Full and Long CoT boost reasoning by expanding intermediate steps—but at a high token cost. Not ideal for latency- or cost-sensitive apps. This paper introduces Fractured Sampling, a practical inference-time technique that turbocharges reasoning without retraining, using far
RESEARCH
-

Qwen3-32B Performance Benchmark Against Leading AI Models
By
–
Qwen3-32B vs. leading models eval https://
artificialanalysis.ai/?models=gpt-4-
1o3llama-4-scoutllama-4-maverickgemma-3-27bgemini-2-5-proclaude-3-7-sonnet-thinkingclaude-3-7-sonnetmistral-medium-3deepseek-r1deepseek-v3-0324grok-3grok-3-mini-reasoningqwen3-32b-instruct-reasoning
… -

Cerebras Qwen3 achieves 99% latency reduction versus o3
By
–
Artificial Analysis measured the time to first token of every reasoning model from o3 to R1. DeepSeek R1 = 103 sec
Qwen3-32B on Cerebras = 1.1 sec We give you R1 level intelligence with 99% latency reduction. Try it: https://
inference.cerebras.ai -
Google Introduces Gemini Diffusion: Faster Non-Sequential Text Generation
By
–
Today we introduced Gemini Diffusion⚡️ (& DeepThink, Veo3, Imagen4, 2.5 updates…).
— Oriol Vinyals (@OriolVinyalsML) 20 mai 2025
It's been a dream of mine to remove the need for "left to right" text generation. It's so fast, that we had to *slow down* the video during the presentation.https://t.co/YW73ZPqJ5a pic.twitter.com/N3BXZs9ELeToday we introduced Gemini Diffusion (& DeepThink, Veo3, Imagen4, 2.5 updates…). It's been a dream of mine to remove the need for "left to right" text generation. It's so fast, that we had to *slow down* the video during the presentation. https://
deepmind.google/models/gemini-
diffusion/
… -
Google Gemini 2.5 Scores 49% on USA Math Olympiad
By
–
A few months ago, the best LLM scored 5% on the USA Math Olympiad test. Models have been rapidly improving. Today, Google Gemini 2.5 scored 49%, which is better than 75% of the people who took the test (roughly the top 250 students in the USA).
-

Pentagon AI Leader Winchell Advances Defense Autonomy Portfolio
By
–
“Winchell most recently led the artificial intelligence and autonomy portfolio for the Defense Department’s Strategic Capabilities Office. Prior to that, he was chief engineer for the Pentagon’s Algorithmic Warfare Cross-Functional Team, commonly known as Project Maven.”
-
50 Trillion Tokens: LLM Growth and Development Speed Call
By
–
0 to 50,000,000,000,000 tokens in just 3 years. 50T Tokens. That kind of scale isn’t just extraordinary, 𝗶𝘁’𝘀 𝗮 𝗰𝗮𝗹𝗹 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱.
— Cerebras (@cerebras) 20 mai 2025
LLM growth is exploding, and speed is the catalyst. pic.twitter.com/XEDog6HOfu0 to 50,000,000,000,000 tokens in just 3 years. 50T Tokens. That kind of scale isn’t just extraordinary, 𝗶𝘁’𝘀 𝗮 𝗰𝗮𝗹𝗹 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱. LLM growth is exploding, and speed is the catalyst.
-
AI Assessment Framework Prioritizes Performance and Safety
By
–
HAI Senior Fellow @aiprof_mykel and his team at @SISLaboratory aim to transform how AI systems are assessed, promoting a more systematic approach to validation that prioritizes both performance and safety.
-
What URLs Should Exist Future World States Algorithm
By
–
There is slightly more to my "What do write about?" algorithm than "What URL will I really want to exist given some future world states" but that is a part of it. Happy to be of service.
-
AI Mode Search: Query Fan Out Technique with Gemini
By
–
AI Mode is our most powerful AI Search but… how does it work? 🔍
— Google AI (@GoogleAI) 20 mai 2025
Using a "query fan out" technique, Search calls on Gemini to break a query into subtopics. Then it issues multiple, concurrent searches and taps into real-time data to generate the most helpful response. pic.twitter.com/KzN2bvCx6UAI Mode is our most powerful AI Search but… how does it work? Using a "query fan out" technique, Search calls on Gemini to break a query into subtopics. Then it issues multiple, concurrent searches and taps into real-time data to generate the most helpful response.
