OpenAI just released HealthBench — a new eval for AI systems for health.

Developed with 262 physicians who have practiced in 60 countries.

Researchers trained an end-to-end visual navigation policy on 2000 hours of uncurated, crowd-sourced data and evaluated it across 24 environments in 6 countries

Recipe: Train a model-based policy on a small amount of clean short-horizon data --> use it to re-label actions for the uncurated dataset --> train your favorite e2e BC model on these labels!

Open-sourced code & data.

MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering

A Gym-style framework for systematically training, evaluating, and improving agents in iterative ML engineering workflows.

Paper.
GitHub.

Cohere presents Aya Vision: Advancing the Frontier of Multilingual Multimodality

- Aya-Vision-8B outperforms Qwen-2.5-VL-7B
- Aya-Vision-32B outperforms Qwen-2.5-VL-72B

Paper.
Models.
Bench.

Notion released AI for Work, a suite of work-centered AI features, including:

— AI Meeting Notes
— Enterprise Search to find answers across tools
— Research Mode to draft docs
— Access to models, including GPT-4.1 & Claude 3.7

Google introduced AlphaEvolve an AI coding agent

It’s able to:

1. Design faster matrix multiplication algorithms
2. Find new solutions to open math problems
3. Make data centers, chip design and AI training more efficient across Google.

AlphaEvolve uses:
- LLMs: To synthesize information about problems as well as previous attempts to solve them - and to propose new versions of algorithms
- Automated evaluation: To address the broad class of problems where progress can be clearly and systematically measured.
- Evolution: Iteratively improving the best algorithms found, and re-combining ideas from different solutions to find even better ones.

Google applied AlphaEvolve to a fundamental problem in computer science: discovering algorithms for matrix multiplication. It managed to identify multiple new algorithms.

This significantly advances our previous model AlphaTensor, which AlphaEvolve outperforms using its better and more generalist approach.

Meta just released new models, benchmarks, and datasets that will transform the way researchers approach molecular property prediction, language processing, and neuroscience.

1. Open Molecules 2025 (OMol25): A dataset for molecular discovery with simulations of large atomic systems.

2. Universal Model for Atoms: A machine learning interatomic potential for modeling atom interactions across a wide range of materials and molecules.

3. Adjoint Sampling: A scalable algorithm for training generative models based on scalar rewards.

4. FAIR and the Rothschild Foundation Hospital partnered on a large-scale study that reveals striking parallels between language development in humans and LLMs.

#DeepSeek presents:Insights into DeepSeek-V3: Scaling Challenges and Reflections on Hardware for AI Architectures

Elaborates on hardware architecture and model design in achieving cost-efficient large-scale training and inference.

Google introduced a notion of sufficient context to examine retrieval augmented generation (RAG) systems, developing a method to classify instances, analyzing failures of RAG systems & proposing a way to reduce hallucinations.

Today "a milestone in the evolution of personalized therapies for rare & ultra-rare inborn errors of metabolism"
—the 1st human to undergo custom genome editing
—outgrowth of decades of NIH funded research.

Paper.

Agents from scratch

This repo covers the basics of building agents:
+ Fundamentals
+ Build an agent
+ Agent eval
+ Agent w/ human-in-the-loop
+ Agent w/ long-term memory
Code (all open source).

Building agents -Combing workflow (router) with agent that can call email tools.
Notebook.
Slides.

Agent evals -Unit tests (Pytest) for triage decision + tools calls (test structured outputs using heuristic eval) and LLM-as-judge to eval email responses
Notebook.
Slides.

Human-in-the-loop -Add human in the loop for approval / editing of specific tool calls.
Notebook.

Memory - Add memory, so the agent learned email response preferences from human feedback
Notebook.

Agent can be hooked into Gmail by swapping out the tools used. Components are also general and can be used w/ various tools / MCP servers.

Qwen introduced Parallel Scaling Law for Language Models

"We introduce the third and more inference-efficient scaling paradigm: increasing the model’s parallel computation during both training and inference time."

"We draw inspiration from classifier-free guidance (CFG)".

