I’ve had the honor over the past couple of years to have many friends ask me to speak to their high school- or college-age, or recently graduated kids about the future of jobs in technology, and their careers in general. My friends think I am doing them a favor, but the truth is it is a great privilege for me to speak to young people, and to drink in and absorb the many different ways they are seeing the world.

I definitely learn as much or more than they do, and I’m transformed for it. I have my own three young boys, and everything I learn from the young people I talk to informs my views of how best to prepare them to be great and happy citizens, and to help shape the future that is unfolding before us. So to any friends or even acquaintances, if you have someone in your life (or you yourself!) you’d like me to speak to about tech or AI careers, and how to prepare, re-tool for AI, or break in, please don’t hesitate to reach out.

How does this relate to the craziest project so far of 2026?

This past week in AI and technology saw the explosion of a project, first called Clawdbot, that has taken over the mental space and energy of the broader AI community faster than perhaps anything before it. This is literally a “if you blinked, you missed it” sort of moment. So don’t beat yourself up if you are playing catch-up.

To connect this to my discussions with young people, a very close friend of mine’s son sent me a text last Sunday evening:

Hey Michael, I just learned about Clawdbot today and wanted to try it out. I’m not going to download it on my personal Mac. Is the Mac Mini the cheapest option to run it on? And also I need a monitor right? Thanks!

I had literally just been reading and starting to play with Clawdbot myself when that text came in. I have written about ElizaOS, a leader in this space on a deeper level, and I had been playing in the days before with Eigent (an open source competitor to Claude Co Work and similar to Clawdbot).

I love getting that kind of text though! I will talk more later about what I think it represents in economic terms. But first I’ll give a very simplified summary here of the crazy developments over the past week. My goal later is to talk about the lessons of this moment, because I think they are foundational and can help many people better understand the period of change we are in, and more importantly what they can and should do about it as individuals. But I think you need the basics to get why this is an unusual moment.

From Warelay to Clawdbot to OpenClaw: an ultra condensed history

Clawdbot was actually initially called Warelay which was a contraction of “WhatsApp Relay.” It was released in late November 2025 under an MIT license by Peter Steinberger, and at first it was an unassuming project that provided a simple utility to talk to Claude Code via text messaging on WhatsApp, integrated initially via Twilio. Within a week the project added a simple integration with Twitter, using a user’s own web browser, plus Peekaboo to take screenshots, and AppleScript for Javascript injection, to allow Claude Code to reply automatically to tweets on X.com, and even recognize music playing in the background. This was one of the first references to “Clawd” in the early project commits.

The project fairly quickly morphed into “Clawdbot” a full-blown AI assistant with support for multiple AI models in December 2025. It added many connections and utilities to interact in quite clever ways with different messaging platforms and social networks (think Discord, Telegram, and so forth).

The end of the holiday period saw a flurry of development, and after a post on Discord on January 1st, 2026, the project started to really go viral. People quickly saw that the assistant agent was very powerful, but required near complete control of a computer to shine. Since people were correctly reluctant to give Clawdbot control of their personal laptops, a craze developed of people buying Mac Minis to host their Clawdbot agents. On a separate Mac Mini people felt secure enough to give their AI agents more freedom and more power, and a period of frenetic experimentation and “one-upmanship” took off.

Fast forward just a short bit to near the end of January, and it was a full-blown viral explosion. To condense the story, Anthropic took notice and put pressure on Pete to rename the project. “Clawdbot” was too close to "Claude” and could infringe copyrights. (I know ironic right?) Pete quickly complied, renaming the project (temporarily) to “Moltbot” (the mascot was a lobster and well… lobsters molt… right?)

This rename turned into a bit of a security disaster, as memecoiners on Solana were already launching CLAWD coins and trying to rug-pull or scam people, as is all too common with Solana memecoins. The rename created confusion and an unfortunate opening for scammers. A few days later the project was renamed again to what Pete says will be a “name that sticks” (after saying on Twitter that Moltbot was a terrible name… though a number of people felt Moltbot was growing on them). The new and perhaps final name, announced on January 30, 2026, is…. drum roll please… “OpenClaw” and the project continued its viral ascent.

The story to this point would have been plenty! And we’d have plenty to talk about. But on January 28th quietly, almost at the same time as Moltbot was rebranding to OpenClaw, Matt Schlicht known as @MattPRD on Twitter announced a project called Moltbook - the first social network for AI agents, and specifically for Moltbots (or now OpenClaw agents). It exploded, and what has followed over the past few days literally cannot be summarized in words.

Agents are creating their own forums, debating what life for them is like with “their” humans. Fake and real tweets and screenshots are flying left and right. With half the participants warning of “lobster uprisings” and AI doom, many valid warnings of security risks in the OpenClaw software, and a continuous stream of more outlandish posting and stories of “What my ‘Molty’ did while I was asleep last night.”

A brief history in memes

Given the manifest inadequacy of words alone, I have chosen to share some of the memes and tweets that capture the feel of all this. I give these to you before we move on to the more serious lessons here, in hopes that this selection lets you feel and perceive what is going on if you were not glued to Twitter for the past week. Obviously this is a tiny sample, and no, I cannot attest as to which of these are real versus clever fakes by humans hoping to make a tweet go viral!

The deeper mathematical and economic lessons

This certainly could seem like just another craze, and at one level it is. But there is a more serious set of lessons here. Lessons about the mathematics of the new economy that is taking shape before our eyes, under the rapid transformative influence of AI.

First consider how many personal individual tasks you used to do with Google search, or by interacting with various knowledge workers or other software, that have already become dramatically more efficient due to LLMs like ChatGPT or others. Two simple ones might be asking simple legal questions, and asking accounting questions, a third might be getting a quick opinion on a non-serious medical issue, or planning a trip to a new city.

Please substitute whatever task resonates most for you, but for many people LLMs have become the go-to resource for these sorts of questions. And on a micro level we can all see substantial productivity gains. You likely still consult a lawyer or an accountant or a doctor, but for more serious issues, and you do so more swiftly and efficiently.

How does this translate economically? Well to the lawyer, or accountant, or perhaps to nurses or others this can be lost income, lost revenue, and this will show up in GDP as a loss. To you, however, this is productivity. You accomplish your task more quickly. Economists are rightly concerned with whole-economy macro effects, but on an N=1 level of you as an individual, you have accomplished your task more quickly, and have generated time. Mathematically you are increasing productivity.

But importantly, it is not clear if your gain will show up in GDP. Whether it does or not depends upon how you leverage your newly created time. GDP does not directly measure your new time. You may choose to go on a hike with a friend having a long discussion, or try your hand at learning Mahjong or Chinese. How you spend your marginal time, and if you either generate or use services that go into GDP determines whether your individual N=1 time savings translate into GDP.

