2025-W23 › 🔗 › 2025-06-06 [2025-06-06]

#agent #apl #clojure #compiler #dspy #formal #fuzzing #interop #news #optimization
- investigating DSPy (Demonstrate-Search-Predict)
    - Pipelines & Prompt Optimization with DSPy
    - found MLflow DSPy Flavor
    - SIMBA (Stochastic Introspective Mini-Batch Ascent)
        - proceeds in a sequence of mini-batches seeking to make incremental improvements to the prompt instructions or few-shot examples
        - Advanced Tool Use: using SIMBA to improve a "Multi-Hop Tool Use" agent from 35% accuracy to 60%
            - in RL for Multi-Hop Research, RL method is compared with SIMBA
        - Correcting Bias in LLMs with DSPy
        - Grok has difficulty figuring out the corresponding paper, even with Deep Research
    - related
        - found TextGrad (code, notebook)
        - found Zenbase
        - found Easy Dataset: Custom Prompts
            - i.e. I can use extra system prompt to generate Q&A pairs then use them to finetune an LM or optimize via DSPy to answer questions without such prompts
- The next phase of jank's C++ interop
    - found compiler-research/CppInterOp: A Clang-based C++ Interoperability Library
        - cppyy (code)
    - jank is C++ (on HN)
- Sharing everything I could understand about gradient noise
- Small Programs and Languages
- APL tutorial
- John Carlos Baez on the shape of the Oort cloud
    - paper: A Spiral Structure in the Inner Oort Cloud
- found Jepsen, a framework for distributed systems verification, with fault injection, written in Clojure
    - Fuzzer Blind Spots (Meet Jepsen!)
    - Jepsen: TigerBeetle 0.16.11
    - The unreasonable effectiveness of fuzzing for porting programs (on HN) (on lobste.rs)
- AI is a gamechanger for TLA+ users
- In which I have Opinions about parsing and grammars
- Experimenting with no-build Web Applications
- Zigar 0.14.1: stream integration and more
- The Lean FRO Year 2 Roadmap
- found No More Shading Languages: Compiling C++ to Vulkan Shaders
- found APLearn: machine learning library
- I Think I’m Done Thinking About genAI For Now