The pace of change and the capabilities are, if nothing else, fascinating.

Whether they’re good or bad, produce meaningful net economic or societal value, contribute to or detract from human wellbeing are all up for debate. It’s not at all clear to me and I have had different ideas on different days.

Below are a few of my thoughts about AI as we stand here at the end of Q1 2026.

Infinite productivity is a compelling illusion

As I dove headlong into Claude Cowork, Perplexity Computer, agent “skills”, and what these tools could mean for the kind of work I do my initial reaction (after some shock) was one of extreme excitement and optimism. It seemed like these things unlocked near infinite productivity.

Every idea could be brought to life in a fraction of the time. I could direct my agents to create plans and strategic frameworks, then execute on them step by step. They would faithfully work independently (sometimes for up to hours at a time) ticking tasks off their self-created list, checking their work, and finally presenting me with imperfect but totally passable work product.

I connected various tools, gave my agents access, spun up new workflows and invented all kinds of neat ways to take advantage of these new capabilities. I was the conductor of an orchestra of agents.

As it turns out, that’s exhausting. It’s more or less the manager’s schedule on steroids, with a hundred 2-5 minute “meetings” a day and a good recipe for frying your brain. Here’s another article on this same problem. Even the hosts of the All-In Podcast, all techno-optimist AI superfans, talked about experiencing it on their latest episode.

This is something anyone using AI agents is going to have to contend with and figure out how to manage. It’s also a pretty inconvenient finding for anyone hoping AI can deliver on its most optimistic promises.

Trying to keep up with AI is exhausting and not very much fun

Speaking of AI-induced exhaustion, trying to stay at the forefront of the wave is wearing me out. I’ve experienced many moments of wonder and excitement, but the predominant feeling is one of overwhelm and futility. Even as someone actively trying to stay current, and spending a lot of time and energy doing it, the technology is advancing at an incomprehensible rate.

For the moment it seems possible to at least stay ahead of the pack, if not at the bleeding edge, with reasonable effort. Most people know, or care, very little about AI. So, for now, you don’t have to be an expert to derive an advantage compared to most peers.

I don’t know how long this dynamic will last, though, for at least two reasons. First, recursive improvement (using AI tools to get better at using AI tools) could create a flywheel whereby an early lead becomes insurmountable, and creates such an advantage that all lesser relative advantages are lost in the noise. Alternatively (and I think more likely in the ultimate) is that the capability of the underlying technology advances to the point where “being good at using AI” is no longer a relevant skill. It will either be so easy that everyone can achieve the same results or even the idea of “using” AI will become obsolete; AI will just do the thing.

For now I think it still makes sense to build skills and adapt to the extent possible while maintaining health and wellbeing. But that takes discipline and explicit boundaries, often intentionally incorporating friction to counteract the allure of infinite productivity discussed above. That’s my plan for now anyway.

We are starting to see recursive self-improvement

I mentioned recursive improvement above in the context of people building AI skills. The more interesting application of the term is related to the AI models themselves, where in early 2026 we seem to be on the cusp of recursive self-improvement.

Recursive self-improvement means an AI that builds the next version of itself all by itself. That may sound far fetched, but the frontier AI labs are already heavily leveraging their AI to build the next versions of their products.

Recursive self-improvement is important because once it is possible then the pace of improvement is no longer governed by availability of human labor. The AI labs can spin up as many instances of their AI researchers as they have power and compute to sustain.

Time will tell how this pans out and we’ll see if the pace of model improvement continues on the exponential curve it’s been on, but there doesn’t seem to be much evidence to believe it won’t. Both Anthropic and OpenAI are rumored to be releasing new models in the next few weeks so we’ll get some new data points.

The pace of change is hard to overstate

My last thought here is just to reiterate that the speed at which this technology is changing is like nothing I’ve ever experienced. If your information is even a month out of date you’re woefully under-informed. A strategy from December 2025 is meaningless in March 2026.

Subscribe now

Does that sounds scary? Let me explain why I see opportunity instead.

First, a little background.

Back in 2023, a study by Eloundou et al. evaluated the potential impact of large language models (the foundation of tools like Claude and ChatGPT) on the labor market. Just last week, Anthropic (makers of Claude) released a paper that takes it a step further: they aggregate task-level exposure scores from the Eloundou paper and compare the theoretical exposure of various fields to the actual observed application of AI tools in those fields so far.

Guess which field has the biggest gap between theoretical and actual usage.

Yep, it’s architecture and engineering.

With 85% of A&E tasks theoretically well-suited to having AI leverage applied, practitioners are applying that leverage to only 5% of tasks! A&E is certainly well known for its glacial pace and resistance to change, so no real surprise here.

Here’s where the opportunity part becomes clear, though.

