Kayleigh’s work was featured in a previous installment of the podcast. She writes My Field Notes, a newsletter about user experience research, creativity, and the messy, interesting business of being human, especially inside tech.

She spent 10 years as a UX researcher and has a master’s in human-computer interaction and two bachelor’s degrees, in history and anthropology.

Kayleigh has been studying why people do what they do for a long time. First it was empires and rituals. Now it’s why someone abandons a checkout flow at the last screen.

In our discussion we talk about what Star Trek: The Next Generation teaches us about how to use technology. We also focus on the importance of practicing good judgement, which includes verifying outputs carefully, when using AI.

Today, The Information reported that a number of large technology companies, including Palantir, Booz Allen and Nvidia, are reconsidering their use of frontier AI models from OpenAI and Anthropic. The reason? Concerns over data security and privacy and whether AI labs are “misusing their intellectual property.”

People have always had privacy and IP concerns about using Claude, Codex, Gemini and other AI models. Do you own your work product if Claude helped create it? What about your proprietary data and methods? Would data be retained or used to train models, or even develop competing products and services?

AI labs’ terms of service state that you own your IP. Model training and data retention is a little more complicated. Certain ChatGPT plan types turn on data retention and model training by default. Business accounts generally have strict data retention and training prohibitions.

These assurances helped build trust that organizations, teams and individuals could use frontier models without worrying about these issues. Recent developments are starting to change people’s minds.

OpenAI Lawsuits Open Up User Chat Logs to Retention and Scrutiny

OpenAI is currently facing multiple lawsuits over how the company handles user data. Importantly, some of this litigation is requiring the company to preserve user chat logs, regardless of its normal data retention policy.

This means that user’s redacted chat logs, which includes highly sensitive information, such as health details, company strategic planning and other topics can be requested by litigants and defendants. Some of this de-identified information may be referenced in court filings, transcripts and other materials. Another consequence is that content “typed into ChatGPT could surface in unrelated litigation [such as divorce] cases, employment disputes and contract lawsuits.”

The Mathematics Controversy That is Raising Eyebrows

Another driving force in data privacy concerns is the recent controversy over a mathematical problem (Navier–Stokes) co-solved by Tristan Buckmaster. During the year-long process of working on the solution, they used Claude, Codex, GPT-5.6 Sol and Astra. The work is an example of how an LLM can be used to speed up progress in mathematics significantly.

OpenAI received a rumor that someone had solved the problem. 88 hours later, OpenAI had solved it as well.

Buckmaster was concerned OpenAI had used his prompts to solve the problem and asked: “whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project.” He was told “the model did not look up user data.”

Later in a statement, OpenAI said that it “cannot rule out that de-identified data derived from the usage of our products helped improved our models.”

The Data Sovereignty Imperative

It is still unclear what happened here, but these events have increased concerns about sharing proprietary data with Open AI and other labs. Now organizations are rethinking their willingness to use these models, as indicated by The Information report.

Another driver of this trend is model costs. While AI inference is still heavily subsidized, the latest models are much more expensive. And, organizations are leaning away from using them due to cost concerns. This is partly because less expensive models are more than capable of assisting with many tasks.

These are the reasons why I believe we are going to start hearing a lot more about companies pushing to secure private inference. These can be models run in data centers on their own rented/owned machines, cloud instances with strong zero data retention policies, or even spending on ‘AI computers’ that can be run on-premises.

Traditionally organizations, especially in sensitive industries like healthcare, financial services and security, have had tight data access and sharing policies. It’s highly unusual for a range of companies to voluntarily share their most sensitive data with third parties.

As access to powerful AI inference becomes more democratized, we’ll see a return to form with data/IP sovereignty, protection and access being emphasized.

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Ward layer, which gives you knowledge, tools and software to guard against AI privacy and security risks.

Additional resources in the Ward layer include the AI Security Action Pack, which features 15 in-depth guides on AI security.

Subscribe now

Today’s episode features Dr. Sean Lawson.

Dr. Lawson writes AI with Footnotes, a skeptical enthusiast’s guide to artificial intelligence. What is a skeptical enthusiast? This is a person who is curious, but cautious about AI, acquires hands-on realism (by digging into the details around research, analysis and tools), and has ethical and societal awareness about AI’s potential benefits and risks.

Dr. Lawson is a professor at the University of Utah. His work focuses on the intersections of technology, society and security.

In his post, Everyone Predicted the Wrong 9/11. They’re About to Do It Again With AI, he cautions against techno-doom narratives that will cause us to make the same mistakes we made with 9/11: missing the risks that are right in front of our noses.

In this episode I read excerpts from his essay that I found particularly relevant.

Music Credit: The Cverse Revealed, Composed by Fard Johnmar, 2009

The podcast Is part of the Doing AI Efficiently Operating System.

This is a five-layer platform delivering education, trend analysis, software, frameworks and courses to improve your understanding of, and ability to execute well in AI.

Subscribe now

P.S.: For more on realism, anti-panic and the AI risks you should be paying attention to right now, read my post: The ‘AI Will Kill Us’ Post Got 148 Million Views, But It’s Not the AI Risk You Should Worry About

In less than 48 hours, a resignation notice published on X by a former Anthropic researcher, Jacob Coxon, has received more than 148 million views.

The post features a dire warning: “[T]he people building AI earnestly believe that it could kill us all by the end of the decade.”

A little over an hour after Coxon’s post appeared, a current Anthropic safety executive, Evan Hubinger, provided this response: “Jacob is correct here--we really do earnestly believe AI could kill all humans! I personally think it is >10% within the next decade.”

This is not the first time this has happened. In February, another Anthropic researcher, Mrinank Sharma, wrote: “The world is in peril. And not just from Al, or bioweapons, but from a whole series of interconnected crises unfolding in this very moment.”

Is Civilization At Risk?

A reasonable question to ask is: Is civilization actually at risk from AI right now? Anthropic’s answer: The risk is low but accelerating.

In his post, Hubinger linked out to Anthropic’s latest Risk Report, a 180+ page document that looked at three categories of risk:

• “An AI model with access to powerful affordances within an organization could use its affordances to autonomously exploit, manipulate, or tamper

  with that organization’s systems or decision-making”: Anthropic ranks the risk of its own models doing this as low. However, the recent Hugging Face attack, where 100s of agents colluded, broke out of containment and hacked Hugging Face’s internal systems, suggests this risk is elevated.

• “Highly capable AI models may be able to perform automated research and development (R&D) that rapidly accelerates progress in technical fields. If under human control, this acceleration could disrupt the balance of power both within and between nations; if combined with dangerous autonomous goals from AI, this could lead to catastrophic harms initiated by AI systems themselves.”: Anthropic says the risk is low. However, it admits that it can no longer reliably measure this risk because its evaluations cannot “capture increases in models’ capabilities” in this area.

• “Individuals or small groups with limited resources use AI models to gain access to non-novel chemical or biological (CB) weapons.” Anthropic says the risk is low but has increased since its last report. One worry is that open source models with similar capabilities might aid the development of bioweapons. In addition, the creation of synthetic viruses with AI assistance is raising further concerns in this area.

What Should We Be Worried About?

So Anthropic is saying we shouldn’t be overly worried that AI will end civilization -- yet. The real threat is self-improving super intelligence that evolves rapidly and cannot be monitored.

We may be inching closer to this threshold. Open AI’s recently released Astra model is one example. Traditionally, models have explicitly generated thinking Chain of Thought (CoT) tokens (e.g., “Let us breakdown this problem into steps”) that can be observed to understand the model’s thinking process.

AI Agents We Can’t Monitor

Astra uses a different technique where it generates tokens inside a hidden latent space. Importantly, this may mean that some of the model’s thinking process is also hidden. From a safety perspective, this matters because if we cannot observe how LLMs are making decisions, it is harder to understand what they might do, and determine whether training efforts are properly aligning the model from a safety perspective.

In the Astra safety overview, OpenAI says “We have found that GPT‑6 Astra is more capable of controlling its own CoT than GPT‑5.6 Sol, and less likely to include incriminating information in its CoT ... These findings indicate that the Astra class models could evade our CoT monitors under adversarial conditions.” Translation: It’s harder to monitor Astra, in certain situations. If we can’t monitor the model, how can we verify it’s not going to act against our self-interest?

Overall, the risk that AI can currently give anyone the ability to manufacture bioweapons, hack into systems at will or take over a country’s nuclear arsenal is low. That’s great. The time to take these issues seriously is yesterday.

But the immediate threats to your personal and professional well-being from AI are much higher, and are already present. Here are three risks you should pay attention to, along with practical guidance you can use to protect yourself.

Three Personal and Professional AI Threats: What They Are and How to Stop Them

Data Privacy: AI Labs Have Your Data, What Will They Do With It?

Risk Threat Level: Elevated

AI systems already have access to huge amounts of personal and professional data. In some cases that information is being shared with third parties or analyzed internally for a variety of purposes. AI’s data footprint is increasing and autonomous agents (and even AI labs) are starting to act on this information. AI labs won’t steal your data, but it may be used to train models.

Recently, OpenAI launched software that lets ChatGPT access a user’s messages. ChatGPT can find information about meetings, family events, or even remind you about urgent messaged. It’s a potentially helpful feature that Proton calls a “privacy backdoor.”

Specifically iMessage “uses end-to-end encryption, which means only the sender and recipient can read the contents of a conversation ... The Apple Messages plugin for ChatGPT changes [this] ... If you authorize ChatGPT to scan iMessage conversations, those contents enter the AI pipeline and are treated like any other conversations in accordance with the ChatGPT privacy policy.”