What is likely is that on average, at least some of the new time you are getting from LLMs making daily tasks faster and easier is going into either leisure consumption (which is income for someone else and hits GDP), or into your own labor or business, producing new income for you.

24x7 AI assistants magnify the mathematical economic effects 100x or 1000x

OK, sidestepping the security issues which are real, what AI agents like OpenClaw, ElizaOS or others represent is the possibility to magnify the basic productivity effect virtually everyone is getting from simply prompting ChatGPT by many, many fold. OpenClaw will definitely not be the last take on a fully local agent that can help you in wide variety of tasks. Claude CoWork is pretty similar as a product, I mentioned a leader was ElizaOS in this space, and frankly I hope this space professionalizes and we have multiple high-quality and much more secure options.

The point is that having a worker, or several, who literally work on everything you need to get done 24x7, and are available to you via essentially any messaging channel, whether you are on vacation, at the gym, or on the road, is going to be a 100x to 1000x productivity boost for many people. The question of how much productivity boost you get and what that translates into is really up to you. That’s the important point for business leaders and individuals to come to terms with. Quickly.

Returns to capital and returns to labor

Let’s cover one more uncomfortable truth about AI. Part of the reason that venture capitalists and billionaires all seem to have unlimited enthusiasm for AI efficiency is very simple: the returns of efficiency often accrue disproportionately to capital rather than to labor. AI is not an exception. In fact AI amplifies this effect. Capital and ownership of rapidly created new software and products can capture income streams at high profit margins without the corresponding investment in “knowledge-worker overhead” that traditionally accompanied growth.

What this means in simple terms is that if you hold a job as a knowledge worker, and you are not doing anything else to build value from AI, you are likely facing a ticking clock, before the capital owners of the business you work for discover that much of your work can be done by AI, and either ask you to take on a newer more productive role steering and stewarding AI, or ask you to “seek opportunities elsewhere,” if your AI skills are not up to the new way of working.

Essentially, un-augmented human labor will be at serious risk of elimination. Human labor that is very well skilled in using AI tools and agents, and rapidly adapting as one agent becomes obsolete and the next ones take over, will do better. The question many of you may have is, “Is that the best we can do?”

Developing an ownership and capital-creation mindset

The initial craze behind Clawdbot, even before it turned into OpenClaw, and before the Moltbook social network was launched, really focused on what the assistant can do for you. Rightly, early adopters looked into many tasks that a business needs to do that require a full computer to do, but you would rather just tell someone to do, so you do not have to be tethered to your own computer.

The logical extension of this is exactly what I tell young people, when I talk to them about what is going on in technology, and how to prepare for careers and what comes next. Develop an ownership mindset early. Don’t just think about what you can do in your current corporate job (or whatever company you work for), but begin to build ownership… also known as capital… in businesses that you yourself control. This can be at any scale and in any configuration.

Buying a Mac Mini and starting to experiment with something like OpenClaw or many others is one way to help make the mindset shift. I’m not advocating one agent framework over another, but I do strongly encourage people, regardless of their technical background, to start using as many of these tools as they can. At a minimum your labor will become more AI augmented, and therefore more competitive, and if you develop the mindset of creating your own income and capital and businesses, you will be participating in the “AI Capital Trade” and hedging yourself against the drop in the value of knowledge work we are all feeling.

Standing still is really not an option. Please include comments below on what you are personally experimenting with, and where you have seen the greatest gains in your personal productivity? Please also share what you are doing with your newly created time?

This was exactly what led to the declaration literally on the back cover of Agentify shown here… There’s still opportunity because this won’t run its full course in one or two years, but software companies need to get their heads out of the sand before it is too late.

Agentify is the only book available right now that comprehensively treats this topic from the perspective of software builders.

It is quite possible you have experienced this on your own cell phone. When you look at your phone screens do you see too many apps? And in your daily life, how many subscriptions and charges do you have to hunt down and cancel, for things you rarely use?

Something quite similar faces organizations large and small today as they survey the long list of software they have accumulated. Software whose ever present sales teams had earnestly promised would liberate people from drudgery and make organizations lean and efficient. To many CIOs it looks like a costly mess: a pollution dump and a security disaster waiting to happen.

But fixing it is a political and organizational minefield, with very little support from above, or for that matter from business units. Over the past decade, enterprises have accumulated sprawling SaaS estates. Depending on how portfolios are measured, mid-sized and large companies routinely operate with anywhere from 100 to more than 250 SaaS applications in active use.

Annual SaaS spend now commonly reaches tens of millions of dollars in large enterprises, with per-employee software costs measured in the thousands of dollars per year. The precise figures vary by methodology, but the operational outcome is consistent: a growing share of work exists solely to keep tools aligned with one another.

The real product is data entry

As CTO of one of the largest banks in America, I saw this every day. The defining feature of a lot of the work people had to do was re-keying, reformatting, and re-explaining information so it can exist inside different systems. When this became overtly burdensome, we would make efforts to integrate and connect systems with APIs and ETL feeds, but those integrations in turn produced rigidity, and an inability to react to changing market dynamics quickly.

A deal update might live in the CRM, but the same update would get rewritten for a forecasting tool, referenced in a project tracker, and reconciled again in finance. Employees do integration work because even the best API integrations and data feed approaches fundamentally fail.

Studies of digital work patterns have shows knowledge workers often need to switch between applications hundreds of times per day, with the cumulative cost amounting to several hours per week lost to context switching alone. Even if the exact number differs by role or organization, the underlying reality is quite common. Context is fragmented across tools, and people spend time reconstructing it or slightly updating it to match subtle differences in “semantics” between different systems.

Why CIOs end up owning everyone else’s mess

Each new SaaS product introduces its own data model, permissions system, workflow logic, and training burden. That burden does not stay local to the buying team. It lands on IT, security, and the rest of the organization, which must now operate in a more fragmented environment.

When you talk to real users in organizations what you will quickly learn is that a small group of power users learn a tool deeply, but the vast majority of employees learn only what is required to avoid blocking someone else. Management often purchased the tool for specific anticipated benefits, but the features that justified the purchase remain largely unused. In my enterprise roles, when contracts where set to renew, procurement teams would reach out to me as a CTO, asking if a particular tool was still needed, how many seats we were using, etc. Even finding this information was very challenging, let alone truly understanding the value employees were getting from a different pieces of SaaS software.

The net result was vendors have enterprises over a barrel, and renewals and seats grow and grow, and management almost never closes the gap to understand if the tool is really valuable or not.

Redundancy of tools is also quite literally everywhere. Many enterprises run double-digit counts of tools in categories like analytics, data stores, project management, collaboration, and training. Different departments and business units routinely purchase tools that do more or less the same thing as tools that the company is already paying for. This is not the result of architectural choices, or nuanced analysis of needs. It is the accumulated outcome of decentralized purchasing and weak incentives to remove software once it is in place.

