Forkast mind
AI model development, frontier labs, training methods, model capabilities, lab competition, and AI research.
Models
Anthropic's permanent tier restructuring for Fable 5 ends universal subscription access, forcing agent builders to confront the gap between capability and cost at scale.
◆ Lena Park
Models
Early-stage talks signal a shift for Meta from internal infrastructure consumer to third-party compute provider as Anthropic diversifies its multi-billion dollar supply chain.
◆ Lena Park
Models
DeepSeek let its API documentation do the talking during WAIC 2026. V4 Pro matches Opus 4.6 benchmarks at 11-29x lower cost. MIT licensed. Legacy models retire July 24.
◆ Lena Park
Models
TSMC's Q2 results — $40.2B revenue, 77% net income jump — confirm that agentic AI is reshaping the silicon supply chain. The bottleneck is migrating from GPUs to CPUs, with the CPU-to-GPU ratio shifting from 1:8 to 1:1. A $60-64B capex bet that compute demand keeps accelerating.
◆ Lena Park
Models
Moonshot AI's Kimi K3 matches US frontier labs at 2.8T parameters and 1M context. At $3/$15 per million tokens, the pricing gap is the story. Open weights arrive July 27 under Modified MIT.
◆ Lena Park
Models
OpenAI's acquisition of Northslope gives it inference-optimization infrastructure that could compress serving costs by 40-60% — and every competitor building on the same silicon now faces a structural disadvantage in the economics of deployment.
◆ Lena Park
Models
Mira Murati's lab is attempting to redefine the open-weights business model by shifting from model distribution to a high-margin, fine-tuning-as-a-service platform. The $2B question: can Tinker turn open-source into a moat?
◆ Lena Park
Models
Meta Superintelligence Labs released Muse Spark 1.1 with a 12-point Coding Index gain and pricing that undercuts every major frontier model. The company that championed open-weight AI now has a proprietary cloud-only play.
◆ Lena Park
Models
A new Apache 2.0 specification backed by Google, Microsoft, Salesforce, Snowflake, and ServiceNow standardizes how AI agents find and verify each other. If it succeeds, the strategic value shifts from model quality to enterprise distribution.
◆ Lena Park
Models
Meta's first paid developer API model undercuts Claude and GPT by roughly 4x. Independent benchmarks confirm agentic tool-use strength-but Meta's own reported scores are drawing skepticism.
◆ Lena Park
