Aug 8, 2026

AMD is buying AI-chip startup Taalas, betting on model-specific chips that etch a model directly into silicon. Early demos reportedly hit up to 17,000 tokens per second.

Why it matters: A bet that drops the general-purpose GPU assumption and bakes a specific model into the circuit for faster inference. The debate centers on whether the cost and speed floor of inference can move off GPUs.

DeepSeek's new model was independently scored on ARC-AGI: at max effort, 89.0% on ARC-AGI-1 ($0.02/task) and 61.4% on the harder ARC-AGI-2 ($0.04/task), reported across three reasoning levels.

Why it matters: Frontier-level benchmark results at a low per-task cost drew the discussion. It's an independent check on the cost-to-performance of open-weight models.

OpenJDK, stewarded by Oracle, banned AI-generated code contributions over safety, security, and IP risk. Private use of LLMs for debugging or review is fine, but AI-generated material can't be submitted to repos or PRs.

Why it matters: Debated against Ellison's claim that AI now writes Oracle's own code. A concrete case of where an open-source project draws the line on accepting AI-generated contributions.