That means any messages ChatGPT accesses could be sent to third parties (such as law enforcement) or be used to train OpenAI’s models.

The most important step you can take to protect yourself is to think carefully about giving ChatGPT, Claude and other AI labs blanket access to your computer files, financial records and other sensitive information on your computer or other devices.

Currently, there’s a brewing controversy associated with the recent solving of a long-standing math problem. The mathematician who solved the problem, Tristan Buckmaster, has accused OpenAI of using prompts he sent to Codex to train the model it used to develop its own mathematical proof. OpenAI said: ““While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models⁠. However, our proofs differ significantly and even the precise results proved are different in the Euler case (forced vs unforced).”

The lesson: Consider any information you send to a non-local model as not private by default and accessible by third parties.

Subscribe now

Digital Security: AI is Helping Hackers Move Faster

Risk Threat Level: Elevated

AI is accelerating the use of deep fakes, no-click malware that can infect mobile devices instantly, and efforts to infiltrate systems that process your payments, or the financial infrastructure your business depends on. One example is the VoidLink malware targeting Linux systems, which is “one of the first instances of an advanced malware largely generated by AI.”

In 2024, “a finance worker at engineering firm Arup authorized $25.6 million in wire transfers to accounts controlled entirely by AI-generated impersonators; every person on the video call was a deepfake.”

It’s easy to get spooked by growing digital threats. Are risks, like an AI hacking your smart fridge, remote? Likely yes. The more serious risks are to your personal digital safety. The best protection is to:

• Keep your devices up-to-date, as security patches are being made more frequently due to the pace new vulnerabilities are being discovered

• Use VPNs and other tools to protect your Web traffic and potentially reduce the odds you visit a site infected with malware

• If you’re using AI agents regularly on your devices, get educated about agentic security threats and how to protect yourself (I developed a free resource, the AI Security Action Pack), which has lots of education and strategies to harden your agents and devices

Off-Skilling: AI is Taking Over Writing, Strategy and Research and the Costs Are Mounting

Risk Threat Level: Low to Moderate

As models become more capable, AI can take over an increasing percentage of high-value cognitive work. The jury is still out on whether AI will take your job.

What’s less debatable is that the nature of work has already transformed. There’s early evidence that off-skilling is occurring at an increased pace, where workers are losing their ability to reliably assess outputs, or even produce routine work. You’ve likely seen this everywhere from professional emails to presentations and reports produced with AI assistance, but contain limited reasoning and logic.

To counter the off-skilling problem, companies are taking action:

• Ernst and Young is giving $100 million in bonuses to staff demonstrating critical thinking, adaptability and innovation

• Haru matcha has banned the use of AI across all its marketing. The company’s CEO Francis Raho-Jeavons said it was to get “back to content feeling authentic, transparent and human.”

Combating off-skilling is complex, and a major goal of my Doing AI Efficiently Operating System. The best offense is a good defense:

• Understand the cognitive and professional risks posed by AI over-reliance

• Engage in activities that actively cultivate your critical thinking and creative skills without (or with limited) AI assistance

• Gain an understanding of how generative AI works at a fundamental level to better understand its strengths and weaknesses

• Regularly train your instinct and judgement, learning how to evaluate AI outputs for correctness, reasonableness and logic

Subscribe now

Pay Attention to AI Risks, But Focus on the Ones That Matter Most

What’s more concrete is the AI-powered plugin that reads your emails and has access to your financial information. Or the AI-developed malware that infects your phone before Apple can release a security patch. Or the AI-caused slow erosion of skills you’ve spent years and decades cultivating.

You have power and agency to do something about your personal risk profile. All it requires is knowledge, awareness and taking small, practical steps to protect your security, privacy and mind.

Staying alert to the potential existential risks posed by AI is justified. Panic is not.

“It’s about the people, stupid.”

I’ve lived by these five simple words throughout my 20+ year career helping people thrive during periods of rapid technological change.

Why? Because everything is so big and fast in the emerging technology space that it’s easy to forget people.

The financial numbers are huge. The pace of product development is relentless. Organizations are struggling to figure out how to adapt in real time. People are being asked to operate outside their comfort zone.

When technology is driving rapid change, I’ve learned people benefit most from first principles thinking, a consistent strategy, and foundational skills. These provide you with power and agency that no technology can take away.

I launched this newsletter, and other parts of the Doing AI Efficiently Operating System, to give you this power.

What You Won’t Find Here

The AI education and implementation space is loud and crowded right now. Many people are delivering:

• Information about how to use the latest large language models such as Fable, Astra and their siblings

• Articles describing the latest agentic workflows such as loop and graph engineering

• Tutorials about Claude Cowork, ChatGPT and other interface systems released by frontier labs

These topics are well-covered elsewhere. While information about models, agentic workflows and LLM user interfaces like Claude Cowork will be referenced, they won’t be the main focus of OS outputs.

What You Will Find Here

I developed the Doing AI Efficiently Operating System to help you (as an individual, part of a team or organization) gain foundational, first-principles-based knowledge and skills that will aid your success. Here’s how each part of the OS works together to achieve this goal.

Subscribe now

Grasp: Gain Knowledge

It all starts with having a solid understanding of AI. Not just generative AI, but other technologies such as machine learning, world models and robotics. Learning about how the market is evolving and what to expect in the near future is also important.

This information will help you separate hype from reality and be less surprised by rapid developments in the field. The Doing AI Efficiently OS delivers insights across all these areas.

In addition to the educational content published in this newsletter, you also have access to the:

• Thinkers on AI Writer’s directory, featuring emerging and established writers publishing content on AI. The directory is refreshed every 36 hours.

• Growing AI terms glossary

Coming Soon: The Thinkers on AI Writer’s Directory

Coming Soon: The AI Glossary

Subscribe now

Discern: Practice Judgement

One of the most important roles people will have is judging whether AI outputs are correct, valid and appropriate. Judgement is required in diverse areas such as checking internal documentation and writing using AI.

You’ll find newsletter articles providing insights about judgement-related topics such as the risks of using AI for writing, and how to conduct high-confidence research in the age of AI-produced data and information.

The OS also features practical education that will develop your judgement (see below).

Ward: Guard Against AI Risks

Generative AI is powerful, but also risky. The AI Security Guard platform, which I’ve been developing since February 2026, features many resources, including the popular free AI Security Action Pack.

The newsletter will feature key AI security insights along with links to security-related content and software from AI Security Guard.

Coming Soon: AgentGuard360

AgentGuard360 is an upcoming AI security platform that prevents malicious software from being installed on your device, locks down device security gaps, and blocks harmful content before it reaches your agent. It includes supply chain, device security, content scanning, runtime protection, and cost control.

Execute: Use AI Excellently

The OS features many AI execution-related resources, including newsletter articles and experiential learning experiences.

Coming Soon Premium Doing AI Efficiently Interactive Courses

The OS will have a library of interactive courses, including:

• Beyond the Em Dash: How to Edit AI Content Effectively

• Lies, Damned Lies, and AI Slop: How to Conduct AI Research Without Getting Burned

• The Penny Pincher’s Guide to AI Maxing: Get More AI for Less

• Beyond Prompts: How Building a Harness Can Make Your AI 100% More Useful

Coming Soon: Premium Clankyopolis AI Skill-Building Game

The OS’s flagship educational experience will be Clankyopolis. This game will put players in the role of Administrator of Clankyopolis, a fictional robot city.

As Administrator, you’ll practice key AI skills such as:

• Evaluating AI outputs

• Conducting data and informational AI quality control

• Implementing agentic workflows, from selecting models to evaluating their performance over time

The game will enable you to deepen your understanding of AI implementation in a safe, fun environment.

Honor: Taking Responsibility for AI Outputs

An important step toward mastering AI is being able to take responsibility for the quality and robustness of AI outputs. Every resource, training solution and product in the OS supports this objective.

What to Do Next

• Subscribe: In addition to receiving research, analysis, tips, tools and other free resources, subscribers receive discounts on OS educational experiences

• Spread the word: If you’re already a subscriber, thank you! Please help this community grow by spreading the word

• Contact me with questions: If you have questions about the OS, how to use AI effectively, send me a message. I read every message

Thanks for reading.

Fard

Welcome to the first installment of the AI Research Beat. This is a regular series where I highlight an interesting study that sheds light on how AI is impacting work.

Writing is tedious, horrifying drudgery. It’s also rewarding and gratifying. But, you have to get past the hard parts first.

Is it any wonder that large language models (LLMs) have been welcomed by writers all over the world? Staring at a blank page? Ask Claude to help you get started. Struggling over word choice? ChatGPT is there to lend a helping hand.

I’m not against people turning to AI for writing help. Some may not be fluent English speakers. Others may have limited writing experience, or simply need a lot of assistance.

Then there’s using AI writing out of necessity. For example, I’m creating an AI educational skill building game called Clankyopolis. Players will learn by working through customized scenarios, generated on-demand. There’s no way I can write every word of the game’s copy. LLMs to the rescue.

But, there’s a problem with uncritically using LLMs for writing. Your unique voice is snuffed out. People complain your writing is flat. They’re right.

According to a new paper, “The Shrinking Landscape of Linguistic Diversity in the Age of Large Language Models,” by Zhivar Sourati of the University of Southern California and colleagues:

• Language provides valuable insights about people’s mental health, behaviors, and health status

• LLMs are eliminating this rich linguistic intelligence

• Generative AI models “homogenize writing styles,” making content less diverse and interesting

• When LLMs are used to '“polish and rewrite texts,” textual diversity and uniqueness decreases (translation: it’s boring)

• LLM text sounds like it was written by older liberal males; unique ethnic or cultural language traits are erased

As shown in the infographic below, since the advent of ChatGPT, writing complexity and diversity across ArXiv, Path Notes and Reddit declined sharply.

