Newsletter Archive

GitHub Copilot is switching from per-request pricing to usage-based billing on June 1, 2026, charging by token consumption (input, output, cached) aligned with API rates. All plans get a monthly credit allotment; paid tiers can buy overages. The shift reflects Copilot's evolution into an agentic platform with multi-step autonomous coding sessions that cost far more than single queries under the old model.

GitHub Blog · 1 min read Industry

OpenAI ends Microsoft's exclusivity rights in their partnership, now free to offer models through any cloud provider while Azure remains first-to-ship. Microsoft keeps a non-exclusive license through 2032 but stops paying OpenAI revenue share; OpenAI's payments to Microsoft cap out in 2030.

Engadget · 2 min read Industry

Google unveiled two separate TPU V8 chips optimized for different workloads—a first in the TPU program's history. The split design lets Google tailor silicon for specific tasks rather than chasing pure scale, positioning TPUs as viable alternatives to Nvidia accelerators for both training and inference.

Tom's Hardware · 1 min read Hardware

Six specific failure patterns that wreck production LLM systems, demonstrated through real customer incidents like a chatbot confidently lying about feature pricing. The piece prioritizes testing for real-world failure modes over relying on benchmark scores.

DZone AI · 1 min read Community

Experts are pushing for an immediate halt on AI chip exports to China following a White House warning about model distillation—a technique where advanced AI models are compressed into cheaper, deployable versions that bypass export controls.

Nextgov · 4 min read Policy

OpenAI and Microsoft are dropping their exclusive AI model deal, freeing both to strike competing partnerships. Separately, China rejected Meta's $2B acquisition of AI startup Manus, and Elon Musk is suing OpenAI, claiming it abandoned its non-profit founding mission.

Bloomberg Tech · 1 min read Industry

Applied Intuition's CEO and CTO discuss their evolution from autonomy tooling into a $15B physical AI platform powering driverless trucks, mining equipment, and defense systems across 30+ products. The real bottleneck isn't model intelligence anymore—it's deploying AI onto constrained hardware safely and reliably.

Latent Space · 69 min read Industry

Microsoft released VibeVoice, a Whisper-alternative speech-to-text model with speaker diarization, on January 21st. A developer ran it on a Mac and transcribed an hour of audio in 8m45s using an MLX quantized version, with detailed performance metrics and working code examples.

Simon Willison · 3 min read Tools