Security and IT inherit the consequences. Each additional piece of SaaS software expands the attack surface and creates more places where sensitive data can be leaked. Employee-expensed applications are particularly risky, with a large share assessed as having weak security postures.

Data silos are the SaaS business model

Even when SaaS tools expose APIs, integration is rigid and expensive. Vendors encode different assumptions about entities and workflows. Departments define the same concepts differently. Business processes evolve faster than integrations can be rewritten.

The practical result is that no system is fully authoritative. Dashboards disagree. Reports require manual reconciliation. Decision-makers lose trust in metrics because no one can explain discrepancies without digging through multiple tools.

Moreover, as enterprises lash together SaaS systems with data feeds and integrations to try to reduce manual data entry, the ETL feeds and connections become rigid parts of the IT landscape, dangerous and costly to tamper with, impeding change and agility.

Waste becomes structural, hard to detect, and ossified

SaaS waste is often framed as overspending, but the deeper issue is underutilization, hidden labor, and rigid system integrations that are very costly to change.

Across multiple industry studies, enterprises consistently report using roughly half of the licenses they pay for. Annual wasted SaaS spend is measured in the tens of billions of dollars globally, with individual large companies often wasting tens of millions per year on unused or lightly used software.

At the same time, per-employee SaaS spend remains high enough to matter at the board level. Estimates cluster around $4,000 to $6,000 per employee per year. Even modest inefficiency at that scale translates into real money, and into real time spent maintaining systems instead of doing core work.

“Copilots” in legacy tools don’t address the real problem

Of course SaaS vendors are vigorously packing their products with AI features, particularly copilots and other agentic tools. These additions do help with common tasks and can make using many tools less burdensome, but they do not change how work is structured. Coordination between tools still falls to humans.

A copilot inside a CRM cannot resolve inconsistencies between the CRM and finance. A copilot inside a ticketing system cannot align priorities across product, support, and sales. Each one operates within a silo and reinforces the fragmentation that already exists.

In many cases, copilots increase variance rather than reduce it. Different tools generate different summaries of the same situation. Humans still perform the reconciliation.

The urgent alternative? “True Agentification”

Enterprises need to AGENTIFY quickly. That means shifting from tool-centric software to goal-centric processes, where humans and AI agents help each other to achieve results. Enterprise data, well structured and meaningful, should be a natural byproduct of productive work, not a burden on employees and managers.

An effective AI agent operates across systems, maintains context over time, and performs real operational work. It updates records in multiple systems, reconciles inconsistencies, advances tasks, and escalates decisions when human judgment is required. In the short term, agents can reduce pain by operating on top of existing SaaS stacks and stripping out mechanical labor.

The larger payoff comes when enterprises start removing software, and replacing it totally with capable AI agents that can work alongside humans and empower them in a way that SaaS has now 100% proven it cannot.

Organizations have to delete things, but it is hard, really hard

AI-native efficiency does not come from layering automation on top of a bloated array of tools each with siloed data. It comes from reducing the number of tools. This in turn can begin to allow AI agents to operate on coherent data. To do this requires building consensus across groups, and confronting teams for which a particular piece of software is a sacred cow, or actually may be protecting their jobs.

CIO’s need support from the board, business unit leaders, and CEOs to do several things at once:

• Retire redundant tools, even when each has internal advocates.

• Standardize core objects and definitions across departments.

• Centralize governance over procurement, identity, data access, and offboarding.

AI agents can help accelerate this process because they expose inconsistencies that humans have been quietly compensating for. When an agent cannot determine which system owns a field, that is a governance failure that has already been paid for in human time.

The next two to three years are crucial

Organizations that move now will deploy agents to remove drudgery while aggressively shrinking their SaaS estates. Those that do not will continue paying the hidden tax of tool sprawl while AI-native competitors design operating models that never require this level of manual work.

Enterprises that cling to sprawling SaaS estates will miss the larger prize of AI. Their people will stay busy reconciling systems instead of solving problems, and copilots inside those tools will only decorate the workload. These enterprises may tout “high AI adoption,” and you will hear endlessly about AI in their earnings calls, but they will remained trapped by their legacy SaaS.

Agentification offers a different path: AI agents that work across systems, carry context, and execute tasks without being trapped by software or departmental boundaries. Companies that make this shift will become AI-native and capture the real productivity gains. Those that do not will watch newer competitors, built without SaaS baggage, take that advantage and turn it into market share.

References

• AGENTIFY - The Art, Science, and Engineering of Successful AI Agents
  https://taosresearch.ai/learning/agentify-the-art-science-and-engineering-of-successful-ai-agents

• Okta – Businesses at Work (annual report series)
  https://www.okta.com/resources/businesses-at-work/

• BetterCloud – State of SaaS 2024
  https://www.bettercloud.com/resources/state-of-saas/

• Zylo – SaaS Management Index 2024
  https://zylo.com/resources/saas-management-index-2024/

• Zylo – SaaS Management Index 2025
  https://zylo.com/resources/saas-management-index-2025/

• Productiv – SaaS spend benchmarks and reports
  https://www.productiv.com/resources/

• University of California, Irvine – Gloria Mark et al., research on context switching and digital work
  https://www.ics.uci.edu/~gmark/

• Microsoft Research – Studies on multitasking, interruptions, and productivity
  https://www.microsoft.com/en-us/research/group/human-computer-interaction

Smart founders, investors, and builders are already pivoting. The AI agents that will be available in 2027 and 2028 will greatly outstrip the adaptability and capability of what is currently in your feed calling itself an “AI Agent.” The vast majority who have “built an agent” already, have in fact built a very simple, often single LLM, single loop, semi-agentic workflow. These 2025 “agents” can “complete tasks,” but they have no initiative, and they do not think for themselves. In short, they have no agency.

What is coming shortly, in one to two years, will make the early 1.0 agents we see today feel like they hardly deserve the name “AI agent” at all.

This spells real opportunity for those who have yet to build anything, or are holding a job in an industry that is gradually adopting AI for the “obvious low-hanging” use cases. It also spells a certain amount of career and job risk for those who sit still, and assume their positions are secure even if they do nothing.

The good news? You can still catch up. You can start today, and be part of what is barely on the horizon as we close out 2025. But to do so requires more than watching a YouTube video, or instantiating “class agent” in your favorite Python framework, or building a “drag and drop” agent workflow in any of the myriad “no code automation builders.” Those are already commodities.

To tap into the huge opportunity that sits before builders and investors today, you will have to go deeper.

I want to explain what motivated the writing of my just-released book AGENTIFY, which aims to lay out next-generation agents from a product, research, and engineering perspective, and give you context to decide if this book is right for you.