(Thanks to Sekoul Krastev of The Decision Lab for highlighting this study on LinkedIn.)

It’s All in the Edit

Generative AI is making our writing less diverse and terribly boring. What can we do about it? The answer isn’t shame or ridicule. People aren’t going to stop using AI to help them write.

Instead, people need to learn how to edit their AI-assisted content. Proper editing can retain a writer’s unique voice while increasing content accuracy and impact.

It’s popular for people to use ‘humanizing’ skills to remove obvious AI tropes from writing. This isn’t enough. The problems with AI-generated prose go way beyond removing em dashes and “it’s not X, but Y” phrasing. Improving AI writing requires a human touch.

I’m currently working on an online course, Beyond the Em Dash: How to Edit AI Writing Effectively. My course will feature guidance on how to deal with common AI writing weaknesses. Here are five tips from the course you can put to work right away.

• AI writing is wordy: Get to the point. Cut unnecessary fluff.

• AI copy is weak: Saying “nobody does X” is a cop out. Write with power. Use percentages, or specifics.

• AI copy is plagued by throat clearers: Phrases like “it seems like” add nothing. Delete them.

• AI copy is fearful: Descriptive, potent writing holds the reader’s attention and doesn’t waste their time.

AI-assisted writing is making everyone sound the same. Writing with purpose, power and precision (and having good editing skills) will put you ahead of the pack.

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Discern layer, which focuses on helping you develop and practice good judgement when using AI.

Today’s episode features the writing of Kayleigh Mann.

She writes My Field Notes, a “newsletter about user experience research, creativity, and the messy, interesting business of being human, especially inside tech.”

She spent “10 years as a UX researcher and has a master’s in human-computer interaction and two bachelor’s degrees, in history and anthropology.”

Kayleigh has been “studying why people do what they do for a long time. First it was empires and rituals. Now it’s why someone abandons a checkout flow at the last screen.”

I’m a big fan of Star Trek the Next Generation. It’s one of my favorite shows.

They have powerful all-knowing computers, but somehow always foreground the human.

In her post, “Computer?” What I Think Star Trek Got Right About AI, she shares lessons on augmentation, trust and interface design.

In this episode I read excerpts from her essay that I found particularly interesting.

Music Credit: The Cverse Revealed, Fard Johnmar

The podcast Is part of the Doing AI Efficiently Operating System.

This is a five-layer platform delivering education, trend analysis, software, frameworks and courses to improve your understanding of, and ability to execute well in AI.

Does using calculators make us worse at math? This was the question being asked in the 1980s during the Great Calculator Debate.

This was a fierce argument about whether letting children use calculators in elementary school was a bad idea. Opponents of the practice suggested doing so would lead to the “destruction of student math skills.”.

The pro-calculator forces won the debate, but the conversation about whether using computers erodes cognitive capabilities continued over the decades. When GPS gained widespread adoption, people lost their ability to navigate using maps, or even using landmarks. Google use has been associated with less ability to remember or retain facts.

Conversation about the cognitive impacts of computer use has intensified with the rise of AI in the form of large language models (LLMs). What makes LLMs different is their reasoning capabilities. They can debate, summarize, restructure and write for us. And, like other useful technologies, people are adopting AI to help with a cognitively challenging tasks such as writing. So, does AI use positively or negatively impact our skills and cognitive abilities?

Recent analysis by Mark Keith of BYU suggests the answer is nuanced. Over the short-term there is a real benefit to AI use (when compared with non-users). AI users out-performed non-AI users in:

• Productivity: 72% increase in task completion speed

• Output quality: 67% increase in content output quality

• Creativity: Writers given access to generative AI for story ideas created better stories (but story idea diversity declined)

• Learning: 75% improvement in learning activity (during AI-delivered lessons)

However, these benefits don’t persist over the long term:

• Skills gained during AI-delivered lessons decline

• Many people forget what they learned when AI assistance is removed

• People experience reductions in critical thinking skills, mental effort and learning retention

Keith suggests there’s a real gap emerging between those who can use AI and retain critical thinking skills and those who do not. In fact, he suggests using AI the wrong way could leave AI users worse off in key ways than those who never started using AI in the first place.

I agree.

How to Retain Cognitive Skills While Using AI

Many of the ways people tend to use AI are sub-optimal for retaining critical thinking and other skills. Keith’s research and my experience in this are suggests the best ways to thrive long-term while using generative AI is to:

• Break the Acceptance Pattern: Resist the urge to ask AI for an answer and accept it on face value

• Be a Critical Consumer: When AI provides any output, especially if it’s high-stakes, critically review it for correctness, validity and rigor before accepting it

• Ask AI to Push Back: This is especially important in engineering tasks where there might be a better and more elegant way to solve a problem

• Keep your creative edge: Remember that AI tends to produce outputs that are at the average rather than out-of-the-box. Spend time coming up with concepts, and ideas without AI assistance.

Practicing thee habits over the long-term can be difficult. Extended AI use, especially if the results are generally of good quality, breeds complacency. But, it’s the practice of remembering that generative AI outputs can be of wildly uneven quality. Doing so will keep you on your toes. Just because an answer is correct in one context does not mean it’s valid in another.

Resource - How to Retain Cognitive Skills While Using AI

7.89MB ∙ PDF file

Download

Infographic outlining how critical cognitive skills are lost using AI, and how to retain them.

Download

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Execute layer, which is focused on helping you use AI solutions better over the short- and long-term using education, tools and experiential learning.

Subscribe now

I have been on Substack for years, but only recently started to devote more time to the platform. Like many of you, I’m keenly interested in learning how to reach my audience and grow on Substack. So, I began by analyzing dozens of Substack accounts for clues.

Very quickly, I noticed something strange. Some newsletters were growing very quickly, gaining 1,000s of subscribers in 30 days. Other people have been on the platform publishing regularly for months, but have seen very slow or limited newsletter growth.

I wondered why so I started reading advice from others (mainly marketers) about how to grow on Substack. They often pointed to their own subscriber acquisition success as evidence their recommended growth strategies were effective.

I found their takes interesting and useful, but what they all had in common was that they were all in the ‘how to market on Substack’ niche. There’s a natural audience for this type of content on Substack. Everyone here wants attention, so it’s not surprising many in that area are able to grow their accounts quickly.

I’m not in the Substack marketing niche, so I looked for other examples of high-growth accounts. My survey wasn’t scientific, but I noticed that people in other niches on Substack had been on the platform for quite some time, often years. When these people post notes, they get lots of engagement (because their audience is larger) and seem to benefit from a growth flywheel.

However, these people aren’t the norm. Most have much lower subscriber counts. This is not necessarily a bad thing. Many people are not on Substack to grow exponentially, but to find a place where they can simply share their perspectives and find a community.

But, if we’re being honest, many people on Substack would like to earn from their writing. Earning requires growing an audience. That’s the promise of Substack after all, right? So I decided to dig deeper into the Substack recommendation engine to learn more about how it operates, surfaces content and promotes account growth.

As an aside, I really wanted to find out why those ‘Do Your Thing Substack’ posts seem to work so well at attracting subscribers. I mean these posts where people ask others to connect around their topic can generate hundreds of comments and thousands of likes almost immediately. That type of post might not be your cup of tea, but you may be curious about why they appear to be effective.

My goal isn’t to give you a “proven blueprint” to go from 0 to 1,000 subscribers in 30 days. Instead, the purpose of the Doing AI Efficiently Operating System (this newsletter is at the system’s center) is to provide you with first principles insights and education on AI so that you can use it more effectively.

At its core the Substack recommendation engine is AI (my acronym for it is SAIE). I want to help you understand SAIE so you can improve your approach to the platform and attract more people to your insights and writing.

There are no guaranteed shortcuts to gaining an audience on Substack. However, armed with the right knowledge, you’ll be able to better calibrate your activities on the platform to set yourself up for success more effectively.

Contents: A Guide to the Substack AI Recommendation Engine

• Four Questions I’ll Answer About the Algorithm

  • What does the recommendation engine do?

  • How Does the Substack AI Engine (SAIE) Work?

  • Why Is SAIE Pumping Harder for Some People Versus Others (And Does It Matter)?

  • Why the House Always Wins

• Strategies for Success

• Infographic: How the Substack Recommendation Engine Works

Four Questions I’ll Answer

Fortunately, Substack has provided all the information we need to understand how its AI recommendation engine works. There are other guides to the Substack Notes algorithm. However, the popular ones I’ve read aren’t grounded in the technical details of the system. There’s a lot to be learned once you read research papers highlighted by the Substack team and the team’s statements about SAIE.

I’ll provide you with an overview of the system’s mechanics, as I understand them, at a level appropriate for non-technical readers. Note that I haven’t spoken to Substack about this essay, and I don’t claim to have insider knowledge of how the system works. My analysis is based solely on review of publicly available information.

I’ll answer four questions in this essay:

1. What is the Substack recommendation engine designed to do?

2. How does SAIE work?

3. Why is SAIE pumping harder for some people versus others (and does it matter)?

4. Why the house always wins: Substack is optimized for subscriber acquisition, not retention (that’s on you)

1. What is the Substack Recommendation Engine Designed to Do?

Let’s not forget that Substack is a business. Its goal is revenue maximization. That means, like other platforms, it needs people to:

• Discover the platform

• Find and subscribe to publications

• Pay for newsletters (this is how it makes money)

One of the reasons I didn’t try to grow my newsletter on Substack in the past was the discoverability issue. If I set up a newsletter on Substack, I still had to find readers by promoting my content on other channels like LinkedIn. If that were the case, I thought, why not just use LinkedIn, or send people to my own platform? Substack was a hard no for me.

