**Gemini 3 Pro Dominates Math & Trading as NVIDIA’s GPT-OSS-120B Eagle Boosts AI Speed**

AI Hardware & Infrastructure

World’s Smallest AI Supercomputer: Tiiny Ai Pocket Lab: Guinness World Records verifies the Tiiny Ai pocket Lab as the smallest AI supercomputer, capable of running a 120B-parameter model locally. The palm-sized device features 80GB LPDDR5X RAM, 160 TOPS dNPU compute power, and leverages "TurboSparse" and "PowerInfer" architectures for efficient inference.

• World’s smallest AI supercomputer: Tiiny Ai pocket Lab—the size of a power bank

AI Models & Benchmarks

Gemini 3 Pro Excels in Novel Math Visualizations: The model demonstrates advanced problem-solving by generating a pure geometry-based proof for the Dirichlet integral—an original solution not present in its training data.

• Gemini 3 Pro is extremely good at generating new math visualizations

Gemini 3 Pro Leads in Vending-Bench Trading Simulation: In a 350-day simulated trading benchmark, Gemini 3 Pro ranks first, followed by Claude Opus 4.5 and GPT-5.2 (tied with Sonnet 4.5), showcasing its performance in financial decision-making.

• GPT 5.2 comes in 3rd on Vending-Bench, essentially tied with Sonnet 4.5

GPT-5.2 Underperforms in RAG Use Cases: Benchmarks reveal GPT-5.2 lags behind GPT-5.1 and other models (Claude, Grok, Gemini) in Retrieval-Augmented Generation tasks, producing shorter, less consistent answers.

• GPT 5.2 underperforms on RAG
  • Detailed analysis

Google DeepMind Updates Gemini Native Audio Model: The latest version improves real-time instruction following, function calling precision, and conversational smoothness, enhancing voice interaction capabilities via the Gemini API.

• Google Deepmind: Gemini rolling out an updated Gemini Native Audio model
  • Official blog post

Model Optimizations & Tools

NVIDIA Releases GPT-OSS-120B-Eagle3 Throughput Model: A speculative decoding module optimized for high-concurrency inference, improving text generation speed. Licensed for commercial and non-commercial use.

• NVIDIA gpt-oss-120b Eagle Throughput model
  • Hugging Face model

Accidental Leak of NVIDIA’s Upcoming Model: An NVIDIA employee inadvertently uploaded the parent folder of an unreleased model to Hugging Face, sparking community speculation.

• Someone from NVIDIA made a big mistake and uploaded the parent folder of their upcoming model
  • Tweet with details

Kateryna Library Detects LLM Hallucinations: A Python tool (pip install kateryna) classifies model responses as Grounded, Uncertain, or Ungrounded by comparing confidence to RAG evidence.

• Kateryna: Detect when your LLM is confidently bullshitting
  • GitHub
  • Official site

Mathematical & Scientific Breakthroughs

AI-Human Collaboration Solves Erdős Problem #1026: The "Aristotle" AI system contributed a novel generalization to solve the problem, earning recognition from mathematician Terry Tao for its "new understanding."

• Erdos Problem #1026 Solved and Formally Proved via Human-AI Collaboration (Aristotle)
  • Terry Tao’s blog post

New Model Releases

Olmo 3.1 32B Think & Instruct Models: AllenAI expands the Olmo family with two specialized variants: Think (deep reasoning) and Instruct (conversational fluency).

• Olmo 3.1 32B Think & Instruct: New Additions to the Olmo Model Family
  • Hugging Face collection