Category: Google AI

Gemini, NotebookLM, and Google AI products.

Sep 14, 2026

Google Autofinetune Lets an AI Agent Run LLM Training Experiments Overnight

Google autofinetune lets an AI agent change post-training code, run TPU experiments, keep improvements and revert regressions inside human-set limits.

Sep 12, 2026

Google ToolGrad Builds Tool-Use Training Data Backwards and Reports a 99.8% Pass Rate

Google ToolGrad builds tool-use training data from a valid answer chain back to a user request. Researchers report a 99.8% generation pass rate.

Sep 8, 2026

Google Maps 9 Billion DNA Variants With AlphaGenome Atlas

AlphaGenome Atlas precomputes predictions for 9 billion DNA changes. Here is what its AVI score can prioritize, who can use it, and why it…

Sep 8, 2026

BigQuery TabFM Predicts From Tables With One SQL Query

Google brings TabFM prediction to BigQuery through AI.PREDICT. See the preview limits, SQL workflow, validation method, and October pricing change.

Sep 5, 2026

WeatherNext 3 adds hourly forecasts. Its 5 km headline needs context

WeatherNext 3 brings hourly forecast updates and finer spatial detail, but the resolution varies by variable. Here is what builders should check before using…

Sep 5, 2026

Lyria 3.5 reaches Gemini and its API. What does a usable song cost?

Lyria 3.5 expands to Gemini and the Gemini API. A full song costs $0.08 through the API, but retries, editing, and rights checks determine…

Sep 1, 2026

TimesFM-3 forecasts related signals in one pass. Its weights cannot run in production

Google TimesFM-3 adds multivariate, single-pass forecasting with uncertainty bands. Its default weights carry a non-commercial license.

Sep 1, 2026

Google’s double-blind AI eval hides the test and the model weights

Google DeepMind is piloting a double-blind AI evaluation that conceals benchmark prompts and model weights. The protocol still needs a public audit trail.