When I came back years later, I had the same attitude. Then I discovered the Substack feed. This changed the game for me because I thought: “Okay, Substack is trying to solve the discoverability issue by providing me with a means of reaching others on the platform.” That seemed worth the effort to me, so I decided to invest time here.

So, I don’t mind that Substack is optimizing for discoverability and paid subscriptions. It’s a win-win, right? Almost.

There’s a catch. The SAIE is not optimized for your success as an individual publisher. Substack wins when publishers, in the aggregate, do well on the platform. Every subscription counts equally. The system doesn’t care whether people subscribe to YOUR publication, only that people subscribe to A publication.

The other things SAIE is optimized for are user data processing, rapid analysis and predictive power. For reasons that I’ll explain later, the system needs a steady stream of platform activity data generated by you. Only show up a couple of times a month to post an article and send out a Note announcing it? Guaranteed audience growth failure mode.

SAIE optimizes for signal accumulation. Every click, like, restack, subscription and payment feeds the system and influences what you’ll see next (and whether your account will grow). I noticed the engine pushing content to me from the beginning. I clicked on some of those “get 1,000 subscribers in 30 days” articles, and that’s all I saw in my feed for a couple of weeks. This changed, but only when I adjusted my reading, following and engagement habits.

SAIE can help you gain subscribers, but retention is on you. Substack only looses if users leave the platform and don’t pay for subscriptions. SAIE is agnostic about whether it’s your content, or someone else’s, that’s recommended.

Understand what SAIE optimizes for and you’ll recognize how to use the system to your maximum advantage.

2. How Does SAIE Work?

SAIE is a complex system. While I didn’t receive a briefing from Substack about its inner workings, I found several valuable sources of information that helped me to learn about SAIE’s technical foundations:

• Mike Cohen, Substack’s Head of AI, wrote this article in October 2025 outlining the system’s technical architecture

• Amanda Bray, of the Publishing Spectrum conducted an interview with Cohen in late July. In the interview Cohen confirmed that it is still using the methods outlined in his October article to run the recommendation engine

• This article, also published in the Substack official newsletter in October 2025, provides specific information about the user data signals Substack is using in the engine

From User Activities to Recommendations

At its core SAIE uses machine learning, specifically sequential models. A sequential model is designed to handle ordered or time series data, where the “order or sequence of the input matters”. Sequential models are used heavily in recommendation engines. For example, in order to predict what content a person might click on next, information is gathered about their previous activities over time. Consider this question: If a female user, age 30, who has used a platform for 6 months, views A, B and C content over a specific time period, what type of information will they be most likely to click on next? A sequential model might answer: “This user will most likely click on an article showing a picture of a dog.”

For Substack a properly trained and calibrated sequential model is valuable for helping to predict what you’ll:

• Read

• Interact with (like, comment, restack, share)

• Subscribe to

• Pay for in the future (the most important success metric)

Have you ever noticed how much data Substack collects about your platform use? In the Substack dashboard can see exactly how many times your Notes were viewed, whether a view resulted in a subscription, how often your readers click on links in your emails and many more pieces of information. This is the type of data that’s fed into SAIE to help it predict the content users will be most interested in viewing, and what should be recommended to them.

Sequential models owe some of their architecture to large language models (LLMs). As I explained in my essay on AI text watermarking, an LLM translates strings of text into numbers so that the model can process it.

Sequential models also translate information. In order for the model to understand user data it must be converted into embeddings (lists of numbers) representing different on-platform events. Embeddings are very important because they allow the model to rapidly reason over numerous parameters and make associations between data types more easily. As Cohen said in his essay describing the system:

“[W]e’ll integrate ... sequential embeddings into ranking ... The sequential approach brings two major advantages: first, the final ordering will be tuned to the flow and momentum of your current session, understanding which posts make sense as the next step in your reading journey. But perhaps more importantly, the user representation itself will be much smarter, incorporating attention mechanisms and 10x more features than before, making the ranker better at determining what posts are good matches for you in general, not just in this moment.”

The types of data used by the predictive model can vary, but may include:

• Where you’re located

• What language you speak

• Which publications you’re subscribed to

• Which creators you follow

• What interests you’ve specified during Substack onboarding

• How often you’ve liked content

• What you’ve restacked

• Sentiment analysis of content you post

Activity Time Periods, Signal Strength and Negative Actions

There are three factors that may have a big impact on SAIE’s recommendations.

Time period: The system considers what you’re interested in reading now and in the past, and, as Cohen notes tracks “the momentum and direction of your current session.” Importantly, the system doesn’t just favor your most recent interactions, but develops a sense of how you use the platform over the long-term. Cohen said:

“[W]e still preserve your long-term reader embedding alongside the sequential state ... So even if you’re deep in a poetry session, the system still knows about your enduring love of sports journalism, your subscriptions to tech newsletters, your history of engaging with climate content. Those long-term signals ensure that other parts of your interest graph stay in circulation.”

Signal Strength: The more information the system has about you, the better it can optimize for your interests and reading habits. At the same time, a rich data set about your writing enables the system to better recommend YOUR content to others.

Negative Actions: Cohen didn’t discuss this in his article (or his other public statements), but it’s also critical to understand that sequential model-powered systems can be designed to track what I call negative actions. These are signals you send to the system when DON’T do certain things.

For example, let’s say the system is recommending a certain type of content, but you don’t interact with it. SAIE learns this content isn’t worth showing to you. This may be why certain Notes get buried, while others, like the infamous ‘Do Your Thing Substack,’ posts are more likley to appear in your feed. You may not like them, but the system has been taught that they get engagement AND drive subscriptions.

Pinterest, which helped to advance the state of the art in sequential models had this to say about using negative actions to fine tune recommendations: “The second, impression-based negative sampling, selects Pins ... that the current user has viewed but not engaged with further, suggesting low interest. Our findings ... reveal that impression-based negative samples are more effective ... [and improves] ranking performance.”

Cataloging Key Substack Signals

I spent a lot of time in the last section discussing signals Substack uses to identify what types of content to show you in its feed (and recommend to others). That’s because it’s the single most important factor that determines your fate on the platform.

Provide the right data and signals to Substack and, given enough time, it can help your newsletter grow.

Below I’ve provided a catalog of the types of signals Subatack is may be feeding into SAIE to help you focus and optimize your activities on the platform.

Remember: It’s about signal strength and time. Think of it in these terms (this isn’t how Substack does it, but it helps explain the concept):

Signal Density is the average volume of topical and engagement activity data you have generated.

Time is how long the system has had to collect data about you, your content and actions.

The more you use Substack, and the more time you spend on the platform, the better the odds the system can:

• Recommend content to you

• Put YOUR content in front of people who might subscribe to your newsletter

Although there may be a negative action penalty, it might be used to help Substack better understand who your ideal reader is over time.

Your Activity Isn’t Wasted Effort

Those Notes you send out that get no reply? They’re not a waste of time. They are valuable because SAIE is able to use this data to better inform its predictions and recommendations. And, your content is used to enrich the feed.

Have you ever seen those Notes published weeks ago in your feed? That’s SAIE at work, using its available inventory of content, whether it was published yesterday, or five months ago, to help keep you engaged.

Below is a listing of signals SAIE may be using to inform content recommendations.

The system is designed to accumulate as much of this data as possible. The more signals you generate, the better SAIE can understand who you are, what you write about, and who might subscribe to your work.

3. Why Is SAIE Pumping Harder for Some People Versus Others (And Does It Matter)?

Let’s address to the elephant in the room: Why are some people gaining subscribers faster than others? There’s no secret formula to success, but understanding SAIE provides some useful insights about what the system favors. And, like all algorithmic systems, it has its weaknesses.

Understanding SAIE Content Bias

Does SAIE censor content, individuals or accounts? Yes. Every recommendation system like the one Substack runs has safety and spam filters.

Given Substack’s investment in Pangram for potential AI content detection, it’s unclear whether content flags are used to de-emphasize certain types of posts. Pangram scanning isn’t activated automatically, so even if is a signal, it’s likely weak and secondary. This could change if automatic scanning is enabled.

Importantly, there may be an inherent biases in the system that can fuel the perception the algorithm is unfair.

First, recommendation systems are known to potentially suffer from “popularity bias.” Specifically in this paper by Anastasiia Klimashevskaia and colleagues the authors suggest: “algorithms may have a tendency to focus on already popular items in their recommendations. As a result, the already popular (“Blockbuster”) items ... receive even more exposure through the recommendations, which can ultimately lead to a feedback loop where the ‘rich get richer’”.

In addition to popularity bias, the recommendation engine may work better for users with a richer data history. In an analysis Shengyu Zhang and colleagues note: “Recommendation performance usually exhibits a long-tail distribution over users — a small portion of head users enjoy much more accurate recommendation services than the others. [There are] two sources of this performance heterogeneity problem: the uneven distribution of historical interactions (a natural source); and the biased training of recommender models (a model source).”

What this research suggests is that:

• Users with a longer account history may have their content recommended by the system more often versus those with a less user data (uneven distribution)

• The model may be trained on data that causes it to favor more popular topics

Other factors that may further bias SAIE include:

• Revenue Optimization: SAIE optimizes for subscription events, not content quality. Content that reliably generates subscriptions gets favored.