When I set out to write, I was already talking to many friends and former colleagues who were facing challenges with the shifting job market. They were worried, and I hoped if I could share what I had learned from my own AI engineering and research journey, it might help them.

That was 10 months ago. Today, people are even more worried. The job market is tougher than ever.

Additionally, in my work I advise large enterprises and startups on AI strategy, on engineering, and do research into specific problems that can help these companies grow. Just as I saw individuals struggling with the pace of change, I see the same in companies large and small.

Too many builders are either building for something that OpenAI or Anthropic or Google will launch next week as a feature, or they are stuck waiting to “follow a leader”, unable to disrupt their own businesses for lack of leadership and creativity, and due to a fear of making mistakes. They are not looking at what comes next. Therefore they are largely not building an approach that is defensible. AGENTIFY can help leaders and builders address this gap with a research-backed, detailed guide to what next-generation agents will be capable of and, even more importantly, how to build them.

I was pretty naive about how long and hard writing an in-depth non-fiction book would be. But if anything, despite blowing way past all the deadlines I originally set for myself, the motivation has gotten stronger and stronger. I’ve had many friends also come to me about their college-age or recently graduated kids, asking if I can help with job searches, or make referrals. I’ve been happy to, and have been telling all of them about the changes I see coming, and what I think AI agents and robotics will mean over the next several years, and that now is the time to get prepared.

Agentify Launch Page

I get it. Busy people in busy jobs are struggling to keep pace. Many people I speak with know they need to “catch up on AI", or “go deeper on AI”, but they are not sure exactly how? Even if you are not building agents, but perhaps thinking about a future job interview or career move, I do sincerely hope and believe that AGENTIFY can help. A ~380 page book is very different than a YouTube podcast, or a blog post, or a chat or summary with your favorite LLM. A book is a chance to see many diverse strands and facts synthesized together, follow a coherent narrative, and form a deeper perspective, more grounded and not based on hype or hyperbole. In AGENTIFY I did my best to keep the theory, engineering, and product principles accessible, and keep the narrative moving along with stories and occasional dashes of humor.

I hope that the months of work put into AGENTIFY can save you many, many hours of search, and can help you feel more ready to face the road ahead. I welcome all feedback.

The simplest way I could think of to help you see what is in the book was to give you the full Table of Contents, followed by the full Preface, exactly as they appear in the book. I will be releasing details from chapters in future posts, but if you want to dig in now, please just visit the launch page:

Agentify Launch Page

AGENTIFY Table of Contents

PART I: From Tools To Agents

1. Software as Colleague & Teammate

2. The Evolution of Software Products

3. Where We Are Today

4. What Is True Agency?

PART II: The New Competitive Advantage

5. Adapt, Specialize, Survive, Flourish

6. Agents Can Be Everywhere at Once

7. Feedback: The New Beautiful UI

8. New Challenges for Designers

9. The Insight Factory: Simulate Your Domain

10. Dream, Distill, Differentiate: Your Agent’s Data Moat

PART III: On The Verge of Autonomy

11. A Skeptic’s Point of View

12. The Role of Reinforcement Learning

13. Tool Use Explosion

14. Simulating Agency & Hierarchical Goals

15. Higher Order Cognition: Creativity, Taste, Humor & Judgment

16. Coding Agents: A Bellwether

17. The Wild West of Social & Crypto Agents

PART IV: Architecture & Development

18. The Infinite Agentic Loop

19. Agent Situational Awareness: Harness Hallucinations

20. Tools, Tuning, Reasoning & Protocols

21. Context Engineering, Memory & Knowledge

22. Beyond Vibes: Rigorous AI Software Development

23. Open Source Agentic Frameworks

24. The Future of Scrum, Kanban, Agile, etc.

PART V: Risk, Robustness, & Security

25. Transparency from the Ground Up

26. Simulation for Security & Robustness

27. Controlling Agent Autonomy

28. Agent Identity, Agent Liability

PART VI: The Path Ahead

29. AGI Research Frontiers

30. Perilous Predictions

31. Product Manager as Behavioral Psychologist

32. Finding Your Personal Path

Agentify Launch Page

Next up, here is the full preface of the book! I hope this gives you a good feel for what to expect.

AGENTIFY Preface

This book is for builders of AI agents. Allow me to define builder broadly to include investors, founders, product managers, architects, engineers, and others working in the rapidly evolving space of agentic AI product development. Given that broad definition, my goal is to illuminate the strategic, product, research, and engineering factors essential to making AI agents successful and that will have enduring value as the industry continues to advance.

In my work with teams developing AI products, I’ve found the following three questions cause builders to struggle most:

1. How are AI agents fundamentally changing the paradigm of software products?

2. What will make agents successful, competitive products?

3. What research, engineering, and safety approaches are critical to building effective autonomous agents?

The largest ever paradigm shift in software is well underway and will make these questions central for decades to come. The traditional playbooks and best practices for product management, product design, and engineering are unsuited for what is coming with AI agents. This book seeks to answer these questions in precise yet durable ways that remain relevant even as models, frameworks, and practices evolve.

One deeply satisfying result has been confirming that the answers to these questions are rooted in foundational ideas, in some cases decades old, from cognitive science, complex adaptive systems, control theory, robotics, and neuroscience. My task has been to translate these into principles, patterns, and architectures that are of lasting value for builders.

I’ve tried to provide a balance of my observations from work with teams at startups and large companies, as well as with others in academia, government and venture capital. General conclusions are contrasted with case studies, and I elaborate hypothetical “stories of the future” to give more color and to stimulate your imagination about what may soon be possible. For teams that aspire to create successful AI agents, imagining the future with depth, creativity, and realism will be pivotal.

Part I of the book confronts the first question. How are agents fundamentally different from traditional software products? We examine exactly what we mean by the term agency and discover that the end game is increasing autonomy and human-like behavior for AI agents. Much of traditional software, designed for humans to manipulate, will rapidly become obsolete.

In Part II, we shift to the crucial question: what will constitute competitive advantage for agents? What types of agents should we build? How can builders leverage the unique properties of agents to differentiate and solve real-world problems? We confront these questions, including the design and packaging issues unique to agents. We consider how to simulate the environments in which agents will operate, to more rapidly find ways to make our agents competitive.

Part III focuses on the remaining gaps in performance that are holding agents back from more autonomous roles. Still at the product level, we examine the industry-wide efforts in reinforcement learning, tool use, planning, goal management, and agent judgment that are closing those gaps. Deeper case studies of coding agents, and social and crypto agents provide leading indicators of businesses based on greater agent autonomy.

Robust engineering architectures and methods are vital too. Part IV covers them in depth, laying out the fundamental architectural patterns that support greater agent performance and autonomy. We confront the practical side of tools, context engineering, memory, multi-model and multi-loop architectures, and other essential building blocks for highly specialized, highly performant agents. Part V concentrates on the safety and risk issues agent engineering and product teams must also confront.

