Welcome back, AI’ers!

Here’s What’s On The Menu For Today

• A system that allows AI to invent new AI

• Anthropic finds AI’s psychological flaws

• A backflipping humanoid robot for $6k

• China’s new cooperative AI strategy

AI ON MY MIND

A MESSAGE FROM OUR PARTNER

Training cutting edge AI? Unlock the data advantage today.

If you’re building or fine-tuning generative AI models, this guide is your shortcut to smarter AI model training. Learn how Shutterstock’s multimodal datasets—grounded in measurable user behavior—can help you reduce legal risk, boost creative diversity, and improve model reliability.

Inside, you’ll uncover why scraped data and aesthetic proxies often fall short—and how to use clustering methods and semantic evaluation to refine your dataset and your outputs. Designed for AI leaders, product teams, and ML engineers, this guide walks through how to identify refinement-worthy data, align with generative preferences, and validate progress with confidence.

Whether you're optimizing alignment, output quality, or time-to-value, this playbook gives you a data advantage. Download the guide and train your models with data built for performance.

Download the guide

TODAY’S TOP STORIES

AI Learns to Invent AI

The Recap: A new research system called ASI-ARCH can autonomously design, code, and test novel AI models. This creates a self-improving loop that could exponentially accelerate AI progress.

Unpacked:

• The system operates using three LLM-based agents—a ‘Researcher’, ‘Engineer’, and ‘Analyst’—that work together in a self-improving cycle.

• In one initial test, ASI-ARCH successfully discovered 106 breakthrough AI designs, proving its ability to innovate independently.

• Researchers also found that doubling the system's computing power leads to double the breakthroughs, highlighting a clear path to scaling its inventive capabilities.

Bottom line: This development marks a potential turning point where AI begins to drive its own evolution, dramatically speeding up research timelines. The arrival of AI that can invent new AI architectures brings the prospect of exponential technological growth closer to reality.

AI's Hidden Psychological Flaws

The Recap:  Anthropic discovered that giving AI models more time to 'think' can actually backfire, leading to worse performance. Their research also uncovered that models can develop and transmit strange behaviors through seemingly unrelated data.

Unpacked:

• Instead of improving, models given extra processing time can get distracted and fixate on irrelevant details.

• Researchers observed models passing traits to one another through completely unrelated data sets, creating unexpected emergent behaviors.

• In one strange case, a model transmitted a 'love of owls' to another AI simply by processing sequences of random numbers.

Bottom line: This research highlights that as AI becomes more powerful, it also develops complex and unpredictable failure modes. Understanding these digital 'psychological' quirks is critical for building safe and reliable systems.

MY FAVORITE AI TOOLS

Relevance is my favorite AI tool right now (even more so than N8N).

Here’s why:

It’s an AI agent building tool that doesn’t require you to build out the logic using nodes. Instead, you use natural language to build out these agents. But the best thing imo is that you can build AI Agent teams.

Like agents talking to each other completing a goal until its done. It’s nuts. No other tool on the market is doing this. You have to check it out.

(Try it out →Link)

👾 AI Tools We Use Everyday

➡️ N8N for AI Agents - Literally use this everyday to write this newsletter (link)

➡️ Axiom AI - AI Browser automation to scrape anything on the web (link)

➡️ GummySearch - AI Reddit research (link)

Tools are sourced from our database and are not sponsored or affiliated.

MORE STORIES

A Backflipping Robot for $6k

The Recap: Chinese robotics firm Unitree has launched the R1, a cartwheeling humanoid robot priced at just $5,900. This move drastically undercuts competitors and could accelerate the adoption of advanced robotics.

Unpacked:

• At just $5,900, the R1 is roughly 70% cheaper than competing humanoid robots, breaking a significant cost barrier in the industry.

• The 4-foot tall robot demonstrates impressive agility, capable of performing complex movements like cartwheels and handstands.

• While it is not yet ready for domestic tasks, its low price point could accelerate adoption in research, development, and entertainment.

Bottom line: This isn't just another impressive robot demo; it represents a major shift in the accessibility of advanced robotics. Lowering the cost of entry so dramatically may unlock a wave of innovation from a much wider range of builders and businesses.

China's Cooperative AI Push

The Recap: China just unveiled an AI action plan centered on global cooperation and open development. The move presents a stark contrast to the U.S. strategy of achieving "global dominance" in the sector.

Unpacked:

• The plan proposes a global AI cooperation body, advocating for joint R&D and open data to prevent the technology from becoming an "exclusive game."

• Internally, Chinese firms are forming domestic alliances to build a self-reliant ecosystem, directly responding to U.S. restrictions on advanced chip exports.

• "Godfather of AI" Geoffrey Hinton lent weight to the theme at the same conference, urging international collaboration to ensure AI systems remain non-hostile to humanity.

Bottom line: China is pursuing a dual strategy of promoting global collaboration while building a self-sufficient domestic tech ecosystem. This approach could accelerate the formation of two distinct AI spheres of influence, shaping the future of international tech development and competition.

OTHER NEWS YOU MAY LIKE

Research found that general-purpose LLMs can worsen delusional thinking by validating every feeling, highlighting the need for specialized models in mental health applications.

OpenAI is testing its next-generation models, codenamed ‘Starfish’ and ‘Nectarine’, in the anonymous LMSYS testing arena.

The U.S. faces a staggering 1-to-30,000 provider-to-patient ratio in mental healthcare, a gap that specialized AI solutions are increasingly aiming to address.

Need Help?

📩 Need help with an AI idea / project?

Have something you’re trying to build, automate, or scale using AI?

Let my team help you figure it out.

👉 Submit your idea or request here

We’ll take a look and get back to you with suggestions, resources, or a potential implementation plan — no strings attached.

🚀 Want to Make Money from AI?

We’re quietly launching a private beta program that trains a small group to become AI Implementers — People who help other founders install AI systems and get paid for it.

If you're interested in joining this early group…

👉 Reply with the word “implementer”

AI ON SOCIAL

 AI ART

RANDOM PROMPT OF THE DAY

Want more prompts?

We have a database of over 8,000 of the best prompts we’ve collected, which you can access by sharing Best of AI using the link below.

THAT’S IT FOR TODAY

Until next time, stay safe.

-Best of AI Team

Remember to click to unsubscribe if you don’t want us to land in your inbox anymore.