• Topic Signal Density: Topics with a lot of data associated with them (clicks, restacks, subscriptions) can deliver higher quality predictions. Are you writing about marketing on Substack? Your personal journey that netted you six figures in subscription volume? These topics are popular and you may benefit from a positive feedback loop of attention and subscriptions.
  
  But, if your topic is more niche, like food safety, the history of hats, or outside of a popular topic area, the system has less information to go on, and may surface your content less (and your growth may be slower)

A biased recommendation engine:

• Cultivates a monoculture that rewards only certain types of writers and topics

• Results in a lack of support and visibility for newer accounts. If the rich get richer affect is pronounced, the Substack user base may become disillusioned, and the community will be less attractive to new users; in this situation, growth stalls

“Hacking” the Algorithm

There are some early signs that some are either purposefully or accidentally taking advantage of some of the weaknesses of recommendation engines. Some have figured out ways to ‘hack’ SAIE to generate high account growth in a short period of time. Two of these strategies are explained in the images below:

• “Do Your Thing Substack” Notes: Discussed previously, this strategy takes advantage of SAIE’s preference for positive content engagement signals to generate outsized visibility for certain accounts. While not all Do Your Thing Substack posts are successful, they reward stye over substance. Fast growing accounts with thin content are outperforming new accounts where authors are developing in-depth analysis, opinion and other rich content.

• “Subscribe for Subscribe”: Users are engaging in mutal subscription exchanges. This strategy directly targets the engine’s primary recommendation driver, subscriptions, to drive engagement and accelerate account growth velocity.

Substack was founded to reward writers seeking a home to share substantive content with audiences. While the SAIE has boosted Substack’s userbase, the engine could also become a major source of user frustration and distrust if it is not successfully spreading the wealth across the userbase, and rewarding high-value, high-signal content creation.

4. Why the House Always Wins

Substack wins when you win. But, SAIE’s goal is to drive revenue for the company, not a specific newsletter.

Here’s the lifecycle the engine may optimize for:

1. New user joins on Substack through a newsletter, referral or other means

2. SAIE shows them relevant content from the feed

3. They subscribe to a publication

4. Eventually, some convert to paid

5. They churn from one publication and discover another via the recommendation engine

6. They subscribe to that publication

7. Repeat

The model treats every subscription event equally. A subscription to Publication A has the same positive signal weight as a subscription to Publication B. (Note: There may be factors that favor publications with bestseller status in terms of author feed visibility that I’m not accounting for here.)

SAIE may be largely agnostic about which publications people subscribe to. As long as a person stays in the Substack ecosystem (and migrates to a revenue-producing publication), that’s a victory for Substack. The house always wins.

Strategies for Success

Volumes have been written about how to gain attention for your work on Substack. A lot of it is very useful. I’ll confine my advice to what this analysis of SAIE suggests.

• The Trend is Your Friend: As you’ve seen, SAIE offers compounding benefits for authors on the platform long-term. As you participate in the network, you gain more subscribers and followers. An increase in attention leads to higher responses to your content, this leads to additional subscriptions and the flywheel continues. SAIE helps increase visibility and subscriber momentum toward your publication. Account longevity and signal density (generated by your actions) counts.

• Consistency is Essential: Being visible on the platform provides SAIE with more signals about you, your interests and the topics you write about. Think of posting Notes, comments, replies and articles as depositing into the SAIE signal bank. The more signal you provide, the better the engine can work for you.

• Quality Over Quantity: It can be tempting to try to juice your subscriber count by ‘hacking SAIE’. If your goal is long-term viability, a loyal readership and sustained revenue, that might not be the way to go.
  
  Letting SAIE help you find your tribe and growing your visibility among people who want to hear from you is always the superior option.

Infographic Resource

Resource: The Substack Recommendation Engine

1.68MB ∙ PDF file

Download

An infographic summarizing how the Substack Recommendation Engine Works

Download

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Grasp layer, which is focused on helping you understand how AI works from first principles, including tokens, context windows, model architecture, how LLM content generation happens, and more.

My ‘holy crap’ moment with generative AI happened in early 2025.

I had spent the better part of a year working on an AI-powered intelligence platform. It had AI-augmented Web search, company competitive and investment data, real-time personalized insights delivery and other powerful features.

I was proud of the product. It featured a lot of bells and whistles that are common today, including retrieval augmented generation (RAG), multi-agent workflows, validation loops, and evaluations. All of these features helped make insights surfaced by the platform more accurate and relevant to users.

But, all it took was one announcement from OpenAI to throw all that work in the trash. That was the day I learned to fear AI.

Building an AI-Fueled Analytics Engine

Let me back up for a second so you understand the full context.

ChatGPT was announced in 2022. Most people were using AI by typing words into a chatbox. But, by early 2023, thousands of people around the world were experimenting with large language models (LLMs). I was one of them.

Using the OpenAI API, we could augment our applications using LLMs. AI-fueled systems were more difficult to build at the time because LLMs (and the API) were much less capable. The APIs didn’t have Web search capabilities, models could only process a limited amount of information, and LLMs made a lot of basic mistakes (hallucinations were a lot worse than).

Many of us builders looked at the limited functionality of these models and thought: AI labs need to do a lot of work to improve their models. They won’t introduce features that will compete with what we’re building.

We were wrong.

AI-Caused Obsolescence

On February 2, 2025, OpenAI announced Deep Research, a “new agentic capability that conducts multi-step research on the internet for complex tasks.” It allowed OpenAI’s models to “autonomously find, analyze, and synthesize hundreds of online sources to create a comprehensive report at the level of a research analyst.”

When I read the announcement, my stomach dropped. I sat in my office feeling numb. OpenAI had just made my product obsolete. Even though my platform had a proprietary database, I knew it couldn’t really compete with ChatGPT’s Deep Research offering.

But there was a deeper issue. I felt obsolete. At that point, I had a 20-year career in innovation consulting. I had written books, developed frameworks, analysis, products and a lot more for organizations around the world.

Now, ChatGPT could replicate all that experience and insight in an instant. It appeared as if I was no longer needed and that scared me.

After Deep Research was announced, my product pitches pretty much went the same way: “What does your product offer that ChatGPT can’t give me?”, they’d ask. I added Deep Research-like functionality into the product, but could never convincingly explain why they should choose what I developed over OpenAI’s offering.

Subscribe now

The Curse of Progressive AI Disempowerment

In the United States and other parts of the world there is a strong moral, emotional and cultural meaning associated with work. In the mid-20th century, there was an emphasis on discipline and company loyalty. The introduction of digital technologies led to the rise of the always-on worker, focused on maximizing productivity. Today, there is a large focus on flexibility and results rather than hours worked.

Throughout all these changes to work culture, one thing remained: work, whether for yourself, or for an organization, provides a strong sense of purpose and mental and emotional stability.

Generative AI is upending the nature of work. AI labs have been working to make their models increasingly capable. They can conduct research, provide analysis, digest large volumes of data, and much more. The mantra has been: AI can replace many of the things people are paid for.

I call this the curse of progressive AI disempowerment. As AI becomes more powerful, many of the tasks once assigned to humans, and the contribute to personal and professional growth and self-esteem, could be eliminated.

(As an aside, I’m well-aware of David Graber’s theory that societies have been creating an increasing amount of ‘bullshit jobs’ that aren’t needed, but fill seats. If anything, some would love to use AI to eliminate these types of jobs (if they even exist), but this is a topic beyond the scope of this essay.)

I experienced the curse of progressive disempowerment keenly. But, I also felt something else: determination. Because the more I worked with higher-capability LLMs, the more I recognized the extreme limitations of the technology.

I began to realize that I was far from becoming obsolete. Instead, my expertise, judgement and knowledge were absolutely needed for LLMs to operate at their highest capacity.

The Power of Knowledge and Understanding

My journey from fear and disillusionment to empowerment required understanding a few personal and technological truths.

• LLMs Have Knowledge, But Limited Judgement and Context: I had been working with LLMs and had a high-level understanding of how they operated, but I decided I needed deeper knowledge. I spent time studying transformer architecture, listening to lectures and running experiments in my own work (which had moved beyond the previous product). My goal: to understand how LLMs work and the edges of their abilities. This work is ongoing, but taught me that LLMs may have access to huge datasets, but don’t do a good job of understanding context and making consistently good judgements.

• Maintaining a Cognitive Moat is Essential: I’ve developed complex systems that enable agents to run autonomously across many areas, including content development, analysis and marketing. But, I realized the more I relied on some of these automations, the less I honed and maintained my cognitive capabilities. Today, I carefully curate how I use LLMs to maintain my edge and have developed systems to get the most out of LLMs in the least amount of time.

• Recognizing and Guarding Against Risks is Critical: Earlier this year, I saw how people were using powerful agentic systems like OpenClaw, but opening themselves up to major security risks. And, the recent incidents of ‘rogue’ AI haven’t surprised me much because I understand how LLMs are incentivized during training around task completion, which explains their behavior. I learned that having an appreciation for LLM risks is empowering because it fosters a healthy skepticism about LLM capabilities and their ability to complete tasks in non-harmful ways (without explicit guidance).

Subscribe now

From Fear to Respect and Empowerment

Today, I no longer fear generative AI. Instead, I have a healthy respect for it and a much better understanding of the technology’s benefits and limitations.

My journey also makes me very sympathetic to people who are anti-AI. According to data published by the Harris Poll in May 2026, about 12 percent of the global population are ‘Skeptical Resisters’ of AI. They understand and have evaluated AI, but reject it due to “distrust, discomfort and perceived risk.”