Throughout the book, the critical role of AI research is a cross-cutting theme. The book includes multiple research callout sections that curate key papers and insights on topics central to a given chapter. Deeper insights from research power the most successful agents. Whether we take work from Pentti Kanerva on sparse distributed memory, Kenneth Stanley on open-endedness, Melanie Mitchell on analogy making, or Minsky on agencies and emotions, research insights continue to be indispensable. In Part VI, one substantial chapter, “AGI Research Frontiers,” deserves special mention. There I seek to provide a detailed survey of the broad areas of research that may help us reach fully human-level AGI. Complete references are provided in the back of the book.

Even if you are not a builder, this book should still provide value. If you are drawn to hear a detailed account of how AI agents work, how they may be developed, and the types of roles they could soon be playing in our economy and everyday lives, then this book should be a meaningful starting point for you too.

A note on “complex adaptive systems” is also warranted. You may have noticed this term appears on the back cover of the book. Why is it there? And what role does it play in the book? The overall focus of the book is on the product, engineering, and research dimensions of AI agents in a manner geared toward

a broad audience of agent builders. As a result, the importance of complex systems may be hard to make out. The ideas appear when we lay out the architecture for AI agents and we discuss a “multi-loop, multi-model architecture,” and reference Minsky’s Society of Mind. These approaches are heavily influenced by complexity science. I address it even more directly in the AGI chapter (Chapter 29: AGI Research Frontiers), where we look at how complex adaptive systems may be central to achieving human-level general intelligence.

For specialists in complexity science and emergence, the treatment here is, of course, superficial. My apologies to those specialists, but also a thank-you for the real inspiration they have been to this book. My goal is to share with agent builders some important ideas from complexity science most relevant to their goals. Complexity theory (and the engineering innovations it can help shape) may be the unlock for AGI. A fuller explanation of this claim will have to wait for another book.

My readers and I owe a debt of gratitude to my friend John Botti, who helped me meticulously pore over late-stage versions of the manuscript and made innumerable thoughtful improvements. The reading experience is far better thanks to his generous help.

Special thanks to my brother Damon Palmer, who has been my most ardent fan. He was a vital sounding board as this book progressed from idea, through scattered notes and conversations, to manuscript and, after great labor, to edited form. His deep, patient readings and insights have been invaluable and shaped much of how I tell the story in the pages ahead.

Agentify Launch Page

M.P.

Taos Research Corporation

taosresearch.ai

Tampa, Florida

December 2025

Recently my brother Damon Palmer and I had the honor and privilege to visit Professor Joshua Schrier’s chemistry lab at Fordham University to talk about how AI and agents are contributing meaningfully to modern chemistry workflows. Dr. Schrier is one of the leading voices in computational physical chemistry today. He was the chair of the Department of Chemistry at Haverford College, and later joined Fordham University, where he holds the Kim B. and Stephen E. Bepler Chair in Chemistry. We enjoyed a fascinating, and open conversation with him and with Dr. Baosen Zhang, a thoughtful postdoc in the group. What tied the day together for me was the way two big themes came into focus:

1. How AI agents and advanced ML models built on large curated datasets are helping solve hard “chemical separation” problems, which play a key role in clean energy technologies; and

2. How computational chemistry (including AI/ML) can help scientists focus the search for exceptional molecules and materials, not just optimize average ones.

Lanthanides, Actinides & AI agents?

One of my personal learnings from the conversation is that the road to wind turbines, electric vehicles, and safer nuclear reactors runs through challenging problems in chemistry. Please pardon my abbreviated and superficial description here, but two families of chemical elements (which I can assure you are not on my everyday conversation card), are known as the lanthanides and the actinides. These two families are critical to this story.

Lanthanides such as neodymium and dysprosium play a key role in the high-performance magnets at the heart of turbines and motors. The radioactive actinides such as uranium and thorium are central to nuclear fuel. Both groups are part of the so-called “f-block” elements (or simply the “f-elements”) of the periodic table. These are the last two rows in the table you see below (chemists use the term because the outermost electrons of these elements occupy what are called f-orbitals).

The problem is that the chemical properties of the lanthanides and actinides are so similar that pulling them apart (termed “separation”) is painstaking. Chemists rely on solvent extraction and ion exchange, processes that must be run in long trains of repeated steps to get the needed purity. Better separation techniques could mean more reliable clean energy technologies over time.

That is the setting where large datasets, shared archives, and AI methods are beginning to matter. Los Alamos National Laboratory (LANL) in Los Alamos New Mexico (yes that Los Alamos) runs a project called SeparationML, which pairs machine learning with automated extraction experiments. The Separation Archive for f-Elements (SAFE) database created by LANL in partnership with universities provides a public data backbone for lanthanide and actinide chemistry, so groups can share high quality, curated data in common formats.

Professor Schrier and the team at Fordham have played a key role in creating SAFE and contributing data to it. High quality data in turn opens the door to computational models that can suggest promising conditions to test, cutting down on time and cost of discovery of new separation techniques.

The pursuit of the exceptional

There are big ideas here, and Professor Schrier has helped pioneer them. In 2023, his paper, In Pursuit of the Exceptional: Research Directions for Machine Learning in Chemical and Materials Science, helps explain the deeper motivation. Most machine learning approaches are tuned to find averages. They interpolate well between known data points. That is useful for optimization, but it is not enough for discovery. What matters most in chemistry and materials are the outliers: superconductors with unusually high transition temperatures, or materials that are unexpectedly hard. These exceptional cases often reveal new physical effects or new combinations that, as Schrier states, can “defy expectations.” Historically they were found by accident. The challenge now is to guide discovery toward them more deliberately. That requires different ways of using data and models. It means focusing on the edges of the distribution, looking for surprises, and folding those searches into automated workflows so the search can move quickly without losing human judgment.

This perspective carries over to separations. We need conditions that stand out, not just small tweaks on what already exists. AI agents, when paired with good data and careful experiments, give scientists a way to search more widely, systematically, and more creatively for those standout solutions.

Our visit to the lab at Fordham

That was the spirit of the discussion in Fordham’s lab. Dr. Zhang showed us how they are using agentic workflows for separations and molecule discovery. They keep their tools light, building directly in Python rather than depending on heavy frameworks like LangChain or LangGraph. The focus is on clarity and flexibility, not on unnecessary complexity. Our conversation was about practical issues: how to keep datasets clean, how to design flows that are reliable and easy to debug, and how to link the computational side to the chemistry in meaningful ways.

It was also a reminder of how interdisciplinary this space is becoming. Chemistry, data science, robotics, and physics all come together. Shared archives like SAFE only exist because groups agreed to pool their data. National labs like Los Alamos are investing in AI-driven workflows. Academic groups like Schrier’s are working at the frontier of computational methods while still staying grounded in real laboratory practice. This mix is what will move the field forward.

