Articles from blog.bytebytego.com

Microsoft scales AI agents by treating the harness—runtime, retrieval loops, identity, and continuous evaluation—as critical as the model itself for production reliability.

OpenAI splits WebRTC into a stateless relay and a stateful transceiver, using the ICE ufrag for routing to serve 900M voice AI users with low latency.

Thinking Machines argues current turn-based AI voice systems have a scalability ceiling and proposes interaction models using time-aligned micro-turns and a two-model architecture for true real-time collaboration.

LLM inference splits into compute-bound prefill and memory-bound decode, driving optimization techniques like batching, quantization, speculative decoding, and disaggregation.

Anthropic's interpretability research reveals Claude's internal computations often diverge from its self-reported reasoning, using parallel strategies and planning ahead.

OpenAI's Codex agent relies on an orchestration layer handling prompt assembly, context management, and a custom protocol for multi-surface support.

Stripe's unattended 'Minion' coding agents merge over 1,300 PRs weekly by leveraging pre-existing developer infrastructure, hybrid blueprints, and layered feedback loops.

Open-weight LLMs now uniformly use Mixture-of-Experts transformers, with key differences in attention mechanisms, expert count, post-training via RL, and permissible licenses.

The X (Twitter) For You feed algorithm uses a Grok-based transformer model and open-sourced components to replace hand-crafted rules with ML, retrieving and ranking posts via in-memory stores and similarity search.