Thoughts, experiences, and technical insights from my journey in software development.
Popular hashtags
1,178 AI staff from OpenAI, Anthropic, Google and Meta signed an open letter asking the US to build tools capable of slowing AI when needed.
MCP drops sessions entirely in the 2026-07-28 spec, switching to a stateless request/response model. No sticky sessions. No shared storage. But migration has a cost — here's what developers need to know.
Bun's AI rewrite was hailed as a triumph: 535K lines ported in 11 days for $165K. Six weeks later, 2,475 open PRs and costs nearing $800K tell a different story.
Formal verification used to cost 10x the coding effort. LLMs are collapsing that barrier — proving Zstandard correct in Lean, end to end, took just 20 minutes.
An ADB maintainer at Google proposes blocking local ADB loopback — a move that could wipe out Shizuku and dozens of open-source developer tools.
Anthropic launches Claude Opus 5: #1 SWE-bench 97%, near-Fable 5 quality at half the cost. Same $5/$25 pricing, double Frontier-Bench vs Opus 4.8.
Etched just raised $300M at a $10.3B valuation — doubling in 7 months. Its Sohu ASIC for transformer inference is a direct challenge to Nvidia.
A Rust tokenizer hits 24.53 GB/s — ~1000x faster than HuggingFace, ~700x faster than tiktoken. Drop-in replacement. It changes how you handle LLM data pipelines.
Google released Gemini 3.6 Flash — 17% fewer output tokens, lower price, higher benchmarks. Flash-Lite hits 350 tok/s at $0.3/1M input, built for scaling agentic workflows.
US AI market share on OpenRouter collapsed from 70% to 30% in a year. Chinese open-weight models dominate token volume — and developers are the first to benefit.
In just 3 days, Alibaba and Moonshot AI unveiled Qwen 3.8 (2.4T params) and Kimi K3 (2.8T params) — both going open-weight. Here's what it means for developers and the global AI race.
PrismML's Bonsai 27B compresses a full 27B-parameter model to 3.9GB, running on an iPhone 17 Pro at 11 tokens/s. The ternary variant retains 95% of full-precision benchmarks across 15 evaluation suites.