Howdy, it’s Barsee.

Happy Wednesday, AI family, and welcome to the last AI Valley edition of 2025.

Today’s climb through the Valley reveals:

• Robotic "NRE-skin" mimics the human nervous system

• SoftBank invests $40 billion in OpenAI

• Meta introduces Self-play SWE-RL

• Stanford-Harvard researchers explain why agentic AI fails

• Plus trending AI tools, posts, and resources

Let’s dive into the Valley of AI…

MINIMAX

MiniMax M2.1 SOTA Multi-language Agentic Coding

MiniMax M2.1 is an OSS coding & agent model for real work. SOTA multi-language coding (Rust, Java, Go, C++, Kotlin, Obj-C, TS/JS), stronger Android/iOS & web dev, faster concise reasoning, lower tokens, and advanced interleaved instruction following.

Learn More about M2.1

New Year Coding Plan Offer

_{*This is sponsored}

THROUGH THE VALLEY

1/ Robotic "NRE-skin" mimics the human nervous system

Researchers have built a new electronic skin called NRE-skin that helps robots sense touch more like humans. Instead of sending constant data, it uses quick electrical spikes similar to real neurons, firing faster when pressure increases. It also has built-in reflexes that let a robot pull away instantly when pressure crosses a pain threshold, without waiting for the main computer. The skin is modular, with magnetic pieces that snap on and auto-detect their position for easy repairs. Powered by neuromorphic chips, it uses very little energy and could make robots safer and more efficient when working around people.

2/ SoftBank invests $40 billion in OpenAI

SoftBank has now finished its massive investment in OpenAI, sending a final payment of about 22 to 22.5 billion dollars and bringing its total commitment to roughly 41 billion dollars. This gives SoftBank an estimated 11 percent stake in the company. The money follows earlier investments and will help fund OpenAI’s long term infrastructure plans, including partnerships with Nvidia, AMD, Broadcom, Oracle, and the Stargate compute project. SoftBank recently sold its 5.8 billion dollar Nvidia stake to free up cash and also bought DigitalBridge for 4 billion dollars to expand its AI push. OpenAI is also in talks with Amazon for another 10 billion dollar investment and recently secured 1 billion dollars from Disney.

3/ Meta introduces Self-play SWE-RL, where one model improves coding by generating and fixing its own bugs

Meta introduced Self-play SWE RL, a system where a coding model improves by breaking and fixing software on its own instead of relying on human-written issues or labels. The model works in two roles. First, it injects bugs into real codebases, then it switches sides and repairs them using tests as the only guide. Because tests give clear pass or fail signals, the reinforcement learning loop stays stable and scales to real repositories with full dependencies. The approach raised scores by 10.4 points on SWE Bench Verified and 7.8 on SWE Bench Pro, outperforming RL methods trained on human issue data.

4/ Stanford-Harvard researchers explain agentic AI fails without real-time adaptation during tasks

A new Stanford-Harvard paper explains why many agentic AI systems look impressive in demos but fall apart in real use. The researchers found the core issue is not bad reasoning but the lack of real-time adaptation. Most agents set a plan, then follow it even when tools fail, outputs change, or assumptions no longer hold. The paper argues that agents need to adjust during execution based on feedback from tool results or correctness checks. It also shows that adaptive agents perform better on long or uncertain tasks, external memory beats long reasoning chains, and rigid tool selection is a major source of failure.

TRENDING TOOLS

• AbleMouse AI edition > An open-source tool for computer vision-powered mouse control

• Manus > SOTA general AI agent, newly acquired by Meta

• Wikiwand > AI-driven contextual learning with Wikipedia

• Bloom > Distil your website into a living brand system and create on-brand assets

• Shortcut > Superhuman Excel Agent that takes you from beginner to pro

THINK PIECES / BRAIN BOOST

• Thinking like an AI

• The moment we stopped understanding AI

• ChatGPT: 30-year history and how AI learned to talk

• The next wave of AI: Physical AI by Nvidia

• Why AI writing is mid

• AI-generated content is changing our language and communication style

• Google has released its Year in Search trends for 2025

THE VALLEY GEMS

What’s trending on social today:

THAT’S ALL FOR TODAY

Thank you for reading today’s edition. That’s all for today’s issue.

💡 Help me get better and suggest new ideas at [email protected] or @heyBarsee

👍️ New reader? Subscribe here

Thanks for being here.

REACH 100K+ READERS

Acquire new customers and drive revenue by partnering with us

Sponsor AI Valley and reach over 100,000+ entrepreneurs, founders, software engineers, investors, etc.

If you’re interested in sponsoring us, email [email protected] with the subject “AI Valley Ads”.