Keep in mind that AEC overall is a two trillion (with a T) dollar industry. A&E services are something like $350B of that.

If 85% of the current work (and $350B of value generated) in those firms can be meaningfully accelerated with AI, then a relatively small number of firms leveraging these tools effectively could serve a wildly disproportionate share of the market.

They will do more business while also delivering better client experiences.

The benefits that accrue to those AI-proficient individuals and firms will compound quickly.

That’s the opportunity.

Subscribe now

So what does this actually look like in practice?

Fair question. Here’s a real example from this past week.

Contract review is one of those tasks that I have to do but really don’t like. Reviewing non-standard client contracts, comparing them against the firm’s standards, flagging the problematic clauses, and producing redlines does not spark joy.

Last week, I used Claude to build a custom contract review “skill” for our firm. I fed it our existing contract review standards—the actual criteria we use to evaluate non-standard agreements. Claude built and tested a working prototype independently, then came to me with a draft for feedback. After a couple of iterations, it now produces specific, actionable contract reviews—complete with proposed redlines in a Word document with tracked changes—all consistent with our existing standards.

The whole thing takes a minute or two.

Obviously I still review the output. My judgment is the 15%. But the 85%—reading every clause, cross-referencing against our standards, formatting the redlines—that’s exactly the kind of work that AI handles well. It’s exactly the kind of work I aim to delegate.

That happened last week, and it’s an example of something that is only just now possible. We had previously used a Gemini Gem to help with this and it was OK but not nearly as capable.

Actually using these tools on an ongoing basis is essential to maintaining fluency and realizing the benefits that are available.

Now multiply that across every tedious, time-consuming task in your business. RFIs. Submittal reviews. Window and door schedules. Meeting minutes. Spec cross-referencing. Drawing coordination checklists.

You got into architecture because you love design, but now you spend your days in email and meeting purgatory? Good news—you can start getting back to the work that actually matters.

Are your clients frustrated by slow responses, high fees, and impenetrable documentation? You can leverage these tools to create a client experience that sparks delight, not derision.

That is the opportunity. Delegate and elevate, as the EOS folks say.

If you want to join the 5% who are actually seizing this opportunity, here’s my challenge to you for this week:

1. Keep a sticky note on your desk.

2. Every time you do a task that feels like soul-crushing data entry or administrative drudgery (e.g., summarizing meeting minutes, reformatting a schedule, drafting an RFI response), simply write it down.

3. At the end of the week, pick just one of those tasks.

4. Open Claude or ChatGPT and type this: “I have to do [one annoying task from your list] for my job. This is how I usually do it: [word vomit a description]. The outcome of all of that work is usually [describe deliverable or end result]. Can you help make that process easier? The specific thing I need to accomplish is [describe deliverable or end result again]—please suggest some ways you could help me to achieve that end result.”

5. Try what it says, evaluate the results, and give your feedback directly to Claude/ChatGPT. This is how you iterate and improve.

Don’t try to overhaul your entire practice. Just experiment with one tedious task and see what happens. You will be shocked at how quickly you can chip away at that 85%.

—Matt

P.S. This is, no doubt, an over-simplification to make a point. Go read the papers for the nitty gritty on the methodologies and what “exposed to AI” really means. Feel free to hand-wave this analysis away as sensational by nit-picking the details. But if you do, I think you’re missing the forest for the trees. And the forest is brimming with opportunity for those paying attention.

The coding revolution

Back in March of 2025 I wrote about building my OneFocus app, a simple task management tool built to my exact specifications, in about an hour, with next to zero coding experience. It felt like magic.

But that was a straightforward HTML proof-of-concept, and it was just the beginning.

Starting around December 2025, something shifted in a very big way. The agentic coding tools (things like Replit) crossed a threshold. They went from “impressive but not ready for prime time” to “genuinely capable and reliable coworker” almost overnight.

I’m not a software developer. I never studied computer science formally. But since that shift I have built dozens of useful, everyday applications, including tools that are now the backbone of how we run our business at CopelandBEC. Not toys. Not experiments. Production tools that we rely on every single day.

Not only that, but the pace of our in-house development is staggering. We make more progress on these tools in a week, working on them on the side, than real software companies used to make in a year.

The tools got so capable so quickly that the bottleneck is no longer the coding. The bottleneck is having a clear idea of what you want to build. If you can describe it properly (and I mean literally speak it into a computer) these tools can build it.

The pattern is repeating

Here’s the thing that I’m thinking a lot about now: the same pattern is now showing up in general knowledge work.

