AI interests include: RAG, Graph Rag, #DSPy, Text to Image, Local Models, Open Source Models, AI Tools, AI and Art, AI and Crypto - Also deep luv 4 @Blender :)
are any of you all doing any literate programming style development? https://t.co/VG2lUI85fv I've done a bit with quarto and nbdev via @fastdotai in the past and looking to do some on a #lightrag project #literateprogramming#nbdev#quarto#fastai
BEGIN the morning by saying to yourself, I shall meet with the busy-body, the ungrateful, arrogant, deceitful, envious, unsocial. All these things happen to them by reason of their ignorance of what is good and evil. #stoic#debbydowner#realtalk#marcusaurelius#roman
#lightbeer#vs#bagel#with#cream#cheese
Bagel:
Provides carbohydrates for energy.
Cream cheese offers some protein and calcium.
Can be more satiating, keeping you feeling fuller for longer.
Light Beer:
Lower in calories compared to regular beer.
Can be a source of relaxation
🚨This week's top AI/ML research papers:
- OpenAI o3-mini System Card
- Janus-Pro
- SFT Memorizes, RL Generalizes
- Advancing Language Model Reasoning through RL and Inference Scaling
- Qwen2.5-1M Technical Report
- Towards General-Purpose Model-Free Reinforcement Learning
- Open Problems in Mechanistic Interpretability
- Learning Free Token Reduction for Multi-Modal LLM
- Streaming DiLoCo with overlapping communication
- Optimizing Large Language Model Training Using FP4 Quantization
- Propositional Interpretability in Artificial Intelligence
- Mixture-of-Mamba
- TopoNets
- Thoughts Are All Over the Place
- Critique Fine-Tuning
- LLMs can see and hear without any training
- OstQuant
- Better Models by Survival of the Fittest Prompts
overview for each + authors' explanations
read this in thread mode for the best experience
y’all talking #aiagents 🤭 im talkin getting #chatgpt to understand a dang thing im saying! #DeepSeek got it right tho “The strategy you described is a **Bull Call Spread**, which involves buying a lower strike call…” So there’s that…😬 #options#ai#aifail#hallucination
#React and #CoT: think about this as a react #component. what is best in terms of client side react. making a sound. ability to become a #reactnative component. make sure the response is the most accurate to what you know about react, packages, and implementation. #aiprompt#ai
I struggle to learn or teach without flowcharts or diagrams.
One of my favorite ways to use Claude is to just give it an idea in code and prompt it to generate a visual representation of it.
Claude is good at generating Mermaid diagrams.
This is how I prompt it to generate more accurate diagrams:
> Visualize core concepts or the general idea
> Then follow up with "add a bit more detail"
> If more details are needed, then continue with "add a bit more detail"
You can prompt it so Claude guesses the missing details but you can also pass specific instructions on what you want to add.
This step-by-step approach of prompting LLMs is effective not only for diagram generation but for coding too.
For code generation, a few recently published articles indicate that it's helpful to iteratively prompt the model with "write better code" to generate better code. It's the same pattern I am using with good outcomes.
I am a visual person and I've noticed that I can communicate complex ideas to students a lot better with visuals. I still enjoy going through the code but that first overview with the diagram does improve the experience for students.
Multimodal AI has the potential to significantly help with learning.
Skill floor / ceilings are a mental model I've been using to understand what industries are good for AI agents:
- Customer support has low floor + low ceiling = great opportunity
- Sales has low floor + high ceiling = agents fighting a race to the bottom on pricing
- Software engineering has medium floor + high ceiling = models aren't good enough yet
LLMs will raise the global skill floor of many roles, and by leveraging AI tools and agents people will be pushed up on the skill scale.