ReSearch: Teaching LLMs to Make Better Decisions Through Search

Baichuan AI has unveiled an exciting open-source project called ReSearch.

This innovative system teaches Large Language Models to improve their reasoning capabilities by actively searching for information when needed.

How ReSearch Works:

ReSearch combines Reinforcement Learning (RL) with Retrieval-Augmented Generation (RAG) to empower LLMs with a crucial skill: determining when to search for external information.

Similar to how humans look up facts when uncertain, these enhanced models learn to:
Identify knowledge gaps requiring external information
Formulate effective search queries
Execute multi-step, multi-hop searches for complex problems
Integrate search results into their reasoning process.

What makes this approach particularly impressive is that the model learns these search patterns without direct supervision.

Sophgo has introduced 1st RISC-V servers that support #DeepSeek R1 models (1.5B to 70B)

Can do 11.8tps for 70B
SRA3-40 computing server use Sophgo's latest SG2044 64-core server CPU.

It also released SRB3-40 storage server & SRM3-40 convergence server on SG2044.

Huge VLM release from Cohere for AI is just in

Aya-Vision is a new VLM family based on SigLIP and Aya, and it outperforms many larger models.

> 8B and 32B models covering 23 languages and two new benchmark dataset
> supported by HF transformers from get-go

Looking for exceptionally strong engineers.

iOS (native, swift, obj-c)
Backend (GCP, Node.js, NestJS, Nx, Kubernetes)

Location: Europe. Remote is OK.

DM for more info @masterrr

Products I'm hiring for:
https://bereal.com/
https://carrotcare.health/

Sales AI market map

Cohere released Aya Vision on Hugging Face

Aya Vision outperforms the leading open-weight models in multilingual text generation and image understanding. 

In its parameter class, Aya Vision 8B achieves the best performance in combined multilingual multimodal tasks, outperforming Qwen2.5-VL 7B, Gemini Flash 1.5 8B, Llama-3.2 11B Vision, and Pangea 7B by up to 70% win rates on AyaVisionBench and 79% on m-WildVision.

Aya Vision 32B sets a new frontier in multilingual vision open-weights models, outperforming Llama-3.2 90B Vision, Molmo 72B and Qwen2-VL 72B by up to 64% win rates on AyaVisionBench and 72% win rates on m-WildVision.

Commerce Secretary confirms the US Bitcoin Strategic Reserve is likely on the cards:

“A Bitcoin strategic reserve is something the President’s interested in and I think you’re going to see it executed on Friday.”

Trump will unveil the Bitcoin reserve strategy at the White House Crypto Summit. "So Bitcoin is one thing, and then the other currencies, the other crypto tokens, I think, will be treated differently—positively, but differently."

The 2024 Turing Award, the Nobel for Computer Science, goes to the inventors of reinforcement learning.

Andrew Barto and former PhD student Rich Sutton’s (famous for his essay The Bitter Lesson) work is foundational for ChatGPT post-training, AlphaGo, robotics and more.

Full list of confirmed attendees for the White House Crypto Summit this Friday with Trump

Alibaba released QwQ-32B a new reasoning model with 32B parameters that rivals cutting-edge reasoning model, e.g., DeepSeek-R1.

HF
ModelScope
Demo
Chat.

Emirates NBD, a wholly owned bank of the Dubai government, launched the Liv X app on March 6, offering cryptocurrency buying and selling services.

The service is based on the infrastructure of Aquanow and is hosted by Zodia, which is supported by Standard Chartered.

Today Anthropic submitted their recommendations to the OSTP for the U.S. AI Action Plan

Anthropic predicts powerful AI systems will appear by late 2026 or early 2027, with intellectual abilities matching Nobel Prize winners, able to autonomously handle digital tasks (text, audio, video, internet browsing), reason independently over hours or weeks, and control physical equipment digitally

They recommend stronger national security actions, including government testing of AI models for security risks, stricter export controls on key chips like the H20, and secure communication channels between AI labs and intelligence agencies

They suggest the government build 50 gigawatts of additional power capacity dedicated to AI by 2027, speed up AI adoption across federal agencies, and improve economic data collection to prepare for AI’s impact on jobs and society

Ai21 launched Jamba 1.6, the best open model for private enterprise deployment

AI21’s Jamba outperforms Cohere, Mistral and Llama on key benchmarks, including Arena Hard, and rivals leading closed models while maintaining unmatched speed and quality. 