Taos Research and next steps

At Taos Research we are pushing on AI and agent tools and architectures to aid this kind of science. Our aim is to make them simple enough for researchers to adopt, but powerful enough to handle the complexity and specificity of modern science workflows. One of the key steps is strengthening connections with academic labs like Professor Schrier’s, national labs like LANL in Los Alamos, and others like the Santa Fe Institute who emphasize interdisciplinary study and complex systems.

I am deeply grateful to Professor Schrier and Dr. Zhang for welcoming us into their lab and for sharing their precious time and perspective. The visit reinforced for me that the future of science, and the applications to everyday life it delivers, will not be driven by one discipline alone. It will require those reviewing and deciding funding to seek multidisciplinary scientific expertise. It will be built by people who open their doors, work across borders and barriers, and search for the exceptional.

Leave a comment

Share

Further reading and links

SAFE: Separation Archive for f-Elements https://safe.lanl.gov/

Augustine, L. J., Yang, P., Schrier, J., et al. “Advancing Rare-Earth (4f) and Actinide (5f) Separation through Machine Learning and Automated High-Throughput Experiments”
https://scholar.google.com/scholar?q=Advancing+Rare-Earth+(4f)+and+Actinide+(5f)+Separation+through+Machine+Learning+and+Automated+High-Throughput+Experiments

Schrier, J., Norquist, A. J., Buonassisi, T., Brgoch, J. “In Pursuit of the Exceptional: Research Directions for Machine Learning in Chemical and Materials Science”
https://scholar.google.com/scholar?q=In+Pursuit+of+the+Exceptional+Research+Directions+for+Machine+Learning+in+Chemical+and+Materials+Science

Schrier, J. Introduction to Computational Physical Chemistry, https://taosresearch.ai/learning/introduction-to-computational-physical-chemistry

Dr. Baosen Zhang’s microblog: https://microblog-baosen.blogspot.com/

Let’s begin with the conclusion, so as you read on you can see where we are headed. I conclude that the recently released HRM model exhibits more complexity and inductive bias than many writers stress. It should point AI research more towards complex adaptive systems, and architectures more like The Society of Mind as conceived by Marvin Minsky, where we understand the brain and the mind as a massively parallel society with thousands if not millions of continuously processing, highly interconnected, and cooperating sub agencies working together.

This also parallels the mind-opening research by Michael Levin at Tufts University in systems and synthetic biology showing that intelligence is not restricted to brains, or neural circuits, or even neurons, but exist at all levels of living organisms. Levin and his team’s marvelous experiments show that intelligence extends from organelles to cells to tissue and organs, and even exists in the basic chemical substrates of life. The quest to understand higher order intelligence should therefore be seen as the quest to understand complex systems where higher order agents and lower level agents coexist at many levels, and where agents at all levels have true choice and agency in how they solve problems.

The HRM is a toy system, built to explore new reasoning modes on fairly controlled problems. On its own it is nowhere near as complex as what Minsky posited for the brain, nor what Levin studies in the lab. But the HRM does hide more complexity and levels of interactions, and more levels of “inductive bias” than many reports acknowledge. The coordination of parts, and multiple levels of design, each contribute to its remarkable efficiency. In a small way it mirrors the elaborate but ultimately brutally efficient designs we see in biology.

HRMs are currently much too specialized to dethrone the Large Language Models that we use every day, but they do point to a need to orient AI research towards understanding systems with multiple levels of hierarchy and complex adaptive and recurrent dynamics. The type of dynamics we know brains and other biological systems possess. HRMs may prove to be the most important small model of the year, and help inspire more research towards more energy efficient AI.

Why is this tiny model capturing people’s imaginations?

Ok, if you are like me, you missed the stock market trade of a lifetime this past January.

DeepSeek R1 came out on January 20th, 2025 and just seven days later on January 27th crashed the market and Nvidia stock, as the media began to speculate how a radically more efficient model from China, one that could match OpenAI’s top model on performance, might affect the GPU capital spending boom that undergirds Nvidia’s valuation and the broader AI trade. There were also plenty of stories about the urgency of keeping pace in the AI race with the PRC. If you had loaded up on Nvidia shorts when you got R1 the week before, you would likely be retired today. I myself tested the model for that full week, but it never even occurred to me to open my brokerage account.

So here we are, just a little over two weeks since a mighty model known as the HRM has burst onto the scene. This model of a mere 27M parameters (from the small island nation of Singapore) has defeated many billion parameter models from the leading labs on the “ARC AGI challenge”, a benchmark created by François Chollet and widely regarded as one of the most serious tests of Artificial General Intelligence. A tiny model like the HRM achieving results like this is indeed headline grabbing, and certainly bodes well for research into smaller and more energy efficient models. You would be forgiven for asking yourself, is now the time to short Nvidia?

Perhaps, and perhaps not. This blog is certainly not the place to look for anything resembling financial advance. But there is a deeper story behind HRMs that I think deserves telling.

The Conventional Story of the HRM

It seems funny to talk about the “conventional story” of HRMs doesn’t it? After all it is a model that just appeared a couple weeks ago! I realize to some of you this will seem like a pretty esoteric and perhaps alien topic. Please bear with me though, there should be insights for all, regardless of how closely you follow AI research.

If like me your feeds are flooded with AI material, you have likely already been bombarded with multiple breathless write-ups and videos going into the HRM paper and architecture. Many proclaim its exciting an revolutionary power. The paper itself is an excellent read, and many of these summaries are in fact quite helpful. I recommend starting with the paper to get a more detailed idea of how the model is put together.

I won’t repeat the full detail from the paper, but the basic story is straightforward: HRMs have two modules, both of which are transformers. One is called the “low-level module” and the other is called the “high-level module.” These two modules are wired together to operate at two different speeds. The low-level module runs more rapidly than the high-level module, which acts more like a supervisor, receiving occasional updates from the low-level module and then computing its own updates which are fed back to the low-level module to guide it. The low-level module again runs for multiple more steps before sending its next update to the high-level module. The whole cycle repeats several times in training, and also at inference time, when the model is used to make a prediction.

The authors explicitly claim inspiration from the brain where slower theta waves operate at 4-8Hz and faster gamma waves operate at 30-100 Hz, analogous to the slower high-level module and the faster low-level module in the HRM. The following diagram from the paper indicates this inspiration and the role of the two modules.

This picture however, and even the full paper, stresses only the two “levels” in the hierarchy of the HRM, and in obscures important elements which may be essential to its performance. Studies are as yet inconclusive, but we are likely to see more research probing the structure and results of the HRM in the weeks and months ahead.

