MIT's AI learns to upgrade itself

Good morning, AI enthusiasts. The path to superintelligence just got a bit clearer, with MIT researchers teaching AI to self-improve without human help.

As models eventually learn to rewrite their own code, training data, and capabilities at scale, is this the first baby step towards the exponential intelligence explosion that experts both crave and fear?

In today’s AI rundown:

• MIT researchers teach AI to self-improve

• AI develops human-like object understanding

• Decode video success patterns for better content

• UK study reveals AI's hidden impact on children

• 4 new AI tools & 4 job opportunities

LATEST DEVELOPMENTS

AI RESEARCH

📶 MIT researchers teach AI to self-improve

The Rundown: MIT researchers just developed Self-Adapting LLMs (SEAL), a framework that enables large language models to teach and improve on their own by creating their training data and instructions for self-updates.

The details:

• SEAL allows models to generate their own "self-edits" — instructions for creating synthetic data and setting parameters to update their own weights.

• It learns through trial-and-error via a reinforcement learning loop, rewarding the model for generating self-edits that lead to better performance.

• In knowledge tasks, the AI learned more effectively from its own notes than from learning materials generated by the much larger GPT-4.1.

• The system also dramatically improved at puzzle-solving tasks, jumping from 0% with standard methods to 72.5% after learning how to train itself effectively.

Why it matters: Self-improving AI is frequently mentioned as a potential lead-in to the leap toward superintelligence. While SEAL (and other research frameworks like Sakana’s DGM) aren’t there yet, they point to a scary but exciting future where models can continue upgrading (exponentially) on their own, going beyond human design.

TOGETHER WITH BASETEN

🚀 Enterprise grade infrastructure for everyone

The Rundown: Baseten’s multi-cloud capacity management (MCM) system brings bleeding edge infrastructure capability to every workload. Launch custom, fine-tuned, or any open source model and scale without limits.

With Baseten’s MCM system, you’ll experience:

• Cloud agnostic infrastructure with no GPU scarcity bottlenecks

• Out-of-the-box worldwide scale with 99.99% uptime

• Performance optimized infra by use case

• Systems that satisfy data residency and sovereignty requirements

Deploy on Baseten’s MCM infrastructure today.

AI RESEARCH

🧠 AI develops human-like object understanding

The Rundown: New research from Chinese scientists just revealed that AI models are spontaneously developing internal 'maps' of the world that closely mirror human conceptual understanding, making a leap from simple recognition to machine cognition.

The details:

• AI models were tested on 4.7M "odd-one-out" decisions across nearly 2,000 common objects, studying how they organize and understand the world.

• The AI naturally developed 66 core ways of thinking about objects, closely matching how humans mentally categorize things like animals, tools, and food.

• The AI’s conceptual map showed a strong alignment with human brain activity patterns, particularly in regions responsible for processing object categories.

• Rather than just memorizing patterns, the research showed that AI models build genuine internal concepts and meanings for objects.

Why it matters: While there is still a large subset of the AI world that believes that models are just “stochastic parrots”, we’re continuing to see more and more evidence of the tech showing hints of genuine reasoning and conceptual understanding — and that machine intelligence may work in similar ways to humans after all.

AI TRAINING

 📈 Decode video success patterns for better content

The Rundown: In this tutorial, you will learn how to use Perplexity's new Labs feature to analyze viral YouTube videos and successful channels to uncover hidden patterns, content strategies, and audience engagement tactics that drive massive viewership.

Step-by-step:

1. Access Perplexity Pro and click the lightbulb icon to activate Labs mode for advanced research capabilities

2. Create a comprehensive prompt: “Analyze viral YouTube videos and successful channels in AI to identify...”

3. Wait 8-12 minutes while the AI analyzes top videos, channel metrics, and engagement patterns across successful creators

4. Review detailed insights and ask follow-up questions for deeper analysis

Pro tip: Frame your initial prompt to focus on specific metrics you want to improve. The more targeted your research request, the more actionable your insights will be.

PRESENTED BY ENEY

🤖 Meet your Mac's new AI companion

The Rundown: After 17 years of crafting trusted Mac software, MacPaw is pioneering something entirely new — Eney, an AI computer being that bridges the gap between you and your digital tasks through natural conversation.

Eney's capabilities include:

• Intuitive assistance using natural language with no app switching or complex commands

• Real Mac actions like system optimization, image editing, meeting transcriptions, and more

• Proactive intelligence that learns your patterns and suggests relevant actions

Join Eney's closed beta and help shape the future of Mac companions.

AI & CHILDREN

🧒 UK study reveals AI's hidden impact on children

The Rundown: New research from The Alan Turing Institute uncovered a digital divide in children's AI use, with 22% of UK kids aged 8-12 already using AI — but with private school students nearly 3x more likely to have access than their state school peers.

The details:

• Private school kids showed 52% usage rates compared to just 18% in state schools, also reporting more frequent use and greater teacher awareness of AI.

• Environmental concerns emerged as an unexpected factor, with some children refusing to use AI after learning about its energy and water consumption.

• The study found children primarily use AI for creativity and learning, with high reports of children feeling the tool helps them communicate better.

• The research also included teachers, with 66% reporting AI use primarily for lesson planning, creating presentations, and designing homework.

Why it matters: This generation of kids is going to grow up with AI just as the previous one did with the Internet. While the benefits to creativity and communication were clear, so was the divide in public vs. private schools — with access and adoption potentially widening the educational gap for more privileged students.

QUICK HITS

🛠️ Trending AI Tools

• 🗣️ AssemblyAI -  Build a voice agent with fast and accurate real-time speech-to-text*

• 🎥 Seedance 1.0 - ByteDance’s new leaderboard-topping video model

• 💬 Chat Mode - Runway’s new conversational interface for generating content

• 👁️ Copilot Vision - AI companion that can see your PC in real time

*Sponsored Listing

💼 AI Job Opportunities

• 🤝 The Rundown - Partnerships Manager

• 🧳 Glean - Consulting Partner Manager

• 🧑‍💻 Perplexity AI - Full-Stack Engineer, Comet

• 🧠 Runway - Applied Research Lead, Language

📰 Everything else in AI today

Nvidia CEO Jensen Huang stated that he “pretty much disagrees with almost everything” Anthropic CEO Dario Amodei has said regarding AI and job automation.

A paper co-authored by Claude 4 Opus critiqued Apple researchers’ recent viral paper that argued LLMs can’t reason, finding flaws in the study’s design.

OpenAI rolled out updates to its Projects feature, with new support for deep research and voice mode alongside improved memory functionality.

AstraZeneca signed a $5.3B AI research deal with China’s CSPC, aiming to use AI to develop new oral medications for chronic diseases.

A new report from the New York Times detailed cases of ChatGPT use reinforcing and fueling user issues like delusions, conspiratorial beliefs, and mental health crises.

Tencent’s Hunyuan released Hunyuan 3D 2.1, an open-source model for generating 3D assets with cinematic textures and realism.

COMMUNITY

🎥 Join our next live workshop

Join our next workshop on Tuesday, June 17th at 3 PM EST with Keanan Koppenaver, Technical PMM at Retool. By the end of the workshop, you’ll know how to supercharge and automate your prospect follow-ups with AI agents.

RSVP here. Not a member? Join The Rundown University on a 14-day free trial.

See you soon,

Rowan, Joey, Zach, Alvaro, and Jason—The Rundown’s editorial team