběhu zkoumatele potkávající magický starý kouzelníky plný moudrosti a různý mimozemský stvoření mluvící mimozemštinou co se musí luštit
byl jednou jeden zivot ale fundamentalni cstice vesmiru, ponified
combine all made games into one
start with existing working games by someone and add my features
centralize results from all benchmarks of all models to one website and do automatic updating?
Exponential made of sigmoids model of technological progress
I'm really recently thinking about creating some site that centralizes information about as many benchmarks of as many models as possible, including as many concrete practical results and anecdotal evidence, including as many proposed comments about various strengths and limitations of the benchmarks and models, and not just LLMs, to be able to see the big picture more clearly
And on this website people could maybe vote how legitimate the benchmarks are or how they feel about the models overall
But adding some sort of more objective evaluations could be nice too
Hmm, as a starter I could add a list of benchmarks sorted by how I like them the most to my burnyverse.com site 😄
But it would be ideal if as much information as possible there would be updated realtime
Sort by voting, popularity on Twitter, validity, burnys opinion, popularity among total amount of models benchmarked
For example be able to see Claude benchmarks, and which models are better overall in code, or in some specific benchmarks, on one "Claude" page with tiger models on the sides - travelling latent space of benchmarks and models in various levels of abstraction with directions
YouTube video upload
Update obsidian (notes, AI)
Já chci nějaký anime co mnohem víc ukáže potenciální extrémní upgrade kognitivních funkcí na vědu/matiku apod.
Sem tam o tom zatím akorát dreamuju v mý hlavě nebo textu, mohl bych se to někdy pokusit víc rozšířit.
Ale teď chci dělat na možným poupraveným AI učím algoritmu pro vědu v praxi v realitě co mě nedávno napadl, než ve sci-fi, to je k tomu blíž. 😄
I write the most useful information dense text when I'm trying to prove a random stranger on the internet that he's wrong. I should hack my brain by creating social media with synthetic bots that are consistently wrong about the stuff that I want to write about.
Alphazero na prší
make minecraft mobs more intelligent, make them speak
nebo by mohlo být cool realtime generovat moby a bloky atd. :smile:
create the biggest possible skill tree for mathematics possible
Create a simplest possible app where you have skill tree in this format, where you can label each node as completed, where you can open and close various subtrees
Arithmetic & Pre-Algebra
- Basic Arithmetic
- Counting
Algebra
- Elementary Algebra
- Linear Equations
Reasoning using chains of thought in language, chains of continuous thought/latent space, graphs of thoughts, chains of images, maybe soon chains of audio/videos... I wonder how soon is some architecture that combines it all, since humans think abstractly, in language, visually, in audio, in video. With fully multimodal base.
[https://www.youtube.com/watch?v=qhYQ20TbtJ8](https://www.youtube.com/watch?v=qhYQ20TbtJ8)
Duolingo for stem in depth
How about increasing image gen prompt adherence using reinforcement learning by generating an image, prompt visual LLM for evaluating prompt adherence and converting that into reward signal
And you could use the LLM to try to extract a lot of more signals to create all sorts of incentives
And in language you could try to use LLM as a judge in non-easily-verifiable domains and get reward signal from that
It will be expensive tho
LLM accuracy will be probably often broken but it should give better signals than just noise
Or instead of actual prompt adherence, you'll index on adherence to the vLLM's world model, which may be too distant from ground truth
And noise might get amplified too much
RL algorithm inferring on the fly which combination of modalities is more effective for various tasks in it's hypergraph of multimodal thoughts?
https://fxtwitter.com/TheTuringPost/status/1906304408415359067?t=QrP_I5vSzaLt-3r42Hyyig&s=19
ještě mě napadá že by možná AIčkám šlo feedovat pouze irl pictures/videos (a/nebo zkušenosti robotů) a pak nějak incentivitzovat ať nějak zhmotní ty jejich naučený absktratní koncepty z latentního prostoru (a jejich mutace) v abstraktnějším vizuálním jazyce, což si myslím že je velká část toho jak první umění u lidí vzniklo, když pralidi poprvý načmárali tvary zvířat v hlíně, když se inspirovali čistě reálným životem, a ne (v kombinaci) s jinými umělci, kde tím jak byli první, tak to jinak nešlo 😄 což by možná v podobný formě šlo nejak aplikovat i na derivace rovnic světa? 🤔
duolingo but for STEM and not just on surface level
Add to minecraft AI generated and AI powered textures and entities
Pak asi spojím všechny moje oblíbený featury z různých her, a featury co mě napadly co jsem ještě předtím neviděl, do jedný superhry
Glorious sandbox with infinite possibilities that can be chosen to be hardcoded or on the fly generated or everything in between, and also with story and survival and automation and speedrun modes
Randomly generated movement
Replicate Factorio
Replicate Terraria
Replicate metin
But this math result is so weird and added so much uncertainty into my model of current AI systems. Someone should reverse engineer circuits in R1!
