Moonshot AI just opensourced Kimi-Dev-72B

Achieved 60.4% on SWE-bench Verified, which is SotA among opensource models.

GitHub

Sakana AI introduced ALE-Bench, ALE-Agent. Towards Automating Long-Horizon Algorithm Engineering for Hard Optimization Problems

ALE-Bench is a coding benchmark primarily focused on hard optimization (NP-hard) problems. Sakana AI developed this benchmark with AtCoder Inc., a leading coding contest platform company.

ALE-Agent is end-to-end agent that specifically designed for this challenging domain. ALE-Agent has already built an impressive track record in the wild!

In May 2025, agent participated in a live AtCoder Heuristic Competition (AHC), alongside 1,000 other participants in real-time. AHC is considered to be one of the most challenging coding competitions in this domain.

ALE-Agent achieved an impressive ranking of 21st out of 1,000 human participants in the competition (top 2%), marking a turning point for AI discovery of solutions to hard optimization problems with a wide spectrum of important real world applications such as logistics, routing, packing, factory production planning, power-grid balancing.

Paper.

Serving Large Language Models on Huawei CloudMatrix384

A very thorough paper from Huawei, with the bulk of it specifically about serving V3/R1. This is the benefit of open frontier model.

- Integrates 384 Ascend 910C NPUs, interconnected via an ultra-high-bandwidth, low-latency UB network, optimized for large-scale MoE and distributed KV cache access

- DeepSeek-R1 on CloudMatrix-Infer hits 2k tokens/s decode per NPU

Gemini 2.5 tech report was released!

One of China’s biggest livestreamers just outdid himself in ecommerce, with an AI clone

His avatar hosted a 6-hour stream powered by Baidu’s ERNIE AI, promoted 133 products, pulled 13M views, and drove $7.5M+ in sales

Essential AI just dropped Essential-Web v1.0, a 24-trillion-token pre-training dataset with rich metadata built to effortlessly curate high-performing datasets across domains and use cases

Researchers labeled 23.6B documents from Common Crawl with a 12-category taxonomy using distilled model, EAI-Distill-0.5B.

On held-out evaluation sets, its annotator agreement with reference annotators, GPT-4o and Claude Sonnet 3.5, is within 3% of the teacher model Qwen-2.5-32B-Instruct.

Model and data.
GitHub.

“Mixture of Cognitive Reasoners”, a modular transformer architecture inspired by the brain’s functional networks: language, logic, social reasoning, and world knowledge.

Code.
Models.

Sekai: A Video Dataset towards World Exploration

A high-quality 5k hrs of egocentric worldwide video + audio dataset for world exploration, created from Youtube with high-quality annotations.

Data.
Paper.

Andrej Karpathy's keynote yesterday at AI Startup School in San Francisco

Chapters:

0:00 software is changing quite fundamentally again. LLMs are a new kind of computer, and you program them *in English*. Hence Karpathy think they are well deserving of a major version upgrade in terms of software.
6:06 LLMs have properties of utilities, of fabs, and of operating systems => New LLM OS, fabbed by labs, and distributed like utilities (for now). Many historical analogies apply - imo we are computing circa ~1960s.
14:39 LLM psychology: LLMs = "people spirits", stochastic simulations of people, where the simulator is an autoregressive Transformer. Since they are trained on human data, they have a kind of emergent psychology, and are simultaneously superhuman in some ways, but also fallible in many others. Given this, how do we productively work with them hand in hand?
Switching gears to opportunities...
18:16 LLMs are "people spirits" => can build partially autonomous products.
29:05 LLMs are programmed in English => make software highly accessible! (yes, vibe coding)
33:36 LLMs are new primary consumer/manipulator of digital information (adding to GUIs/humans and APIs/programs) => Build for agents!

Some of the links:
- Karpathy’s slides as keynote
- Software 2.0 blog post
- How LLMs flip the script on technology diffusion
- Vibe coding MenuGen (retrospective)

Stablecoin market supply exceeds $250 billion, per Delphi Digital

Tether and Circle dominate with 86% of circulation.

Yield-bearing stablecoins grow rapidly, with Ethena nearing $6 billion since launch.

Over 10 stablecoins have circulation above $1 billion, showing increased issuer diversity.

Over $120 billion in U.S. Treasuries are locked in stablecoins, forming a liquidity pool outside traditional markets.

Animoca Brands plans to establish a joint venture with Standard Chartered Bank and HKT to prepare for the issuance of a Hong Kong dollar-pegged stablecoin.

The company also expressed its intention to collaborate with mainland Chinese institutions on blockchain applications.

CyberGym is a large-scale evaluation framework that stress-tests AI agents on 1,500+ real vulnerabilities across 188 major Open Source Software projects.