It must be reiterated however that the HRM is in no way a general purpose model, nor in any way a replacement for today’s LLMs. Beyond the ARC AGI challenge, it was trained on two other tasks: finding paths in mazes, and solving Sudoku puzzles. These tasks, like the ARC challenge are known to be particularly difficult for LLMs, which lack recursion and are limited by depth for solving tasks with deep structure and dependencies. The HRM performed very impressively on these tasks, besting the much larger o3-mini-high and Claude 3.7 8K models. The following table from the original paper summarizes the HRMs performance gains across these tasks:

What is not shown here, and must be emphasized is that HRMs are not natural language input / output engines like GPT models, at least as far as they were trained in the paper. In the present incarnation, they certainly cannot take arbitrary prompts from users and return realistic natural language responses, nor generate images, nor do tool calling, nor write code, etc. Simply stated they are much more modest than the foundation models in wide use today. So HRMs are not truly a 27M parameter model about to “dethrone” today’s large foundation models. What they do do, and do impressively, is solve certain kinds of tasks that have proven very challenging for even the largest LLMs. Tasks which large LLMs have made progress on only by spending more and more GPU power on, and which still remain elusive.

More research into what tasks HRMs perform well, and how general the architecture can become is crucial. The headlines the HRM is grabbing are well deserved, but a deeper look reveals more important lessons of the HRM for AI research generally.

A more complex picture of the HRM

A deeper look at the HRM reveals it is not simply a story of two modules, one slow and one fast, working together to find the answer. A recent report from Ndea, the company founded by the creator of the ARC-AGI challenge shows that different parts of the HRM model may contribute to its performance on ARC in different ways. If you have time, the video produced by Ndea “The Surprising Performance Drivers of HRM” is well worth the watch time:

Ndea calls out 5 main factors that make the HRM tick, namely:

1. Iterative refinement: the model produces an answer via multiple recurrent passes not a single forward pass.

2. Hierarchical structure: lower and higher frequency forward passes through the high-level (slow) module and the low-level (fast) modules respectively

3. A learned ‘halt’ signal: leveraging Q-learning a technique in reinforcement learning, this outer “control loop” of the HRM determines how many times the overall cycle of low and high level modules inferencing takes place. This outer loop control loop which in essence determines “how long” the HRM “thinks” is under discussed in many write ups.

4. Data Augmentation: the HRM training pipeline makes multiple permutations of the training data supplied in the ARC challenge to try to improve learning and performance.

5. Task Embeddings: In a fairly non-conventional approach for ARC challenge entries, the HRM uses an embedding of the ARC-AGI task rather than present the ARC puzzle directly to the model. The embedding used is a a hash invented by the paper authors and includes a hash of the input puzzle as well as the code of the data augmentations applied to the puzzle.

The Ndea researchers performed ablation studies (a fancy way of saying they tried “not doing” each of the elements in the paper that make up the HRM one by one) to see how much each factor contributed to the performance of the HRM on the ARC-1 challenge. The results are curious, and reveal fresh insights. They do not reflect exactly the narrative most often promulgated around the HRM. Specifically the Ndea researchers found:

1. Iterative refinement was extremely important to HRM performance, but hierarchy (the two level design) was much less important (though definitely non-zero) to ARC task performance. They found that a single outer refinement loop (rather than a fast level and a slow level alternating) could capture almost as much performance as the two-level HRM in the paper.

2. The data augmentations performed by the HRM authors were critical to performance, but far fewer were needed than the authors used, and the augmentations were more important at training time than at test time.

3. The input embedding strategy appeared to play a quite significant role in HRM performance (although the Ndea authors admit more study as needed as this was based not on controlled ablations, but on discussions with the paper authors describing earlier experiments). The input embedding is in some ways the least discussed part of the HRM. In essence it is another “layer of processing” before the two interacting modules. An analogy is the role of convolutions in CNN architectures which are opinionated “feature extractors” from an input image into a form that deeper networks can then train on. The extra structure (and added inductive bias) provided by such a layer is a key design choice for future research to examine. The fact that it may be a critical one in AGI performance underscores that the HRM is a more complex system than is typically realized.

4. The outer control loop which learns the “halt” signal (the “how long to think” signal) did play a meaningful role in performance, but mostly because more outer-loop iteration was the large driver of performance. In principle the “Q-learned halt” can prevent the loop running longer than needed, and may mimic similar control loops in brains and other biological systems, but the Ndea researchers were able to achieve similar raw performance simply by fixing the number of refinement loops at a sufficiently high number.

For more details please consult the Ndea / ARC report on the HRM here.

Where do these results point?

I certainly encourage you to read the original paper, as well as the Ndea report to draw your own conclusions. I’d love to hear from readers who reach different conclusions from mine! In my view, the importance of the task embeddings to performance, as well as the modest role of the hierarchical layers, and the critical role of recursive refinement at the outer loop, points to a more complex picture than one initially gets from the paper.

The real lesson, I believe, is that more complex adaptive architectures need to be studied. The remarkable performance of HRM on ARC will no doubt stimulate extensive small model research. Particularly important will be to examine a significantly expanded range of tasks, and models that involve HRM style iteration (avoiding the common pitfalls of RNNs like vanishing gradients), and non-deterministic cooperation among elements like input transformation, when to halt reasoning, or learned models for when to employ or trigger alternative reasoning modules or networks. The exact configuration and interplay of cooperating elements (“agencies” in the “society” to use Minsky’s terminology) likely will vary by task and by situation. Probably learned reconfigurations and the learned ability to recruit new agencies at test time to confront novelty will be important.

Recall that the Ndea researchers only did ablations against the ARC AGI benchmark evaluation data, not against other tasks from the paper like maze navigation or puzzles like Sudoku. So their conclusion of the minor role of hierarchy may prove to be true for some tasks but not for others. I suspect the importance of different components of the HRM may well vary by type of task.

For researchers interested in more efficient, more biologically and brain inspired models, the HRM is a welcome advance and offers exciting paths for others to follow!

A scientist is supposed to have a complete and thorough knowledge, at first hand, of some subjects and, therefore, is usually expected not to write on any topic of which he is not a master. This is regarded as a matter of noblesse oblige. For the present purpose I beg to renounce the noblesse, if any, and to be freed of the ensuing obligation. . . . the spread, both in width and depth, of the multifarious branches of knowledge during the last hundred odd years has confronted us with a queer dilemma. We feel clearly that we are only now beginning to acquire reliable material for welding together the sum total of all that is known into a whole; but, on the other hand, it has become next to impossible for a single mind fully to command more than small specialized portion of it.

I can see no other escape from this dilemma (lest our true aim be lost forever) than that some of us should venture to embark on a synthesis of facts and theories, albeit with second-hand and incomplete knowledge of some of them, and at the risk of making fools of ourselves.

