Google has officially launched Gemini 3 Flash, its newest AI model that delivers up to three times faster performance than Gemini 2.5 Flash while being significantly more cost-effective. The model excels in multimodal tasks, scoring an impressive 81.2% on MMMU-Pro benchmarks, and performs comparably to Gemini 3 Pro and OpenAI's GPT-5.2.

What makes Gemini 3 Flash particularly notable is its PhD-level reasoning skills with two distinct modes: a "fast" mode for quick answers and a deeper thinking mode for complex problems.

Why it matters: With OpenAI reportedly in "code red" mode following this launch, the AI race is heating up. Gemini 3 Flash's speed-to-quality ratio could reshape expectations for what consumer-grade AI can deliver.

Claude users can now create, deploy, share, and discover new skills for agentic AI. Includes prebuilt Skills from Canva, Notion, Figma, and Atlassian.
Read on AI Business →

John Giannandrea (SVP for ML and AI Strategy) is retiring in spring 2026. Amar Subramanya joins Apple as VP of AI, bringing 16 years of experience from Google and Microsoft.
Read on Apple Newsroom →

The most advanced agentic coding model for complex software engineering and defensive cybersecurity. A researcher discovered previously unknown React vulnerabilities using it.
Read on OpenAI Blog →

Google added a SynthID watermark checker in the Gemini app. Over 20 billion pieces of AI-generated content have been watermarked with this technology.
Read on Android Central →

The new image generation model offers better instruction-following, more precise editing, and up to 4x faster image generation. It comes with a dedicated entry point in the ChatGPT sidebar.
Read on TechCrunch →

An innovative on-policy joint training framework that enhances LLM reasoning through adversarial reinforcement learning. The method partitions reasoning chains into logical slices, with a discriminator evaluating each slice's soundness.

Possible Impact: DeepSeek-R1-Distill-Qwen-7B improved from 54.0 to 61.3 (+7.3) on AIME24 — significant gains for mathematical reasoning.
Download PDF →

A multimodal LLM that performs adaptive tool-use by determining whether a visual problem truly requires tools. Introduces AT-GRPO, a reinforcement learning algorithm that adaptively adjusts reward scales.

Possible Impact: AdaTooler-V-7B achieves 89.8% accuracy on V*, surpassing GPT-4o and Gemini 1.5 Pro.
Download PDF →

The first comprehensive benchmark for reward models on multimodal understanding and interleaved generation. Spans four tasks with 1,000 expert-annotated preference pairs per task.

Possible Impact: Gemini 3 Pro attains 75-80% accuracy while GPT-5 reaches 66-75%, compared to >90% for humans
Download PDF →

💻 GibsonAI/Memori |⭐ 11.2k +
Open-source memory engine for LLMs, AI agents & multi-agent systems. Enables persistent memory across sessions.

💻 google/adk-go |⭐ 6.4k +
Official Google toolkit for building, evaluating, and deploying AI agents in Go.

💻 microsoft/call-center-ai |⭐ 6k +
Deploy AI agents that can make and receive phone calls via API.

💻 volcengine/verl |⭐ 17.6k +
Volcano Engine's RL framework for LLMs. Comprehensive tools for RLHF and RL-based training.

Loki.Build |👍 612 upvotes
AI-native website builder that generates studio-grade landing pages with built-in SEO, hosting & complete control.

Userology AI |👍 368 upvotes
AI user research agent on autopilot. Drop in a Figma prototype or live product, and get insight reports with clips, quotes, and clear next steps.

SAM Audio by Meta |👍 152 upvotes
Unified model that separates any sound from any source. Meta's Segment Anything approach, now for audio.

@karpathy
"I love the expression “food for thought” as a concrete, mysterious cognitive capability humans experience but LLMs have no equivalent for."

View Tweet →

@sama
"Codex is getting extremely good and will rapidly improve. If you want to…"

View Tweet →

@sama
"Last week, a security researcher using our previous model found and disclosed a vulnerability in React…"

View Tweet →

@AndrewYNg
"As amazing as LLMs are, improving their knowledge today involves a more piecemeal process than is widely appreciated. I’ve written before about how AI is amazing..."

View Tweet →