Nedávno jsem ještě přemýšlel jak by možná šly některý deep research problémy částečně zaizolepovat na surface úrovni bez editování LLMka samotnýho. Někdy ve final reportu outputne link co neexistuje, nebo co hodí 404, nebo claim se zdrojem co ten claim nepodpiruje. Napadlo mě, že by ještě po vygenerování reportu mohl projít ty claimy a zdroje, a udělat extra double checking, a když tak to opravit. Ale je taky možný, že to opraví hůř, ale myslím, že overall kvalitu by to spíš mohlo zlepšit. :D
https://fxtwitter.com/skdh/status/1905132853672784121?t=dUgCJpriZbsWdtTGSakmdA&s=19
Thinking about quick ductape fix for hallucinations for now where you backtrack autoregression and try other tokens if outputed link doesn't exist, outputs 404 page, is irrelevant and source of claim isn't there according to classifier, or sources are missing from generated text, until it's valid, but maybe that could also cause infinite loops. And give certainity score to all claims and sources depending on their token probability, or add it via LLM. And do additional automated double checking process after report is generated.
mechinterp circuits simplicity as RL reward function
PINNs applied to some new system
PINN navier stokes bias add to video gen models?
Given the multiple of extremely capable AI tools that's out there, each with its different strengths (i.e., o1-pro, Gemini 2.5, Claude 3.7, Grok 3, etc), we kind of need a meta-AI tool that knows how to manage and combine the outputs of all these! We need UI for the AI power-user!
Multiagent RL
Try bio data
Selforganizing AI plus RL
quantum ML, hybrids, reasoning on top of QM systems
evolutionary ai
langrangian neural networks
collect all definitions of intelligence
Active inference
combine physics informed neural networks (cooling coffee, neural ODEs, hamiltonian/langrangian neural nets) with CoT reasoning (or latent space reasoning)
use LLM evalutation in reward function (interestingness?)
langrangian/hamiltonian neural nets (or combined etc, with other physics inspired neural nets etc.) trained on a lot of physics problems and look for generalization (use insights from how to incentivize generalization)
wiki write outline of AI / AI pages https://x.com/burny_tech/status/1908975009944731965
deep RL / alphazero on games (snake, prší?) (make comparative analysis of different architectures)
one deep RL model on many games and look for generlaization
Langrangian neural networks on standard model?
A discord full of LLMs you can go to and chat shit or chat high brow with a community with AIs that keep you up to date with stuff, direct you where impactful, let you share any type of media, moderate seamlessly, collaborate etc
replicate paper
I suspect we will soon figure out more general superhuman reasoning by some more general multimodal LLM RL, since GRPO is pretty new and currently somewhat work in limited ways in more easily verifiable domains like math and competitive code, but maybe we will need to switch the transformer architecture for something else
https://x.com/jmhessel/status/1899909893324468444
[GitHub - BurnyCoder/burnycoder-repos: List of BurnyCoder repos](https://github.com/BurnyCoder/burnycoder-repos)
upload vibecoded recorded stuff as unlisted
Change to continuous fluid
[GitHub - BurnyCoder/emergent-complexity: Interactive generative art visualization where particles move based on the Clifford Attractor and user-adjustable parameters for particle count, trail opacity, repulsion strength/radius, color influence, density threshold, and anti-cluster force, forming complex patterns on a canvas. Made with HTML with Canvas API, Tailwind CSS, JavaScript.](https://github.com/BurnyCoder/emergent-complexity)
Add more random noise
Mamba + Transformers + RWKW + graph neural networks
master algotihm - hybridize or find mathematical similarities or create customizable lego or shapeshifting agorithm: between Symbolists Connectionists Bayesians Evolutionaries ReinforcementLearners CausalInferencers
Merge the Connectionists, Symbolists, Bayesians, Evolutionaries, ReinforcementLearners, CausalInferencers, DivergentSearchNoveltyMaximizers
https://x.com/burny_tech/status/1909938195543838747
Rubik's cube ai
Since circuits inside LLMs like Claude 3.5 Haiku are different than what LLMs say when asked to explain how they got to an answer, I suspect the same holds for the models in the new reasoning paradigm. Someone needs to do graph attribution mechanistic interpretability on the reasoning models as well! I bet Anthropic will release something like that soon.
