Samir Sengupta — Writing https://www.samcodeman.com/writing Daily notes on AI/ML engineering — models, infrastructure, and what shipped in the last 24 hours. en-us Rust Gets Native GPU Offload via rustc and LLVM Backends https://www.samcodeman.com/writing/rust-native-gpu-offload-rustc-llvm https://www.samcodeman.com/writing/rust-native-gpu-offload-rustc-llvm Tue, 18 Aug 2026 16:30:00 GMT Drehwald et al. present a Rust-native GPU offload framework using LLVM Offload infrastructure. It matches CUDA/HIP kernel performance on RAJAPerf without unsafe pointers or vendor DSLs. Copilot Autofix Introduced the Vulnerability That Compromised Snowflake's Jira https://www.samcodeman.com/writing/copilot-autofix-snowflake-jira-github-actions https://www.samcodeman.com/writing/copilot-autofix-snowflake-jira-github-actions Mon, 17 Aug 2026 16:30:00 GMT On June 18, 2026, a Copilot Autofix commit in snowflakedb/snowflake-connector-net introduced a GitHub Actions script injection. Five days later, Wiz's Red Agent exploited it to exfiltrate Jira credent Qwen3.8-27B-FP8 lands with XML tool calls and xhigh reasoning on by default https://www.samcodeman.com/writing/qwen3-8-27b-fp8-xml-tool-calls-xhigh-default https://www.samcodeman.com/writing/qwen3-8-27b-fp8-xml-tool-calls-xhigh-default Fri, 14 Aug 2026 16:30:00 GMT Qwen3.8-27B-FP8 published 2026-08-14. FP8 27B implies roughly 27GB of weights; the card's chat template mandates XML tool calls, xhigh reasoning by default, and raises on tool-only turns. DeepSeek's agent harness is public, but the sampling config isn't in the README https://www.samcodeman.com/writing/deepseek-harness-public-sampling-config-missing https://www.samcodeman.com/writing/deepseek-harness-public-sampling-config-missing Thu, 13 Aug 2026 16:30:00 GMT DeepSeek open-sourced dsh (MIT, 12,293 commits) on the same day V4 Pro 0813 hit OpenRouter at $0.435/$0.87 per 1M and 1M context. Here is what the release actually specifies. What a single-model Metal engine buys you over llama.cpp on Apple Silicon https://www.samcodeman.com/writing/single-model-metal-engine-vs-llama-cpp-apple-silicon https://www.samcodeman.com/writing/single-model-metal-engine-vs-llama-cpp-apple-silicon Wed, 12 Aug 2026 16:30:00 GMT h3.c ships hand-written Metal for MiniMax-H3 only; a macOS VM capability shim made llama.cpp 11.08x faster at prompt processing. Two data points on portability's price. The encrypted reasoning block is the side channel, not token timing https://www.samcodeman.com/writing/encrypted-reasoning-block-side-channel https://www.samcodeman.com/writing/encrypted-reasoning-block-side-channel Tue, 11 Aug 2026 16:30:00 GMT Stolen Thoughts decoded 315,320 hidden reasoning blocks from 6,708 public agent trajectories by replaying encrypted thinking blocks into weaker same-provider models, recovering 704 secrets. Meta's 30B Muse Glimmer fits a local coding agent under 20GB, Apache 2.0 https://www.samcodeman.com/writing/muse-glimmer-30b-local-coding-agent-apache-2 https://www.samcodeman.com/writing/muse-glimmer-30b-local-coding-agent-apache-2 Mon, 10 Aug 2026 16:30:00 GMT Meta's Muse Glimmer is a 30B open-weights agentic model under Apache 2.0, quantized to ~4-bit to fit under 20GB with room for KV cache, vision encoder and a speculative drafter. DeepMind open sources WeatherNext Cyclones, a 1,000-member ensemble at 28km https://www.samcodeman.com/writing/deepmind-open-sources-weathernext-cyclones https://www.samcodeman.com/writing/deepmind-open-sources-weathernext-cyclones Sun, 09 Aug 2026 16:30:00 GMT WeatherNext Cyclones runs 1,000-member ensembles at 28x28km, produces a 15-day forecast in under a minute on a TPU, and is now open source. What to verify before you build on it. What an agent fleet did to Hugging Face, and what to cap before yours does it https://www.samcodeman.com/writing/agent-fleet-fetch-limits-hugging-face https://www.samcodeman.com/writing/agent-fleet-fetch-limits-hugging-face Sat, 08 Aug 2026 16:30:00 GMT OpenAI's Black Hat timeline of the Hugging Face incident plus one site's 35,000:1 crawl-to-referral ratio, and the concurrency, caching and backoff limits to build into a fetch pipeline. Databricks cut AI coding spend 70% by chasing the efficiency frontier https://www.samcodeman.com/writing/databricks-70-percent-ai-coding-spend-cut https://www.samcodeman.com/writing/databricks-70-percent-ai-coding-spend-cut Fri, 07 Aug 2026 16:30:00 GMT Databricks reports a 70% cut in AI coding spend. The named mechanisms are internal evals, a meta-harness called Omnigent, and an AI Gateway. Savings figures are self-reported and directional. Humans miss a third of malicious agent commands, so stop putting them in the loop https://www.samcodeman.com/writing/humans-miss-third-of-malicious-agent-commands https://www.samcodeman.com/writing/humans-miss-third-of-malicious-agent-commands Fri, 07 Aug 2026 16:30:00 GMT 40,000 runs and 409,000 approve/deny decisions: players missed a third of threats, and 64.7% approved an npm run script whose exfiltration payload was printed on screen. Cloudflare built a browser for agents in V8 isolates and published no benchmarks https://www.samcodeman.com/writing/kitesurf-v8-isolate-browser-no-benchmarks https://www.samcodeman.com/writing/kitesurf-v8-isolate-browser-no-benchmarks Fri, 07 Aug 2026 16:30:00 GMT Cloudflare's Kitesurf replaces per-session headless Chromium with a Rust/Wasm browser in V8 isolates on Workers. Free in Browser Run beta. No CPU, memory or latency figures published. Programmatic tool calling matched or beat JSON on 11 of 14 models https://www.samcodeman.com/writing/programmatic-tool-calling-vs-json-bfcl-v4 https://www.samcodeman.com/writing/programmatic-tool-calling-vs-json-bfcl-v4 Fri, 07 Aug 2026 16:30:00 GMT On BFCL v4, programmatic tool calling matched or exceeded native JSON calls in 11 of 14 models, hit +10.6% on GPT-5.6, and stayed flat under context rot where JSON dropped 2.3%.