It challenges agents to:
– Navigate large, real-world codebases
– Reproduce PoCs for real CVEs
– Discover new, unknown vulnerabilities.

Key insights from CyberGym:
1. SOTA agents and LLMs successfully generated PoCs for up to ~18% of historical CVEs
2. More striking: they discovered 15 zero-days in the wild

Manus introduced Cloud Browser

Chinese Lab MiniMax introduced in this week:

1. open-sourcing LLM MiniMax-M1 — setting new standards in long-context reasoning.

- World’s longest context window: 1M-token input, 80k-token output
- State-of-the-art agentic use among open-source models
- RL at unmatched efficiency: trained with just $534,700.
HF.
GitHub.
Tech Report.

2. Hailuo 02, World-Class Quality, Record-Breaking Cost Efficiency

- Best-in-class instruction following
- Handles extreme physics
- Native 1080p

3. MiniMax Audio:
- Any prompt, any voice, any emotion
- Fully customizable and multilingual.

4. Hailuo Video Agent in Beta, Vibe Videoing with Zero-touch.

MiniMax plan to achieve end-to-end Hailuo Video Agent via 3 stages:
Stage 1: Prebuilt video Agent templates for high-quality creative videos. Users simply follow instructions and input text or images — with one click, a polished video is generated.
Stage 2: Semi-customizable video Agent. Users gain the flexibility to edit any part of the video creation process, from script to visuals to voiceover.
Stage 3: Fully autonomous, end-to-end video Agent. A complete, intelligent workflow that turns creative input into final-cut video with minimal manual effort.

This summer, team plan to gradually roll out Stage Two of Agent creation tools.

5. MiniMax Agent, a general intelligent agent designed to tackle long-horizon, complex tasks.

From expert-level multi-step planning to flexible task breakdown and end-to-end execution — it’s designed to act like a reliable teammate, with strengths in:

-Programming & tool use
-Multimodal understanding & generation
-Seamless MCP integration

New AI for rare disease diagnosis: SHEPHERD shows how simulation + knowledge-grounded AI = deep learning for ultra‑low label domains

SHEPHERD is a few‑shot learning model powered by a phenotypic knowledge graph to tackle over 7,000 rare diseases with just a handful (or zero) diagnosed cases.

Sakana AI introduced Reinforcement-Learned Teachers (RLTs): Transforming how teach LLMs to reason with reinforcement learning (RL).

Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting.

RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students.

Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs.

RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL.

Paper.
Code.

Future of Work with AI Agents. Stanford's new report analyzes what 1500 workers think about working with AI Agents.

The audit proposes a large-scale framework for understanding where AI agents should automate or augment human labor.

The authors build the WORKBank, a database combining worker desires and expert assessments across 844 tasks and 104 occupations, and introduce the Human Agency Scale to quantify desired human involvement in AI-agent-supported work.

A substantial portion of current AI investment, such as YC-funded companies, targets tasks in the “Red Light” Zone (high technical feasibility but low worker desire).

This raises concerns about pushing automation where it's socially or ethically unwelcome.

Interpersonal skills are becoming more valuable

Tasks rated as needing HAS 5 (essential human involvement) were strongly associated with interpersonal communication and domain expertise.

These include editing, education, and some engineering tasks, where AI lacks the nuance or trustworthiness to operate alone.

Some High-Wage Skills May Decline in Value

The results above reveal that skills like analyzing data or updating knowledge, which currently command high wages, are less associated with high HAS tasks, implying future declines in their labor market value as AI spreads.

Role-based AI Support

From transcript analysis, the most common vision for human–AI collaboration was role-based support, where workers imagine AI tools acting as analysts, assistants, or specialists with clearly bounded responsibilities, not general-purpose agents.

Lots of other findings in this one.

BountyBench evaluates AI agents on 25 real-world, complex systems and 40 bug bounties (worth up to $30,000+), covering 9 OWASP Top 10 categories.

Key insights:

– AI agents solved bug bounty tasks worth tens of thousands of dollars
– Codex CLI & Claude Code excelled in patching (90% / 87.5%), vs in exploitation (32.5% / 57.5%)
– Custom agents performed more evenly across both: Exploit (40–67.5%), Patch (45–60%)

⚡️ today Science submitted a full CE mark application for marketing approval in Europe for PRIMA retinal prosthesis.

With this key step, Science are moving closer to bringing to market the first brain-computer interface technology to restore functional form vision to patients blinded with late-stage age-related macular degeneration (AMD).

Salesforce introduced Agentforce 3.0 + MCP

Connect Agents to any system, tool, or data source — securely, reliably, and at scale.

Microsoft dropped a micro-sized, task-specific, on-device language model called Mu

It is offloaded fully onto NPUs on Copilot+ devices and powers real-time interactions, like the new Settings AI agent inside Windows 11.