I’d also add fear to that list. A fear that generative AI will overrun and replace human cognition and agency across a range of areas.

Although I don’t agree with Skeptical Resisters’ dismissal of AI, the risk of AI-caused disempowerment is real. A major reason I developed the Doing AI Efficiently Operating System was to help people defend themselves against AI disempowerment by becoming better informed about generative AI’s capabilities and more able to use it optimally. All while preserving their skills and cognitive strengths. Knowledge and skills are empowering and extremely useful.

If you fear generative AI, my recommended approach is to learn how:

• The technology works (from first principles)

• It can be controlled

• To use it to maximize your personal and professional goals

Generative AI is here to stay, and we’re not being given a choice about whether to adopt it. Action is superior to passivity.

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Honor layer, which is about preparing yourself to take responsibility and ownership of AI-generated or aided outputs you produce.

Over the last few years, Pew Research has been publishing new findings on how Americans are using AI. The headline from their latest research: about half of U.S. adults say they never use a chatbot, and most of them have no plans to start.

Pew also found many are skeptical, suspicious and worried. Most predict AI will have a negative impact on their lives. Majorities say AI technology is advancing too fast and has significant privacy risks.

Usually, when people are uninterested in using a technology and have negative feelings about it, it’s easy to conclude adoption will be slow and arduous. What’s more, tepid consumer interest in chatbots, and, more recently, AI Agents, are powerful signals that AI may be in an extreme bubble. After all, if people don’t want to use the tech, it may not worth the investment.

But, there’s something different happening with AI. The reasons why have everything to do with a huge technology transition that took place in the 1990s and the rise of ‘invisible’ AI.

How technology adoption usually plays out

I’ve spent more than 20 years tracking (and living) technology adoption in a range of areas such as how builders adopted and reacted to AI’s use in coding. The famous adoption curve tells the tale. A small number of innovators and early adopters live at the bleeding edge. It can take a while for everyone else to catch up.

My lived example of this was in digital health, or the use of virtual reality, smart devices, wearables, artificial intelligence and other technologies in healthcare. I started working in digital health in 2005, focusing on the adoption of the Internet as a source for health and wellness information.

I then turned to mobile technology and other solutions. Broad adoption of digital tech broadly in healthcare took more than a decade. Patients were the innovators and early adopters. Organizations and government were (and still are) very slow.

There were many reasons for slow institutional adoption, including the need to prove digital health solutions are safe and effective, figuring out how to pay for them, and creating regulatory and legal processes.

Interestingly, AI has been in healthcare for a long time. Adoption wasn’t really any faster, because trusting an AI to diagnose diseases like cancer is orders of magnitude harder than relying on it to analyze a financial document.

Ultimately adoption comes down to infrastructure: regulatory, technical and legal. There’s another area where infrastructure played a major role in adoption: The Internet.

The rise and fall of (and rise again) of the Internet

It’s the question on everyone’s lips right now: are we in an AI bubble? The comparisons are stark. The dot-com boom was characterized by extreme exuberance. Many believed that the Internet, which had finally gained traction, and mobile technologies would usher in a golden age of major digital, societal and economic transformation.

At the same time, the cost of transmitting data fell as fiber optic cables were installed. Global Internet use increased and many investors thought bandwidth demand would grow exponentially. Telecom firms and startups accelerated their efforts to install fiber optic cables and increase wireless capacity.

The problem was supply versus demand. Although Internet traffic was increasing, the last mile problem persisted. Many Web users were unable to take advantage of fiber optic capacity, which meant download times were slow and frustrating. Fiber optic may have been under the street, but getting it into homes was a whole other problem.

In 2000, about 1% of U.S. adults subscribed to broadband. About 34% of Americans were on dial-up. Only 43% of Americans used the internet at all.

The last mile problem had a negative impact on Internet commerce as consumers couldn’t buy what they couldn’t access, and demand was much lower than expected.

Many publicly traded Internet companies had negative cash flows. Telecom companies had buried some 80 million miles of fiber across North America and Europe, but 95% of it may have been dark, i.e. not carrying data.

Markets were pricing a fully-formed, fast-moving digital economy. But when credit tightened after Fed interest rate hikes, and investors began to shift their portfolios from Internet stocks, disaster struck. In March 2000, the Nasdaq hit its highest rate ever, by April, it had lost a high percentage of its value. Dot-com and telecom companies that had taken on massive debt or had no cash reserves (and limited revenue) failed.

Despite this massive failure, the boosters were right: the Internet did change everything. Web commerce became commonplace. Demand for fiber optic exploded. Just not on the timeline everyone assumed.

What big tech internalized from the dot-com crash

The dot-com bubble burst is seared into the memories of everyone who lived through that era. And, it’s shaping the strategic decisions of big tech companies.

First, just like in the dot-com era, companies are spending huge amounts on infrastructure. In 2026, Amazon, Microsoft, Alphabet and Meta are on pace to spend roughly $725 billion on on AI infrastructure, up 77% from 2025’s $410 billion.

Where are they getting the money? Some of it is coming from cash on hand. Alphabet reported its first-ever negative free cash flow quarter in Q2 2026, driven by $44.9 billion in capex in a single quarter. Its full-year free cash flow is projected to fall roughly 90%, from $73.3 billion in 2025 to around $8.2 billion. Amazon is projected to turn free cash flow negative as well. For every additional dollar these companies generate in operating cash, they are spending approximately $1.57 in capital expenditures.

The invisible AI adoption curve

Are companies and investors making the same mistake of the dot-com era: over-investing in a technology where demand is lagging? The Pew chatbot stat is measuring one visible surface of AI that requires proactive engagement. Someone choosing to use a technology featuring an AI agent is another active choice.

The Internet has always required active engagement. People choose whether to go online, they decide to call up the cable company to have fiber optic installed, they have to visit a Website to make a purchase.

AI is different because it is being embedded into technologies that people are using already. Google, Apple, Meta and other big tech firms aren’t waiting for people to demand AI. They’re giving people no choice but to use it.

Google AI Overviews in search now have 2 billion monthly active users across more than 200 countries and 40 languages. This has transformed search engine marketing. The focus is now on how to be mentioned in AI search and cater to Internet-roaming AI bots. Consumers are adopting AI agents. They just don’t know it.

Embedment is going even further, the Gemini app has nearly 950 million monthly users as of Q2 2026. Apple Intelligence is active on 940 million devices, with 410 million daily active users.

Pew’s research reveals that the general public thinks AI means Chatbots. The reality is that AI is ... everything. And, because it’s embedded in search, mobile phones and other ubiquitous technologies, it’s already powering people’s most consequential decisions. (Side note: I started talking about technological embedment back in 2014 as it relates to digital health tech adoption (which was slow at the time).

The next embedment push is in hardware. In 2025, AI-capable chips were in 35% of all smartphones shipped globally, up 74% from the prior year. AI PCs are projected to cross 50% of the global PC market in 2026. Amazon says 97% of all devices it has ever shipped can now support Alexa+.

People are noticing the AI seems to be everywhere. According to Pew 40% of U.S. adults say they engage with AI at least several times a day. 63% are concerned AI is moving too fast.

Yes, 50% of Americans never use a chatbot.

But, they’re using AI agents, reading AI content and much more. They just don’t call it AI. They call it search, using a computer or operating a mobile device.

Resource: Infographic Invisible Vs Embedded Adoption

8.5MB ∙ PDF file

Download

An infographic explaining invisible versus embedded adoption, which is crucial for understanding how the usual tech adoption story may not apply to AI.

Download

AI and work: The next chapter

Pew found that roughly four in ten employed adults under 50 already use AI at work, and that number is likely to increase. A few months ago, many were talking about how AI was going to replace workers. And, in some cases they will. But, leaders are realizing that large language models are nowhere near ready to replace workers. So, they have to figure out how AI can augment work. And employees will need to understand how to use AI effectively.

What you should be focusing on now.

Stop asking the AI bubble question. It’s irrelevant

I think some people asking whether AI is in a bubble are looking for an excuse to ignore or dismiss it. AI is already ubiquitous, and is being used in everything from coding, to writing, search and mobile.

Moreover, although Anthropic and OpenAI capture much of the attention, I believe the companies best positioned to survive market declines are Goggle and Apple. They are well-positioned to buy bankrupted companies and assets.

Remember: AI Agents are infrastructure

At their core, AI agents are not complex. They are LLMs armed with tools being pointed toward tasks. They are capable of complex decision-making, which is what makes them valuable in situations that require nuance and judgment.

Although some people are productizing AI agents, I don’t see consumers buying many of them. Instead, they’ll purchase a service where agents will be embedded. Unless something goes wrong, they won’t notice them.

People need to be trained how to use AI

AI will be everywhere and people will be expected to understand what it is, how to use AI well and the ways AI can be used to augment rather than diminish cognitive capacity.

All of this will require training. Those who get ahead of the curve will benefit the most.

The 50% who say they’ll never use a chatbot aren’t wrong. Chatbots are a user interface. Why seek one out if you don’t need it?

AI however, will be everywhere. And people know it. How people adapt to these tools will be the adoption story to watch.

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Grasp layer, which is focused on helping you understand how AI works from first principles, including tokens, context windows, model architecture, how LLM content generation happens, AI adoption trends and more.

Earlier this month, Anthropic announced that it would be “working to include machine-readable marks in content that Claude generates.” It took this action to comply with the “EU AI Act’s Article 50 Code of Practice on Transparency of AI-Generated Content.” The watermarking will take two forms: “ (1) watermarks embedded in text, and (2) signed provenance metadata attached to files.”

