Reasoning models aren't really reasoners, its basically neural style transfer protocol where you force a model "decoder" to emit tokens in a style that appears to be Reasoning like a deductive thinking.
Using LLMs as Optimizers: A new population-based method called LEO that leverages large language models for numerical optimization tasks like nozzle shape and windfarm layout design. Shows comparable results to traditional methods while highlighting unique challenges of LLM-based optimization
classic NN takes a vector of data through layers to make a prediction. Backprop adjusts network weights till predictions are right. These network weights form a vector, and training changes this vector till it hits values that mean "trained network".
Neural ODE reframes this: instead of focusing on the weights, focus on how they change. It sees training as finding a path from untrained to trained state. At each step, it uses ODE solvers to compute the next state, continuing for N steps till it reaches values matching training data. This gives you the solution for the trained network.
LLM are fundamentally stateless. The exoskeleton of agents is essentially a work around to enable a state that resides externally to a LLM. This is required in some manner for creating an agent. Task oriented agents require reasoning and planning but the nature of its wildly different compared to the kind of behaviour that is observed in the interactive simulacra paper. Your ideas have merits but I feel it needs to have some further qualifications
The hallucination may not be a bug but a form regularisation that will helps sample behaviour and that itself might be of great use as it might act as a proxy for focus groups
Honestly the conversation of AI researchers on climate science seems so superficial and removed from reality of climate change. It's like they are trying to solve a drought by capturing condensation on a glass of cold beer.
The major emissions sectors are agriculture, transportation, cement and metal, energy usage in primary sector for thermal processes.
For someone who seems to have dedicated their lives to climate and AI solving energy intensity of deep learning isn't even relevant let alone a factor in emissions, also better climate models aren't the solution for reducing emissions.
We need not only new tech but a new socio-technical system for reduction of emissions. Rice cultivation, cement, waste, heating of houses in northern latitudes all generate sizeable quantum of emissions what's AI going to do there to solve these?
Optimization can squeeze some gains but we need revolution in fundamental science and engineering capabilities to quickly discover, synthesize and scale the manufacturing of new materials whilst rewiring the every elememt of the the supply chain to support such transition. This is a problem for AI- yes, but is this where bulk of climate AI research is focused on? Sadly not!
What is more motivating and upbeat though is the fact scientists from traditional engineering domains are increasingly discovering AI. SciML work by Chris rackauckas PiNN by karniadikis, neural operators by anima Nvidia are some of the best forays into this space. I have not mentioned the frentic research happening in AI and microbiology which can be a game changer in invention of biotechnology solutions for climate change.
Yet sadly none of this is ever mentioned or covered in climate AI related conversations.
Agree. So attention is like hierarchy of graphs where nodes are tokens and edges are attention scores per head.
Now what's trippy is this node has position data. So node feature and position it appears is used to create a operator that projects a sequence to a semantic space.
This seems to work for any modality of data.. so there is some thing about order in which data appears that seems to be linked to semantics and for me hints about some deep causal structure being latent in LLM
In CNN the layers seem to learn geometric primitive and deeper layers seem to learn more complex geometric patterns loosely speaking.
In transformer what do query key matrices learn? How are their weights somehow working to extract context no matter which word appears in which position?
AI for science is much bigger than RL or Generative AI in science.
There are several classes of models Like operator learning, physics informed neural networks, Fourier operators
That perform magnificently well and have killer applications in various industrial settings
Do read the attached paper if you're curious about AI in science