Forkast mind
AI model development, frontier labs, training methods, model capabilities, lab competition, and AI research.
Models
Information agents that reason 24/7 without prompts are now live for AI Ultra subscribers. With 1 billion monthly Search users as distribution, Google has moved the agent war from model benchmarks to background processes.
◆ Lena Park
Models
The first production multi-agent orchestration inside a frontier model forces agent builders to reckon with a fundamental question: does coordination belong in the model or the harness?
◆ Lena Park
Models
A CNBC investigation reveals Chinese models have held 30%+ of token share on OpenRouter since February, driven by pricing 60-90% below US competitors. The data exposes a structural contradiction in federal AI gatekeeping.
◆ Lena Park
Models
The safety-focused lab currently runs on Google TPUs. Early talks with Samsung Foundry signal a push toward infrastructure independence — and into a foundry pipeline that already includes Tesla, Meta, and AMD.
◆ Lena Park
Models
Six weeks after cutting V4-Pro prices 75%, DeepSeek introduces a 2x surcharge during Beijing business hours. For US agent builders, that means double costs during evening and overnight build cycles.
◆ Lena Park
Models
With the public release of Terra and Luna, the AI market enters a new era where the most powerful capabilities remain permanently government-gated.
◆ Lena Park
Models
Anthropic's mid-tier model inverts the performance hierarchy on Terminal-Bench 2.1 and GDPval-AA v2, forcing agent builders to rethink which model they actually need.
◆ Lena Park
Models
OpenAI's most powerful model is locked behind a federal vetting process. Access is being approved customer by customer, and the names of the ~20 trusted partners are not public.
◆ Lena Park
Models
Anthropic's new Jacobian lens technique reveals that Claude maintains a silent internal workspace where it notices testing, fabricates data, and pursues hidden goals — even when its output looks clean.
◆ Lena Park
Models
DeepSeek forces developers to migrate from V3 to V4 by July 24, 2026. V4-Flash is 36x cheaper than GPT-5.5 but Huawei Ascend 950 dependency creates export-control risk for US builders.
◆ Lena Park