(context: [On the Biology of a Large Language Model](https://transformer-circuits.pub/2025/attribution-graphs/biology.html)
AI designed reinforcement learning reward functions for any task on the fly to train fully general model using RL
Jepa
try all possible architectures on snake RL [Before you continue](https://gemini.google.com/app/ea0f411c5871f53d)
Neural Architecture Search
to chce evoluční reinforcement learning from human feedback
neuroevolution s moderníma technikama
picbreeder s moderníma technikama
[https://www.youtube.com/watch?v=_2vx4Mfmw-w](https://www.youtube.com/watch?v=_2vx4Mfmw-w)
myslím že cesta je v primárně evolučním searchu co je hodně pushed přes novelty (hybridně s jinýma metodama)
ale aby se to nerozteklo do chaosu a šumu, tak nějak pořád grounded přes nějaký accuracy signál s nějakou ground truth, možná zákony toho jak fungují koherentní tvary 😄
musí to být pořád nějak grounded v nějakých lidských vzorách
tím jak je moc high entropy vzorů co lidskej mozek interpretuje jako šum 😄
parallel communicating threads with omnimodal graphs of thoughts that synthesize at the end
o3 is now now thinking in images, executed python code, search results etc., when will we think in audio, video, 3D images, game engines etc.? :D
Přednáška na RL irl
Hudba
biologically inspired algorithms, alternatives to backprop
https://towardsdatascience.com/feedback-alignment-methods-7e6c41446e36/
[[2212.13345] The Forward-Forward Algorithm: Some Preliminary Investigations](https://arxiv.org/abs/2212.13345)
[Hebbian Learning: Biologically Plausible Alternative to Backpropagation | by Reut Dayan | Medium](https://medium.com/@reutdayan1/hebbian-learning-biologically-plausible-alternative-to-backpropagation-6ee0a24deb00)
All MLST episodes in notebooklm
https://www.sciencedirect.com/science/article/pii/S0004370221000862
honestly im getting more and more reinforcement learning pilled recently, but i def think it's not the whole picture
i think more ideal learning algorithm will be more of a hybrid (maybe add in some divergent novelty search or something, and evolution) and more biologically inspired (hebbian learning?)
and more ideal architecture would be more biologically inspired (liquid foundational models?) and more neurosymbolic (dreamcoder?)
but maybe biology just isnt optimal and we will find some general algorithm that will be much more effective than biology that differs from biology even more than the current mainstream meta
https://fxtwitter.com/deedydas/status/1913588236959859095/
[https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf](https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf)
Hamiltonian neural networks + attention [HNN-Transformer Integrated Network for Estimating Robot Position | IEEE Conference Publication | IEEE Xplore](https://ieeexplore.ieee.org/document/10316909)
DreamCoder but atomic elements are ZFC axioms with valid transformations [https://www.youtube.com/watch?v=qtu0aSTDE2I](https://www.youtube.com/watch?v=qtu0aSTDE2I)
Dreamcoder for physics
AI architectures based on fluid dynamics
DreamCoder + Reinforcement learning (neuralyl guided program synthesis trained by reinforcement learning)
Combine theoretical mathematical AI frameworks
remix interstellar
combine many genres of music
Interview Tim on intelligence and everything
Music pick most used there note configurations in the background and notes on the same scale to create good sounding melody
Brute force není effective v praxi no
Velká část AI oboru je o tý chytrý redukci toho šíleně gigantickýho prostoru možných funkcí 😄
Do loss funkce jde dát přesnost a velikost programu jako dvě objektivy, což některý architektury dělaj
Hmm, možná by šlo do loss funkce nějak přidat i generalizaci... Jak relativně velkou podmnožinu trénovacích dat to predikuje... To by možná šlo přidat do DreamCoderu...
"
In the AI field you have to first approximation these camps:
- Connectionists like to mimic the brain (neuroscience): artificial neural networks, deep learning, spiking neural networks, liquid neural networks, neuromorphic computing, hodgkin-huxley model,...
- Symbolists like symbol manipulation: decision trees, random decision forests, production rule systems, inductive logic programming,...
- Bayesians like uncertainity reduction based on probability theory (staticians): bayes classifier, probabilistic graphical models, hidden markov chains, active inference,...
- Evolutionaries like evolution (biologists): genetic algorithms, evolutionary programming
- Analogizers like identifying similarities between situations or things (psychologists): k-nearest neighbors, support vector machines,...
Then there are various hybrids: neurosymbolic architectures (AlphaZero for chess, general program synthesis with DreamCoder), neuroevolution, etc.
And technically you can also have:
- Reinforcement Learners like learning from reinforcement signals: reinforcement learning (most game AIs use it like AlphaZero for chess uses it, LLMs like ChatGPT start to use it more,...)
- Causal inferencers like to build a causal model and can thereby make inferences using causality rather than just correlation: causal AI
- DivergentSearchNoveltyMaximizers love divergent search for novelty without objectives: novelty search
And you can hybridize these too with deep reinforcement learning, novelty search with other objectives etc.
I love them all and want to merge them, or find completely novel approaches that we haven't found yet. :D
"