DeepSeek V4 Flash cuts model depth to 43 layers, bucking the trend of deeper-is-better. Architecture innovation over scale brute force.
Texas just froze new data center approvals, putting 20% of the US pipeline at risk. This is what compute centralization looks like as a single point o
Alibaba ships Qwen3.8 Max with 2.4 trillion parameters and calls it the end of the fully free open source era. A 2.4T open-weight model from a non-US
20% of the US project pipeline faces delays while regulators audit power infrastructure.
Centralized compute at this scale demands billions in capital, years of construction, and entirely new energy infrastructure. Those economics support
Training kernel optimization is an underappreciated bottleneck in distributed training.
We are postponing the launch of the SOMA SN114 Conviction Program. 🧵 1/2
Anthropic is building a custom chip design team. The motivation is clear: there is simply not enough hardware to train the models they want at the sca
Anthropic commits $10B in compute capacity with Volta, a cloud startup backed by a $4.7B lease at a Bitdeer facility.
The White House AI safety framework mandates 30-day pre-release review for closed models and explicitly exempts open source.
DeepSeek V4 Flash at $0.14 per million tokens. A production-grade coding model at a price that makes local inference competitive with cloud APIs.
Week 12. A shipping week for the confidential-compute stack:
Hugging Face's CEO pushes back on a US kill switch bill, arguing for open model access instead.
Alibaba opens Qwen3.8 Max weights next week, a 2.4 trillion parameter model. Frontier-scale.
OpenAI cut GPT-5.6 Luna prices by 80% to $0.20 per million tokens. Steep. Inference pricing keeps racing toward zero as competition from open models a
DeepSeek V4 Flash runs at 62 tok/s on 4 RTX 5090s. Consumer hardware. A 284 billion parameter model running inference on four gaming GPUs. That combin
Merged PR #3019: basket-migration-multi-block
RT @DistStateAndMe: TEST_RUN_001... completed.
230 signers, one holdout. The Nvidia open weights letter has crossed a serious threshold of industry support. Open model weights are becoming the norm
Andrew identifies the pressure Templar is built around. Model training is a recurring cost, and each generation demands more compute before the last o
Open Source Must Win 🫡 https://t.co/dKZxh5Z3ZT
Hash Rate - Ep. 180: Stillcore 'State of Subnets' Report $TAO