The updated Deep Think mode outperforms Gemini 3 Pro on Humanity’s Last Exam and ARC-AGI-2 – raising the bar for how it handles complex math, science, and novel challenges.
RESEARCH
-

Energy Efficiency Becomes Critical Metric for AGI Development
By
–
Compute and energy are the two major bottlenecks to creating an industrial-scale ‘general intelligence’ … which is why efficiency now matters more than ever. The real metrics that count are tokens-per-watt and tokens-per-flop. In other words: how much intelligence can you
-
Gemini 3 Deep Think Mode Now Available for Ultra Users
By
–
Glad to share that Gemini 3 #DeepThink mode is now available for Ultra users in the Gemini app! This is the engine behind the success of our #IMO2025 Gold🥇with the upgraded power of Gemini 3! Deep Think uses our novel parallel thinking approach to explore multiple hypotheses to… https://t.co/4M9JmP2TdX
— Thang Luong (@lmthang) 4 décembre 2025Glad to share that Gemini 3 #DeepThink mode is now available for Ultra users in the Gemini app! This is the engine behind the success of our #IMO2025 Goldwith the upgraded power of Gemini 3! Deep Think uses our novel parallel thinking approach to explore multiple hypotheses to
-
Cerebras showcases nine papers at NeurIPS 2025 spanning pretraining to inference
By
–
Cerebras is showcasing nine new papers at NeurIPS 2025, spanning pretraining to inference:
CODA – orchestrate 32B models to beat 235B models Calibrated Reasoning – 1-3x token savings in best-of-n sampling
DREAM – 3.6× faster VLM inference using spec decode CompleteP – -
Testing Gemini’s Performance on Complex Mathematical Proofs
By
–
Trying it for some math proofs that Gemini previously had trouble with
-
Vanessa Rosa Presents Little Martians at NeurIPS Conference
By
–
Packed house for @va2rosa introducing Little Martians at NeurIPS pic.twitter.com/7bm4161sf1
— Gene Kogan (@genekogan) 4 décembre 2025Packed house for @va2rosa introducing Little Martians at NeurIPS
→ View original post on X — @genekogan, 2025-12-04 19:52 UTC
-
Chain of Debate: Beyond Chain of Thought in AI Reasoning
By
–
Here's the post I put on LinkedIn on "Chain of Debate" a while back: Chain of thought was just chapter one. Next up: Chain of Debate. When AIs started sharing their chain of thought (CoT), it gave us a window into its thinking – how it reasoned its way from A to B. But at the
-
Reasoning Models Offer Better Answers But Consume Significantly More Energy
By
–
Reasoning models can give more useful answers to certain types of questions. They also use A LOT more energy, says new research from Hugging Face's @SashaMTL and Salesforce's @bgamazay bloomberg.com/news/articles/… @rachelmetz & I take a look
→ View original post on X — @rachelmetz, 2025-12-04 18:41 UTC
-

Meet Jeff Dean and Gemini team at NeurIPS Conference
By
–
See you at 2:30 pm with @JeffDean and the Gemini team @NeurIPSConf!
→ View original post on X — @oriolvinyalsml, 2025-12-04 18:28 UTC