PDF parsing is solved (again). Mistral's new OCR API
— parses 1000-2000 pages for $1
— achieves state of the art results on tables, multilingual
— supports structure: images, bounding boxes, scans, equations

90% of the world's organizational data is in PDFs.

Trump just now signs an Executive Order establishing the Strategic Bitcoin Reserve and U.S. Digital Asset Stockpile

The Reserve will be capitalized with Bitcoin owned by the federal government that was forfeited as part of criminal or civil asset forfeiture proceedings. This means it will not cost taxpayers a dime.

It is estimated that the U.S. government owns about 200,000 bitcoin; however, there has never been a complete audit. The E.O. directs a full accounting of the federal government’s digital asset holdings.

The U.S. will not sell any bitcoin deposited into the Reserve. It will be kept as a store of value. The Reserve is like a digital Fort Knox for the cryptocurrency often called “digital gold.”

Premature sales of bitcoin have already cost U.S. taxpayers over $17 billion in lost value. Now the federal government will have a strategy to maximize the value of its holdings.

The Secretaries of Treasury and Commerce are authorized to develop budget-neutral strategies for acquiring additional bitcoin, provided that those strategies have no incremental costs on American taxpayers.

IN ADDITION, the Executive Order establishes a U.S. Digital Asset Stockpile, consisting of digital assets other than bitcoin forfeited in criminal or civil proceedings.

The government will not acquire additional assets for the Stockpile beyond those obtained through forfeiture proceedings.

The purpose of the Stockpile is responsible stewardship of the government’s digital assets under the Treasury Department.

Google co-founder Larry Page is building a new company called Dynatomics that’s focused on applying AI to product manufacturing

Page is reportedly working with a small group of engineers on AI that can create “highly optimized” designs for objects and then have a factory build them.

Chris Anderson, previously the CTO of Page-backed electric airplane startup Kittyhawk, is running the stealth effort.

MALT: Improving Reasoning with Multi-Agent LLM Training

A new multi-agent post-training method that uses credit assigned synthetic data to improve the reasoning capabilities and self-correction rates of a generator, critic, and refinement model working together.

Paper.

Cortical Labs announced the world's first biocomputer

CL1 merges real living neurons with a chip to solve complex problems and redefine research.

FoundationStereo: A Revolutionary Breakthrough in Stereo Vision

NVIDIA researchers have just introduced FoundationStereo, a groundbreaking foundation model for stereo depth estimation that works right out of the box without any fine-tuning.

Stereo vision enables computer systems to perceive depth and create 3D representations of the world using images from two or more cameras - similar to how humans perceive depth through two eyes.

Code.

This capability is crucial for:
- Robotics and automation
- Autonomous vehicles
- AR/VR applications
- Industrial inspection
- Medical imaging

The Breakthrough
Until now, high-quality depth estimation from low-cost cameras has been a persistent challenge in industrial applications, often requiring investments of $10,000+ in specialized structured-light scanners.

FoundationStereo changes the game by:
Zero-shot generalization: The model works across diverse scenarios without domain-specific fine-tuning
Handling challenging cases: Exceptional performance on transparent objects, reflective surfaces, and complex lighting conditions
Leveraging monocular priors: Adapting rich geometric priors from vision foundation models
Massive training dataset: Built on a 1-million stereo pair synthetic dataset with unprecedented diversity and photorealism
Technical Innovation
The researchers developed several novel components:
Side-Tuning Adapter that leverages internet-scale knowledge from monocular depth estimation models
Attentive Hybrid Cost Filtering with 3D Axial-Planar Convolution
Disparity Transformer for long-range context reasoning