Deschide articolul original într-un tab nou Schimbarea acestei opțiuni se aplică imediat și reîncarcă pagina.
LECTURA
← Înapoi la Urmăritori

Subhadip Mitra

subhadipmitra.com/feed.xml
Vizitează site-ul

Cel mai recent articol

Ultima verificare

Urmăritori 0 urmăritori
Articole: 10

  1. Your OpenAPI Spec Is Not an Agent Interface: Compiling Governed MCP Tool Surfaces

    (Subhadip Mitra)

    Code: github.com/bassrehab/api-mcp-compiler · Docs: api-mcp.subhadipmitra.com · Install: pip install api-mcp-compiler TL;DR: pip install api-mcp-compiler compiles OpenAPI 3.x, Swagger 2.0, and WSDL 1.1 into MCP tool surfaces that are designed rather than…

  2. Beating FP16 with 4-bit Weights: A Portable W4A16 GEMM in Triton

    (Subhadip Mitra)

    Code: github.com/bassrehab/triton-kernels · Kernel: huggingface.co/kernels/bassrehab/w4a16 Last time I fused an entire Mixture-of-Experts forward pass into Triton and it beat Megablocks.

  3. The Activation-Cone Blind Spot, or Why Your Jailbreak Defense Can’t See Prefilling

    (Subhadip Mitra)

    TL;DR: I evaluated five jailbreak-defense paradigms across seven instruction-tuned models (7B-31B) and five attack families.

  4. What Runtime Interpretability Actually Costs, Part 1: The Case for Measuring It

    (Subhadip Mitra)

    TL;DR: The claim that activation probes are too expensive to run in production is folklore.

  5. DeepSeek DSpark: Speculation Is a Scheduling Problem

    (Subhadip Mitra)

    Last year I built speculative decoding from scratch and wrote about it here.

  6. Loop Engineering: Convergence Is Not Correctness

    (Subhadip Mitra)

    TL;DR: The June 2026 discourse on loop engineering is right about the shift and wrong about the hard part.

  7. Attention Is All You Bid: Advertising in Embedding Space

    (Subhadip Mitra)

    TL;DR: OpenAI is showing ads in ChatGPT.

  8. Beating CUDA with Triton: A Fused MoE Dispatch Kernel for Mixtral and DeepSeek

    (Subhadip Mitra)

    Paper: Cross-Platform Fused MoE Dispatch in Triton (arXiv:2605.23911) · Code: github.com/bassrehab/triton-kernels In my last post on Triton kernels, I optimized individual operations: RMSNorm, SwiGLU, INT8 GEMM.

  9. Confessions vs. CoT Monitoring vs. Probes: Three Bets on Model Honesty

    (Subhadip Mitra)

    TL;DR: OpenAI bets on confessions (ask the model to self-report).

  10. OpenAI’s Confessions Paper Has a Blind Spot. Here’s What Fills It.

    (Subhadip Mitra)

    TL;DR: OpenAI’s new “Confessions” paper trains GPT-5-Thinking to self-report misbehavior - reward hacking, scheming, sandbagging - via a decoupled RL objective.