NESTFRONTIER / ARCHIVE
Technical analysis, model releases, research, and deployment notes.
Unconventional AI released Un-0, an image generator running on coupled oscillators instead of neural networks. FID 6.74 on ImageNet 64x64. The hardware to make it 1000x more efficient does not exist yet.
DeepSeek V4 charges $0.14/M tokens. Claude charges $5.00. That 35x gap is structural, not temporary, and it's forcing AI labs to confront whether their pricing was ever real.
Anthropic told the Senate that Alibaba ran 28.8M exchanges through Claude using 25K fake accounts. Nearly double the previous record set by three labs combined.
Krea 2 drops open weights for a 12B image model with a two-checkpoint workflow: train on Raw, generate fast on Turbo. Zero synthetic training data, 2-second inference.
Alibaba's Qwen-AgentWorld is the first language model built to simulate agent environments, and training in its fake worlds beats training in real ones.
Anthropic updated its privacy policy to collect government IDs and facial geometry through Persona, a Thiel-backed firm whose code was found on a US government server months after Discord dropped them.
A 3B parameter model from Sina Weibo claims to match Claude Opus 4.5 on math benchmarks. The AI community is split on whether this is a breakthrough or benchmark gaming.
Noam Shazeer, co-author of the Transformer paper, left Google for OpenAI less than two years after Google spent $2.7 billion to bring him back. The move came 10 days after OpenAI filed its IPO prospectus.
A new 0.22B parameter model matches or beats 10B-level inpainting models like FLUX and SD3.5 using less than 2% of the parameters and 15x faster inference.
Sakana AI's Fugu is a 7B model trained to orchestrate frontier LLMs. It beats GPT-5.5 on SWE-Bench Pro while using 6x fewer tokens than existing multi-agent frameworks.