A Newsletter Focused on *What AI and Robotics Can DO*

"The age of generalist robotics is here." Jensen Huang

Many people struggle to keep up with all the new AI updates. When people ask me how to adapt to so many updates, I encourage them to see them through a broad lens: AI is working toward full/general human-level intelligence.

Another lens one can take is to think through the capabilities of the technology. What is AI technology capable of?

In this newsletter I try to offer an overview of such capabilities. I drew from Claire Zau’s newsletter, The-Decoder, Tech Crunch, ArVix, and Interesting Engineeering.

This first release is free. Future releases will be available to subscribers.

Supervisors

AI agents are growing increasingly capable of supervising other agents that carry out multiple step tasks. Manus from China is an AI system designed for independent thinking, planning, and execution without human oversight, setting it apart from traditional AI assistants. Built on a multi-agent architecture, Manus operates like an executive overseeing specialized sub-agents (CZ – 3/17/25)

OpenAI has launched new tools for building agents.

Applications: A business assistant that can activate multiple assistants. A classroom aid that can work with multiple AI classroom aids.

For more on agents, see AI Agents: Evolution, Architecture, and Real-World Applications and A Comprehensive Survey on Multi-Agent Cooperative Decision-Making: Scenarios, Approaches, Challenges and Perspectives

Coaching

Initially launching on the Xbox mobile app, Copilot for Gaming serves as a second-screen companion, offering real-time gameplay analysis, strategy tips, and in-game guidance. CZ – 3/17/25)

Applications: AI systems can help students learn in real-time, provide suggestions for teachers in real time, and help business leaders manage difficult situations. It s important to note that the Xbox app only supports digital applications now, but improved multimodal vision capabilities mean it could support live environments

This article focuses on AI ‘wingmen’ bots to write profiles and flirt on dating apps, but the idea can be easily expanded to support student learning, managing clients, and trying to close sales deals. CZ – 3/17/25)

Creative Writing

Sam Altman teases model with advanced creative writing capabilities: The model is said to demonstrate remarkable skill in metafiction, capturing nuanced themes of grief, memory, and AI consciousness with literary depth CZ – 3/17/25).

Link to the original Twitter/X post.

Applications: The positive application is more challenging here, because if AI can duplicate creative writing abilities then it can do a lot of “unique human work” that

Multistep Deep “Research”

A number of tools (ChatGPT, Google, Perplexity) have improved their multi-step “Deep Research” tools.

How do these work?

Step 1: Initial Prompt

The user poses an initial request or question that requires in-depth research.

- Example:

"Find the major causes of urban poverty in America and explain their historical roots."

Step 2: AI Analysis and Research

The AI conducts a preliminary scan of available data, references its training knowledge, and gathers initial insights.

- Example:

The AI identifies initial factors like housing discrimination, economic inequality, or educational disparities.

Step 3: Re-prompting (Clarifying and Expanding)

Based on initial insights, the AI re-evaluates what information might be missing, unclear, or incomplete. It internally generates new prompts to refine and expand its research.

- Purpose:

- Clarify ambiguities: "What exactly constitutes historical roots here?"

- Fill information gaps: "How did specific policies historically contribute to urban poverty?"

- Example (internal re-prompt):

"Research specifically how redlining policies from the 1930s impacted economic inequality in urban America."

Step 4: Follow-up Prompting

After acquiring refined information, the AI proactively anticipates further questions, implications, or depth the user might require. It generates follow-up prompts to deepen or enhance the original query.

- Purpose:

- Provide contextual depth and additional layers of understanding.

- Anticipate user needs by asking follow-up questions the user might not initially consider.

- Example (follow-up prompts):

- "What were the lasting impacts of redlining on wealth accumulation among minority groups?"

- "What were the progressive-era reform efforts aimed at reducing urban poverty?"

Step 5: Final Synthesis and Delivery

The AI synthesizes the findings from the initial prompt, re-prompting, and follow-up prompting into a comprehensive and coherent final output.

- Example:

An organized summary detailing historical causes of urban poverty, specific historical policies, and relevant contemporary implications.

Applications: Very good research papers for students and companies (once they train the deep researchers on their own existing research and what is available in other proprietary databases)

An AI-written paper has already passed peer review.

Coding

Dario Amodei, CEO of Anthropic, made a splash when he said 90% of code will be written by AI in 3-6 months, with maybe almost all of it written by AI in 12 months. Cognition AI, the company behind AI coder Devin, has raised hundreds of millions on a $4 billion evaluation.