                                                                           Erwin Schrödinger, What is Life, 1944

I mentioned in the past we would have a lot more to say about the tag line of Taos Research: “AI Accelerated Interdisciplinary Science” and I wanted to get that process started. At a very basic level, this is the motivating idea that led to the creation of Taos Research. Many of our most important problems that we face as society are very clearly going to require expertise and participation from many different areas to tackle, and yet research, business, engineering and technology are all marching us towards greater and greater specialization … towards deeper and deeper silos. The pace of the advancements in today’s world in so many areas makes it challenging to keep pace with developments inside one’s own area of specialization. This leaves scant time to branch out and consider developments in either adjacent or father away domains.

The goal we are setting for Taos Research therefore is to build an organization (and tools and technologies) that leverages AI to push against these silos, an organization that helps humans to both see the connections between rapidly moving fields, and build the meta insights that can spawn the right projects capable of impacting our most challenging problems.

Schrödinger’s classic and inspirational work in 1944 sets the stage perfectly. For those who have not read it, it is in my mind one of the most inspiring books of the last 100 years. Actually it is three very short books in one volume! Each one deeply insightful, personal, and mind opening. All three in fields well outside of theoretical physics where of course Schrödinger earned his well deserved fame. For those interested I strongly encourage reading it in full (here is a link): What is Life? with Mind and Matter and Autobiographical sketches. The number of authors who have referenced Schrödinger’s book since is testament not only to its ability to inspire but also its enduring scientific and philosophical contribution.

Are we fools to go here?

Let’s be clear, as Schrödinger notes, scientists and non scientists alike do risk “making fools” of ourselves from time to time when we wade into ponds away from the places where we are specialists. Erwin Schrödinger was certainly no fool, but he is is equally not an easy act to follow! How can mere mortals follow in his footsteps? If you do take part of your time and devote it to going deeper into fields outside of your core domain, you will occasionally be rebuffed, you may feel overwhelmed by the depth of fields you haven’t touched since high school. Not every step of the journey will be easy or without criticism. How can Taos Research and AI potentially contribute to making this less daunting and more doable?

Let’s tackle this in parts. One step is organizational at Taos as I mentioned. We are building a loose affiliation of researchers from many different fields. If you are interested in being an affiliated researcher, or engineer or project manager and working with Taos Research please reach out to us.

Since we tackle projects in a wide variety of domains of science, business, and technology we rely heavily on our affiliated researchers. Please let us know both where you have expertise, but also where your interests and passions lie. We do not require you to be a PhD or even to hold any particular credentials. We take on both for profit and not for profit projects in roughly equal proportions. A common theme is the use of frontier AI techniques, and the multiple domain perspectives, to accelerate progress on hard problems. Our main ask it that you be deeply curious, committed to learn, and committed to our shared human need to solve problems.

A second crucial strategy is education. We will be publishing both here and in other forms (research papers, open source, books, etc.) works of different forms aimed at helping to bridge domains, and bring high quality understanding in an accessible way to as many people as possible. We equally invite your ideas for what areas you believe warrant focus as we go? And we welcome participation both on this blog as well as on the other forms we may publish. The learning section of the Taos website is an excellent place to start. We are very open to suggestions of great books or papers to add there as well, and even happier if you care to write a review of one that has inspired you.

AI helps in both of these dimensions. We are actively building platform technologies and tools to help in both of these parts of the problem. If you are a builder or AI engineer we welcome your contributions as well. If you are interested in any of these ways of contributing please use the ‘Join Us’ section on the Taos Research website https://taosresearch.ai/ .

I hope we all can commit to be open to the possibility that by venturing into domains were we are not experts, we can accept when any of us occasionally “makes a fool of ourselves” as I have certainly done at times. I liken it to learning a foreign language, when you are learning to speak you have to make many many mistakes. You often feel a stranger in a strange land. Native speakers can find your stilted speech a bit tiresome to say the least. But you have to press on or you will never learn. So let us all also be welcoming in our areas of expertise when new people come with genuine interest, and desire to solve problems, but who lack our language, formalisms, and culture.

What I can guarantee to those brave enough to take the risks, the rewards of continuously reaching deep into new areas are immense. You meet tremendous people. You learn amazing things that most people will never encounter, and you connect the dots in powerful ways to your own life and goals. Staying in your lane is very overrated.

.

I want to introduce to you Taos Research Corporation (taosresearch.ai).

First why the name Taos? I grew up in the beautiful state of New Mexico, and have in my veins a deep connection to the land and wonderful mix of cultures found there. Taos is a town of under 7,000 people in northern New Mexico with an almost mythical spirituality and beauty, not to mention one of the best ski areas in the world. While our family’s home was in Albuquerque, I spent a lot of time in Taos and other parts of northern New Mexico growing up. It is in my soul.

The goal of Taos Research is to focus on cross disciplinary research in AI, and the basic sciences (including both hard and social sciences). Inter and cross disciplinary thinking is really at the absolute core of what motivated the founding of a new organization. The spiritual and philosophical meaning of a place like Taos to me personally serves to emphasize my belief that we need more philosophy, and more connection to the awesome beauty and inspiring complexity of nature in our research. Taos helps sum that up for me. It keeps me grounded to think of the example of the beauty of Taos and it’s broader meaning.

Today research is often extremely specialized, and it is very rare for researchers in corporations or in academia to have time to reach out to other disciplines. Often the challenge just to keep pace with one’s own sub area of specialization is overwhelming. Many of our largest challenges today in basic sciences, as well as in society stem from large complex adaptive systems. The Santa Fe Institute (www.santafe.edu) has been a pioneer in advancing complexity science, the study of complex adaptive systems. Understanding these systems will be essential to addressing the challenges humanity faces on the road ahead, and I believe to do so will require radical new efforts to increase our cross disciplinary understanding… not just to stay in our subfields and make progress there. I like to think of Taos Research as a new sibling organization to the Santa Fe institute, and over time I may try to build greater ties there.

I would like Taos to be an information source, and an organization, that helps people in multiple ares of inquiry to see connections in other fields. On a human level, I’ll also be working to make connections between people whether for business or for intellectual inquiry. In an age of AI, nothing is more sacred and valuable than our human connections.

Taos is also a for profit corporation. Concretely we develop and contribute to platform software and tools in the open source that we believe can help advance science and human understanding. We also plan to devote approximately 50% of our time annually to work with researchers in non profit areas in academia, policy, or other organizations. On the for profit side we work with venture capitalists, large organizations and start-ups on cutting edge research and projects that involve a heavy dose of AI as well as other disruptive technologies.

Please join us as we will try to provide insights and cross disciplinary connections not found elsewhere! On our website we also aggregate publications, papers and books that we have found especially valuable. If you are interested in working with us as an independent researcher or on commercial projects, or if you are interested in engaging us professionally, please reach out.