Self-Improving AI & The Talent War
In this episode:
Self-Improving AI & The Talent War — Andrej Karpathy officially joins Anthropic to work on self-improving AI, sparking debate over recursive improvement risks and regression without human oversight. Meanwhile, China restricts overseas travel for AI researchers at Alibaba and DeepSeek, treating expertise as a national security asset amid a documented 89% collapse in China-to-US researcher migration.The Crowded Reasoning & Coding Race — The SWE-rebench leaderboard returns with 110 fresh tasks and growing demand for local model evaluations over frontier ones. Benchmark fatigue is setting in as practitioners argue that standardized scores are meaningless without domain-specific evals, cost-per-fix metrics, and real-world edge case testing.Big Models on Small Hardware — Krasis v1.0 runs Qwen3.6-35B at reading speed on an 8GB laptop GPU using mixed-precision HQQ quantization with a pure Rust hot path. At the other extreme, the 26M-parameter Needle model beats Qwen3-0.6B on CPU tool routing by always attempting action rather than falling silent under ambiguity.Harness Engineering & Agent Tooling — Harness engineering emerges as the discipline of wrapping AI coding agents in rules, hooks, memory, and orchestration scaffolding to reduce errors and automate multi-session workflows. New tools like piia-engram for persistent developer identity, Tactile for accessibility-first computer use, and OpenHack for agent-driven security review expand the ecosystem.Security: One Bad Character — A trailing slash bypassed AWS API Gateway JWT auth on a fintech app, exposing full account data and enabling unauthorized wire transfers due to a path-matching mismatch between the routing and auth layers. The pattern repeats across frameworks: trivial input normalization gaps defeat sophisticated auth middleware.Training Smarter: Sparse Deltas & Private Fine-Tuning — Hugging Face's delta weight sync exploits bf16 arithmetic sparsity to shrink per-step model syncs from gigabytes to megabytes, enabling fully disaggregated RL training across clouds. PACZero achieves near-baseline fine-tuning accuracy at privacy levels where all differential privacy baselines collapse to chance.Beyond the Cloud: Mesh Networks, Neuromorphic Chips, and When AI Breaks — Reticulum emerges as the most promising mesh networking stack for its transport-agnostic, cryptographically-addressed architecture spanning LoRa to Tor. A neuromorphic computer combining quantum tunnelling with brain-inspired design tackles combinatorial optimization, while Starbucks scraps its AI inventory tool after real-world data coordination failures.Keywords: agent-orchestration, ai-failures, anthropic, api-security, auth-bypass, automation, aws, benchmarks, china-ai-restrictions, coding, delta-sync, developer-tools, differential-privacy, edge-computing, edge-inference, fine-tuning, hardware-efficiency, harness-engineering, karpathy, krasis