"In this paper, we hypothesize that the effectiveness of CFG lies in its double computation."

"We propose a proof-of-concept scaling approach called parallel scaling (PARSCALE) to validate this hypothesis on language models. "

"parallelizing into P streams equates to scaling the model parameters by O(log P)".

"for a 1.6B model, when scaling to P = 8 using PARSCALE, it uses 22× less memory increase and 6× less latency increase compared to parameter scaling that achieves the same model capacity".

GitHub

OpenAI introduced AI agent codex.

it is a software engineering agent that runs in the cloud and does tasks for you, like writing a new feature of fixing a bug.

U can run many tasks in parallel. Starting to roll out today to ChatGPT pro, enterprise, and team users.

Arc is hiring a Chief Scientific Officer!

We think is the most dynamic and ambitious biology research organization in the world.

They want to build the world's first virtual cell + use it to develop cures to complex diseases like Alzheimer's.

IBM has released report the Agentic AI in Financial Services

highlights

1. AI agents enable a paradigm shift from responsive to adaptive technology services, creating more accessible, personalised #banking services & experiences for customers.

2. Organisations must adopt a measured, phased approach balancing #innovation with comprehensive risk management, through thorough assessments, clear governance structures, talent development, & continuous monitoring.

3. Successfully managing agentic AI requires seamless cooperation across business functions, supported by transparent governance structures and communication channels.

4. Understanding where new risks emerge and implementing appropriate controls is essential, recognising agentic AI as a fundamentally different technological paradigm requiring new approaches to #governance & controls.

5. Early integration of #compliance considerations validates use cases before significant investment while ensuring alignment with organisational risk appetite.

6. Organisations must develop holistic AI literacy programs—including philosophy, linguistics, law, and anthropology—to formulate and responsible #AIstrategies and address potential model biases.

The UK tax authority has announced that, starting from January 1, 2026, crypto asset companies operating in the UK must comprehensively report user and transaction data, including user identity, address, tax identification number, and details of each transaction, in compliance with the global Crypto Asset Reporting Framework (CARF) to combat tax evasion and enhance transparency.

Violators will face a maximum fine of £300 per user.

Adobe introduced HUMOTO, 4D dataset for human-object interaction, developed with a combination of wearable motion capture, SOTA 6D pose estimation vision models, LLM, and the professional refining works of multiple animation studios.

HUMOTO features:
1. Over 700 diverse daily activities
2. Interactions with 60+ objects, 70+ articulated parts.
3. Fine-grained text annotations
4. Detailed hand and finger movements.

Microsoft introduced Magentic-UI — an experimental human-centered web agent

It automates your web tasks while keeping you in control —through co-planning, co-tasking, action guards, and plan learning.

Fully open-source.

Google is launching their own coding agent, Jules, at I/O today

It lets you make changes to your GitHub repos with English prompts in a VM using Gemini 2.5 Pro. It's Google’s version of Devin.

Here's the leaked official repo of prompts it can do.

ByteDance presents the Pre-trained Model Averaging strategy, a novel framework for model merging during LLM pre-training.

Found that merging checkpoints trained with constant learning rates not only achieves significant performance improvements but also enables accurate prediction of annealing behavior.

Code might be posted there

Claude 4 is here ! Anthropic Prepares to Launch Claude 4

The lineup will reportedly include 2 versions:

1. Claude Sonnet 4 - a faster version for everyday tasks.

2. Claude Opus 4 - a powerful model for complex problems and creative work.

Based on leaked information, these models are currently in a closed testing environment marked "not intended for production use" and subject to strict rate limitations.

Several intriguing features have been discovered in the configuration files:

"show_raw_thinking" / "show_raw_thinking_mechanism" - functionality that potentially allows users to observe the AI's thought formation process.

The models appear to operate in an environment based on a popular web framework.

A sophisticated "feature gates" system is being employed to precisely control which capabilities are available to different users.

Configuration includes specific parameters for various interaction methods and model performance metrics.

The technical JSON contains numerous rule sets and session parameters suggesting enhanced personalization capabilities.