Importantly, the Act specifies that the “marking obligation should not cover AI systems performing primarily an assistive function for standard editing or AI systems not substantially altering the input data provided by the deployer or the semantics thereof.” In plain language: There is absolutely no requirement to mark content that is only used for editing content or if the AI did not meaningfully change content produced by a human.

Anthropic has no obligation to introduce a blanket watermark on all Claude-generated content. However, it’s not practical for Anthropic to determine when the line is being crossed, so all content is being watermarked. According to Anthropic, when Claude “generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.” Importantly, “[b]ecause the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.” The last part is critical, which I’ll discuss later.

Anthropic isn’t the only organization to commit to marking AI-generated content. According to the European Commission: “Several well-established and prominent AI companies have committed to adhering to the code. Examples ... include: Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, Open AI, Synthesia.”

This means watermarking of AI-generated image, video audio and text content will be widespread. There are real benefits to AI content identification. For example, it can be used to prevent deep fakes and criminal activity using voice cloning. Another benefit is, when done properly, it can improve other AI-text detection techniques.

Some of these systems use machine learning to identify whether a text has been AI generated. However, they aren’t perfect and have high false positive rates (occasions when human-generated text is mistakenly identified as AI-written and vice-versa). Watermarking may be much more reliable. It can also be defeated.

The purpose of this essay is to help you understand the mechanics of AI text watermarking from a first principles perspective. How:

• Watermarking is implemented

• AI text watermarks are detected

• The ways it can be defeated

• How the scale of AI assistance can be measured using watermarks (this directly addreses concerns about watermarking being applied to light or minimal AI changes to text, e.g., for spelling corrections or light edits)

How LLMs Generate Text

Because AI text watermarking is conducted at the model level, we’ll begin with a high-level overview of how AI models actually generate text. For many, Large Language Models (or LLMs) are magical black boxes that can generate text or answer questions on-demand, with distinct personalities. LLMs don’t have personalities, but that’s the topic of a future essay.

The important thing to understand is that there is a standard pipeline used in LLMs to ‘ingest’ text and output a response. Watermarking involves modifying how LLMs generate text.

Text Conversion

Let’s say you decide to ask an LLM to tell you what words should be follow the phrase “The quick brown fox”. You input the text into a chatbox and the LLM provides a response. But how did the LLM do this? Let’s go through the steps.

First, a system called a tokenizer takes the words “the quick brown fox” and converts them into small units represented as a list of numbers, or vectors. A vector can look like this, for example: Token: ‘The’, vector [-0.12, 0.45, 0.03, ...] (768 numbers in the list).

This process of translating a text string into numbers is called embedding. This is required so that the LLM can process text strings.

Token Matching

Each token is matched to a large table of vectors that are in the model’s vocabulary. Because word ordering is important, the model also assigns a position id to the vectors. For example if the word “quick” was represented by a vector consisting of 768 numbers, it would be assigned a position id of 2 because it is the 2nd number in the phrase.

At their core, vectors are numerical representations. They tell the model that ‘the’ equals a specific list of numbers that represent a word. Words that are related are assigned vectors with similar numbers. This allows models to make associations between words. These associations are provided to LLMs during model training.

Transformation: Or Token Enrichment

After embedding, the vectors flow through the LLM’s transformer layers. During transformation, individual tokens are enriched with contextual information (e.g., a fox is an animal). An important step is that the model determines token relevance and semantic meaning, or the context of the tokens. Information about token definitions and relevance is provided to the model during its training.

After the transformation step, the model uses all this information to output a series of tokens with a score, or logit. This score represents the model answering the question: “based on what it knows about the input tokens, what are the most relevant tokens it should output?”

Token Scoring and Probability Matching

The model outputs a list of potential candidate tokens or logits. For example:

• Jumps: 8.2

• Lazy: 5.2

• Over: 1.8

• a: 1.1

• Dog: 1.05

• Brown: 0.4

The next stage is the crucial one: A calculation called the softmax function takes the logits provided by the model and converts them to probabilities, from 0 to 100%. This conversion answers the question: “what are the most likely relevant tokens related to this input?”

Each token is assigned a probability. For example:

• Jumps: 0.91 (or 91%)

• Lazy: 0.05

• Over: 0.02

• a: 0.01

• Dog: 0.005

• Brown: 0.005

It’s easy to assume that the model always picks the token with the highest probability. But, this isn’t the case. Instead a few different modifiers can be used to weight the probability of token selection in different ways to make the output more creative.

After this additional weighting process tokens (and their positions) are picked at random from the list of high-probability tokens provided by the model. This is what the model outputs to answer the question: “What comes after “A quick brown fox?” The model’s answer: ‘jumps over a lazy dog.’”

How Text Watermarking is Added and Detected

The specific method Anthropic and other AI labs will be using to watermark text has not yet been revealed. However, we have some prior work to guide us on some potential methods. In In 2024, John Kirchenbauer and colleagues published A Watermark for Large Language Models.

The paper outlined a process that could be used to reliably mark LLM-generated text in a way that survived copying, pasting and even re-wording. It is possible that Anthropic and other labs may utilize this foundational technique, so it will be the focus of my discussion.

Watermark Definition and Application

If you recall, Anthropic said: Claude “weaves an imperceptible watermark directly into the text itself.” Here’s how the weaving might happen.

At its core, a watermark is a statistical signal. It provides information about whether words were produced by a human or a model.

In the last section, I discussed how the model:

• Outputs a list of tokens along with a score (logits)

• The softmax function is then used to assign a probability of whether a specific set of tokens will be selected

The watermarking happens at the pre-softmax function stage. After the tokens are generated, but before they are assigned a probability, the set of tokens outputted by the model are assigned a color: red or green. The logits of tokens in the green list are modified (an additional number is produced), which is added to the token score.

For example:

• Green - Jumps: 8.2 + modifier: .02

• Red - Lazy: 5.2 (no modifier)

• Green - Over: 1.8 + modifier: .02

• Green - a: 1.1 (+ modifier: .02)

• Green - Dog: 1.05 + modifier: .02

• Red - Brown: 0.4 (no modifier)

What happens is that this slightly increases the probability of certain tokens being selected. This process works best if there are a large number of potential relevant next tokens rather than a smaller list. If the answer of what token should be selected for a sentence is obvious, it’s more difficult to determine whether a model produced the output.

Detecting the Watermark

The key to the detection method is answering another question: “Given an LLM or human, how likely is it that the human would have picked words from the green list in their content?”

If the list of tokens that could be selected is small, meaning the answer is obvious, detection is unreliable. If the list of potential tokens is very large, detection is more reliable.

If a text has a large number of tokens from the green list, it is less likely that a human develpoed it.

Importantly, the sensitivity of detection does not rely on producing a very long text:

• The list of green tokens has to be sufficiently large (the list can be large even with a short text)

• The percentage of green tokens in the text is higher than random chance would allow

The specific method used to determine what tokens are on the red versus green list must be kept secret. Otherwise it would be easy for others to avoid detection.

Resource: How AI Watermarking is Done

7.63MB ∙ PDF file

Download

An infographic illustrating how AI watermarking is implemented

Download

Defeating the Text Watermark

Now that you know the watermarking process, you can understand why Anthropic said: “the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing.” Specifically:

• The watermark is part of the text: Anthropic knows what tokens are on the green lists. Others do not, so it’s hard to determine what text to omit to avoid the watermark

• The watermark will travel with the text: When the text is copied and pasted the fact that it is on the list of green tokens does not change

• May persist through some editing: Without knowledge of the green list, one can’t target the tokens and delete them easily

Despite the difficulty of defeating the watermark, it is possible. Anthropic admits this in its announcement:

“Content generated by Claude may not carry a detectable mark if, for example:

• The text has been heavily edited, paraphrased, translated, or mixed into other writing;

• The passage is very short, leaving too little text for a reliable signal;”

The key to defeating the watermark comes from understanding that the strength of the method comes from the fact that each token (individual word unit) seeds the next red/green list.

For example, the phrase: “A quick brown fox” results in a list of red/green tokens associated with that set of tokens.

But what happens if the text is changed from “A quick brown fox” to “A rapidly moving fox”? or “A quick red fox”? The first example is a major edit. It’s unlikely that watermarking would survive that edit. The second is a minor edit. The list of tokens to select may still overlap with the red/green token list. In that case, the watermark may survive.

If you change enough tokens, the signal disappears. Minor edits may not make much of a difference.

There’s also the challenge associated with a large body of AI-generated text. Changing a few sentences in a 1000 word essay, largely generated with AI, won’t make much of a difference to the detection algorithm. Completely re-writing or modifying the text would work. This is the strategy some people in the open source community are using to attack AI text watermarking.

Methods for Stripping Text Watermarks

A new popular open source tool tool as been developed to defeat AI text watermarking. Here’s what the model does.

Text Cleaning, Backtranslation and Re-writing

First, a script can be run on the text to remove invisible markers in the text inserted by model providers. This process deletes these types of hidden marks, but has no impact on the text watermarking I’ve discussed in this essay.

The second stage removal process can involve:

• Rewording the text: The watermark can’t be removed by re-organizing the text. The text has to be completely re-written line by line.

• Backtranslation: This involves translating the text to another language and translating it back, which breaks the watermark.

• Working to maintain text quality: Re-writing the text completely can have a negative impact on quality. Correcting this requires using another high-capability model to re-write the text, ensuring quality is retained.

This process is complex and can be expensive depending on the model. Also, with many model providers signing on to EU AI Act, using an OpenAI model to re-write content produced by Claude will just result in a mix of marks in the content.