Despite having almost no coding experience myself (I know a bit of html and C++), I’ve been able to build a few basic games with only text prompts. A friend of mine with no coding experience has built multiple apps and chrome extensions.

Human Communication

Efforts to improve human-AI communication/interaction are accellerating. Cartesia Raises $64M to Advance Real-Time Voice: Their Sonic 2.0 model achieves ultra-low latency (90ms full models, 40ms real-time) and industry-leading voice cloning, capturing accents and tonal nuances CZ – 3/17/25).

Anthropic is prepping a voice mode for Claude

Regal.ai has made millions of phone calls.

Communication isn’t just language, but Alibaba Releases AI Model That Reads Emotions to Take On OpenAI: Alibaba’s Tongyi Lab has released R1-Omni, an open-source AI model capable of inferring human emotions from videos while also describing clothing and surroundings, enhancing computer vision capabilities.

Story Telling

Gemini 2.0 Flash native image generation: Generates illustrated stories while maintaining character and setting consistency. Supports multi-turn dialogue-based edits (conversational image editing), allowing iterative refinements. CZ – 3/17/25)

Robotics

“The age of generalist robotics is here,” Jensen Huang

See more at Endgadget

Rapid advances continue in robotics. In tests with real robots, a new model, FP3, learned completely new tasks with just 80 examples, achieving over 90% success even when handling objects it had never seen before.

Dexterity, a startup specializing in industrial robots with “human-like” finesse, has raised $95 million at a post-money valuation of $1.65 billion, per Bloomberg.

Google DeepMind, Google’s AI research lab, on Wednesday announced new AI models called Gemini Robotics designed to enable real-world machines to interact with objects, navigate environments, and more.

Google co-founder Larry Page is building a new company called Dynatomics that’s focused on applying AI to product manufacturing.

Today, NVIDIA revealed Groot N1, a foundation model for humanoid robotics. Nvidia claims that Groot N1 features a “dual system architecture” for “thinking fast and slow,” inspired by human cognitive processes.

*Deployment*

The technologies are advancing faster than we can deploy it, and this means that few people are aware of how quickly things are advancing. As Ethan Mollick often notes (including today), even if AI development stopped, it would probably take at least a decade to deploy these technologies into society.

Deployment is happening, however.

Every part of Amazon is affected by AI/emerging AGI

How the AI Talent Race Is Reshaping the Tech Job Market: 36% of IT job listings in January were AI-related, particularly in tech giants, finance, and consulting firms. AI adoption is spreading beyond tech—retail, utilities, and healthcare are hiring for AI-integrated roles. CZ – 3/17/25)

As AI nurses reshape hospital care, human nurses are pushing back: Hippocratic AI and others are developing AI nurses like "Ana" to handle admin tasks, answer patient queries, and monitor vitals 24/7, easing nurse workloads. However, National Nurses United warns AI is being used to "de-skill" and replace caregivers, compromising care quality. CZ – 3/17/25)

See also: Human Digital Twins in Personalized Healthcare: An Overview and Future Perspectives

Google is going to release new open models for drug discovery and has launched a new health-care related features search.

GM has teamed-up with NVIDIA to integrate robotics into self-driving cars, factories, and other robots.

The Pentagon has begun to give AI agents a role in decision making, ops planning

Education

US K-12 schools are doing more to support AI understanding and integration, but it’s all still pretty basic (mostly Level I bot prompting instruction, simple tools such as MS/SchoolAI/etc).

China’s six-year-olds are already being offered AI classes in school in a bid to train the next generation of DeepSeek founders: Beijing schools will introduce AI courses for students as young as six, covering chatbot use, AI fundamentals, and ethics, with at least eight hours of AI instruction per year. The Ministry of Education has already piloted AI programs in 184 schools and announced plans for a 2025 white paper on AI education. CZ – 3/17/25)

See also: LLM Agents for Education: Advances and Applications

AGI

My favorite topic :)

Most, including Demis Hassabis, are saying AGI within 5-10 years, but it could be sooner.

Half of adults believe the systems are already smarter than them, which is probably true.

It is fair to say that most people don’t believe language models alone won’t get us to AGI, but those with more aggressive predictions of the time-frame for AGIs arrival understand that.

Costs

The costs of the most powerful intelligence is rapidly dropping. ErnieBot, released by Baidu (China) claims ChatGPt4.5 capabilities at 50% of the price.

Better AI

AI is going to keep getting better. Dr. Tim Dasey noted: