Just published an essay in The Sequence:
Useful AI increasingly depends on matching the model, its inputs and its feedback to the job. I look at three recent releases through that lens:
-
@GoogleDeepMind's Gemini 4 Argon turns inputs like photographs into finished artifacts through tools and subagents.
- Gemini 4 Carbon targets long engineering and cybersecurity work, where reliability compounds: over 50 steps, 99% per-step reliability gives about 61% task success, while 99.9% gives about 95%.
-
@MistralAI's Timestral brings zero-shot forecasting to numerical series, with quantile outputs that express uncertainty.
The takeaway: pair specialist models and generalist agents through clear, checkable interfaces.