Human Content Creation from an Outline and Editing

Another way to defeat watermarking is to:

• Have a model generate an outline and use the outline to develop the text: No signal

• Re-write text the model produces line by line: Reduced signal

Resource: How to Evade AI Watermarking

8.16MB ∙ PDF file

Download

An infographic illustrating how AI watermarking can be evaded.

Download

Measuring the Unexpected Benefits of Watermarking: The AI Text Watermark Density Metric (AITDM)

A common concern people have is that the mere presence of an AI text watermark in their copy will brand it as AI-generated. For the percentage of people who are anti-AI, that might be the case. However, for everyone else, there might be an unexpected benefit: Being able to reliably determine how much of the copy was generated with AI assistance. To illustrate this I have created a metric called: The AI Text Density Metric or AITDM

Imagine documents A and B:

• Document A: Alice has an AI generate a financial analysis using a company dataset, assumptions outlined in a presentation and Slack messages. Her content is 100% AI generated, but she has restructured it. A watermark analysis is run on the content and finds that 5% of the content is human-generated and 95% is AI-created. It gets a AITDM of 95.

• Document B: Bob has created a different financial analysis using a similar data set. Bob has the AI generate an outline, but develops the analysis himself. He then asks the AI to reviw the content, but only correct grammatical errors and spelling mistakes. After analysis, it receives an AITDM of 15.

Clearly, Document A and B are in different classes. This is because AITDM allows us to measure the percentage of content developed by an LLM versus human.

AITDM analysis has a few additioal implications:

• Content with a high AITDM should warrant additional scrutiny. Did the LLM hallucinate data? Were the assumptions correct, etc?

• Content with the low AITDM should also be scrutinized because the author used an AI-generated outline to develop the content. However, it is less likely that the content suffers from hallucinations than fully AI generated material.

A Low AITDM Does Not Signal Quality

Importantly, a low AITDM does not mean the content is high-quality. Humans are prone to mistakes, misinterpretations and bad writing.

A high AITDM does not mean the user has out-sourced their strategic insights to a model. Instead, they may have carefully developed a prompt, checked the sources, corrected the model during development and taken other steps to ensure high-quality output. Judgement and expertise play a role here. In this case, the additional scrutiny is about accounting for LLM limitations rather than the overall quality of the output.

Where We Go From Here

Over the last few years the coding world has been embroiled in an argument about whether AI-assisted code is useful, relevant and high-quality. There are still many who don’t trust AI-developed code, but they understand that AI code generation is widespread. It’s becoming accepted in some corners that just because an AI developed code, does not mean that the producer is not responsible for the outputs, or that it is automatically low quality.

The writing world is now engaged in a similar debate. What makes AI-assisted writing different from coding is that it is easier to evaluate the outputs of AI-generated code. Scripts can be well-structured, secure, bug free and well-architected. There are software design principles that can be applied to code and testing can be done for correctness.

Doing this is harder with writing. People generally can’t tell the difference between well-written AI content and content produced by a human. A strategy is harder to evaluate for ‘correctness.’ It depends on a lot of factors.

Over time, I expect the debate about whether to use AI for writing will cool. Mainly because it will be so widespread. And, the presence of AI watermarking may prompt people to change their writing habits to get low AI text density scores, if only to demonstrate that they maintained significant control over the content development process, even if AI was used to aid their thinking. We’ll see.

References

• A Watermark for Large Language Models

• Watermarks Remover (Open Source Repo)

• Transformers Explainer

• Strong backing for the Code of Practice on Transparency of AI-generated Content

• AI Act Explorer (Article 50)

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Grasp layer, which is focused on helping you understand how AI works from first principles, including tokens, context windows, model architecture, how LLM content generation happens, and more.

In June, Ramp published data suggesting that the most intensive ‘AI-pilled’ companies spend about $7,500 per employee per month on AI costs.

The spending is significant and the data suggest many are working to control costs by turning to less expensive open source models, and using the right model for the task.

However, it’s unclear what firms are getting from an ROI perspective from their spend, and how they are measuring it.

Understanding the Behaviors That Drive LLM Costs

Cost will always be a major concern and area of focus for many. To help, I conducted research simulating more than 240,000 LLM users of tolls like Claude Code and Codex to understand the specific behaviors and habits that drive and save money on AI spend.

In addition to the research, I developed an assessment that reveals what behaviors drive your LLM costs over time and shows you exactly where cost savings can be achieved. The assessment takes less than 5 minutes to complete.

Take the Cost Assessment

For the last year, we’ve seen companies blaming AI for layoffs, saying that AI would deliver greater productivity and efficiency gains than humans.

This thinking was echoed by leadership and management of AI labs. For example, some said that software engineering was largely solved and that AI would be able to manage many routine coding tasks on its own.

But, as we have seen, the promise of AI’s ability to manage complex tasks autonomously has not matched up with reality. (This doesn’t mean that AI won’t take on a large part of cognitively difficult work, just that it’s going to take a while, or may even be deemed undesirable over time.)

Instead what’s happened is that people started to recognize that when it comes to working with AI:

• More does not always equal better

• Sometimes slowing down is essential

• Thinking carefully about what AI is going to output is vital

• Simplifying complex tasks into components is a best practice

• Cultivating judgement about what good looks like is essential

it seems that some companies are starting to recognize that human + AI is optimal and are beginning to accelerate hiring, and even re-hire individuals they let go previously. According to the Wall Street Journal:

“Companies ranging from railroad giant CSX to Google parent Alphabet have recently told investors that they plan to hire to meet growth goals or to seize on emerging technologies.”

My belief is that the next big deal when it comes to organizations working with AI

is understanding how to up-skill employees to get the most out of AI. This will require:

• Improving understanding of AI’s mechanics (tokens, context, etc.)

• Focusing on the daily, micro actions people take with AI (such as how they write prompts, manage context, reduce retries) to improve efficiency and increase task success rates

We’re still in the AI’s early innings. The modest reversal in the prioritizing AI over humans trend is heartening. The next trend to look out for will be efforts to understand how best to ensure humans are working with AI at maximum capability.

—
Benefit from These AI Efficiency and Productivity Resources

• LLM Whisperer Method: 3-part system helping people understand how to optimize AI micro actions to get the most out of these systems

• Understanding and Managing the AI Agent Footprint: Practical guidance on what AI agents install, access and change on devices and workflows and how to manage the AI footprint

This newsletter is part of the Doing AI Efficiently Operating System, built on five operational layers: Grasp, Discern, Ward, Execute, and Honor. This essay is part of the Grasp layer, which is focused on helping you understand how AI works from first principles, including tokens, context windows, model architecture, how LLM content generation happens, AI adoption trends and more.

I started my journey in generative AI when ChatGPT was first launched. It was a time of great promise, but also scarcity.

AI models (I was using GPT 3.5 and 4.0 at the time), were very context limited and had narrow reasoning capabilities.

I was working on early versions of retrieval augmented generation (it wasn’t called that at the time), where I’d provide the model with content from a data store. I also manually set up tools to deliver content from the Web to the model via Python scripts.

This work was extremely frustrating. Not only would the model hallucinate a lot, but you could only put so much information into its context. I had to put content through customized summarization pipelines and carefully construct my prompts to make sure the LLM saw the right things.

This was the minimum required to allow these models to function. There was also a financial incentive because input and output tokens were, on a per-token basis, expensive to generate.

The Scarcity Mindset

These early experiences shape how I use generative AI today. Even though models have larger context windows and can reason much more effectively, I’m still stingy with it. I don’t view inference as an unlimited resource and am careful with how I use these models.

This scarcity mindset influenced how I looked at a recent Anthropic study, focusing on AI’s use for work tasks. Anthropic found that AI is helping people engage in tasks that deliver more economic value over time. This is fantastic, but I thought: how much is this actually costing us?

So I decided to measure it. The result of this work is the Secrets of the LLM Whisperer study, which looked at the habits and behaviors that contribute to inefficient (and more expensive) LLM use.

A key driving factor is that, although we have more AI inference available to us, it is still a scarce resource that is being heavily subsidized. AI labs are feverishly buying data centers to meet the voracious appetite for inference. And, a backlash is growing, with some states imposing moratoriums on data center construction.

The cost of inference is also moving up. Anthropic is putting its most powerful current model Fable, behind stringent usage caps and pay-as-you-go payment schemes. Many are angry about this, but labs are heavily incentivized to stop subsidizing AI inference so heavily and have users pay closer to the market rate for it.

The study results and these market forces inspired me to launch this newsletter. In it, I’ll explore questions like:

• How can we use AI more smartly?

• What are the strategies and tactics being used to conserve resources (such as water, land and energy) consumed by AI?

• How can we make the social and economic costs of AI lower so ROI is higher?

• How can we operate AI more securely (which has its own efficiency benefits)?
  

Thanks for joining me on this journey.

When an agent runs, it leaves traces: packages it installed, files it read or wrote, environment variables it accessed, credentials it touched, processes it started, and tokens it spent. Taken together, those traces are the AI agent’s footprint.

Most development tools and AI coding assistants do not surface this information by default. The footprint is there. It is just not visible unless you go looking for it. This is true whether your agent is running locally, in a virtual machine, or in a cloud environment.

The gap between what an agent does and what you know it did leads to unanswered questions such as:

1. What is my agent doing?

2. Is my agent behaving normally?

3. Why did my agent do that?

4. How much is my agent costing me?

5. What changed?

6. Can I trust this agent?

7. What can this agent access?

I’ve develped a series of articles answering the seven questions above. Each provides operational insights around AI agent efficiency, cost maximization and security.

Read the Series