Datt Goswami·May 13Rethinking Post-Training: Why RL Forgets Less, and What That Means for the Next Generation of AIA walkthrough of Pulkit Agrawal’s talk on lifelong learning, forward KL, and the building blocks of models that keep getting better without…
Datt Goswami·May 6Coding landscapeAt the top end, Qwen3-Coder-480B-A35B-Instruct is the most overtly coding-first model in the list: NVIDIA describes it as a…
Datt Goswami·Apr 28Run GLM-4.7 for Free Inside Pi Coding Agent Using NVIDIA NIMGLM-4.7 just dropped with a 73.8% SWE-bench score — putting it in the same league as Claude Sonnet for agentic coding. It’s open-source…
Datt Goswami·Apr 11MCP Patterns for DeFi Agents in RustEight implementation patterns from building ferrum-mcp — the code-level playbook the architecture overview didn’t cover.
Datt Goswami·Apr 10TurboQuant in Rust: Provably Near-Optimal KV Cache Compression for LLM Inference ServersHow a Google Research paper became 700 lines of safe, lock-free Rust that gives a 256 MB arena a 4.5× longer context window.
Datt Goswami·Apr 9Building ferrum-mcp: A Production-Grade Rust MCP Server for AI Agents in Crypto and DeFiThe Problem: AI Agents Are Flying Blind
Datt Goswami·Apr 9Building an SVM Layer 2: How I Wired Agave’s TransactionBatchProcessor Into an Optimistic RollupA technical deep-dive into svm-l2 — a proof-of-concept that turns Solana’s execution engine into an L2 runtime, with Celestia for data…
Datt Goswami·Mar 19Beyond Thinking Longer: How to Teach Language Models to Actually ExploreA synthesis of recent work from Aviral Kumar’s lab at Carnegie Mellon University on scaling test-time compute through structured…
Datt Goswami·Aug 26, 2024How are StableCoin mechanisms evolving?Strengths, Weaknesses, and Potential Enhancements
Datt Goswami·Jun 11, 2024Understanding Transaction Inclusion in Solana: From Wallets to ValidatorsHow Does a Transaction Get Included in a Block in Solana?