With coding, the progression was:

1. AI chatbots were helpful assistants for small, discrete tasks (debugging, learning, building simple tools that required additional work and technical ability to deploy)

2. AI developed into semi-autonomous workers (agentic tools that plan, execute, and check their own work across many steps and then deploy it in the real world)

The change from #1 to #2 facilitated a sudden, dramatic expansion of what’s possible for one person to bring into being.

I wrote in January about Claude Cowork, Anthropic’s tool that brings this agentic capability to everyday office work. You give it a task involving documents, spreadsheets, data analysis, file organization. It plans the work, executes it, checks its results, and iterates until it’s right.

This week Perplexity launched something called Computer, which orchestrates multiple AI models across research, coding, design, and project management in a single cloud-hosted system.

These are not incremental improvements to chatbots. They represent a fundamentally different way of working, where you describe a desired outcome and a capable system goes off and does the work.

The same threshold that coding crossed in December is now being crossed for knowledge work more broadly.

I don’t think the before/after line is quite as bright here, as I think this type of knowledge work is a lot “fuzzier” and more general than coding, but in concept it’s the same idea. This kind of thing (you describe outcome -> agent does complex work and produces decent results) simply did not exist before 2026; now it does.

It’s about leverage

In life we all have the same 24 hours in a day. There are always more ideas than hours, always more things I want to build, explore, or improve than I have capacity for. Good ideas wither simply because there is no time or energy left to pursue them.

What one person can create in those same 24 hours, though, is expanding. Rapidly.

The capacity to do things—to build a tool, analyze data, create a document, research a market, design a workflow—is exploding. Not in some theoretical future. Right now, in February 2026.

This means the constraint is shifting. It’s no longer primarily about bandwidth. It’s about vision.

Can you articulate what you want? Can you describe the thing clearly enough that a capable system can go build it? Can you look at your business, your workflow, your career and identify the highest-leverage opportunities?

That’s the new skill. Developing that skill means, first and foremost, putting in the work to learn and practice. It also means developing taste, judgment, expertise, and a deep understanding of how your work actually creates value.

What this means for you

I’m learning as I go, the same as everyone else. I’m playing with these tools every day, breaking things, being surprised by what works and what doesn’t. Chasing curiosity, incorporating lessons, and trying again.

But here’s what I’m increasingly confident about:

If you work in professional services—architecture, engineering, consulting, law, accounting, whatever—the window to get ahead of this curve is open right now. The tools are here. The capabilities are real. And most of your competitors are still figuring out ChatGPT fundamentals or pretending none of this is happening at all.

The people who will thrive will be the ones who combine domain expertise with the curiosity to experiment, genuinely question the core value proposition of what they do every day, and cultivate a willingness to rethink how work gets done.

Be well, talk soon.

-Matt

Subscribe now

Claude Cowork.

If you haven’t hard about this new tool from Anthropic here it is in a nutshell: it’s a semi-autonomous digital coworker that lives inside your computer.

You can read the details from Anthropic directly.

This is not just another chatbot, or an incremental improvement to an existing one. It’s a little glimpse into the unevenly-distributed future that’s already here.

So as not to over-hype Cowork specifically, it’s a “research preview”. It’s still early days. There’s an infinite number of things it cannot do, or cannot do well. It’s not AGI, whatever that means.

But it’s nonetheless very powerful, and it is one more data point to help illustrate the story of how AI is impacting work.

In broad strokes, you assign a folder to share with Cowork, and all the files related to its task go in there. These include reference files you start with, and also files that Cowork will create along the way. Then you give some instructions. Cowork asks for clarification as needed, and then goes ahead to plan and execute the work you’ve delegated.

Cowork can perform complex multi-step tasks, check its own work, and make revisions as necessary before ever presenting you with a deliverable.

I recently tasked Cowork with making a responsive web-based viewer for large format construction drawings. I gave Cowork the original PDF file and instructions. It was able to view the drawings, understand the logic of how the various sheets related to one another, and create an interactive website in minutes.

When I asked it to pull out individual details from the drawings so they could be linked to directly, it went through an exercise of cropping the specific areas as individual images, but some of the crops were cut off. When I pointed this out, it autonomously checked all the crops and adjusted them as needed. It performed this process recursively until it achieved the desired outcome, only checking in with me when it had completed its work successfully.

This semi-autonomous process is exactly how Claude Cowork itself was made. Anthropic has said that all of the code for Cowork was written by AI itself, using their own agentic Claude Code tool (Cowork is essentially Claude Code made for non-developer folks). Anthropic built and shipped Claude Cowork in about a week and a half, over the holidays.

Having used Replit, which is another agentic AI coding tool, for several months now I can attest to the power of this type of AI application. The ability for the tool to work for extended periods of time, across several steps, checking its own work as it goes is fundamentally different than what most people are used to with AI chatbots (to the extent that most people are used to AI chatbots at all).

...and now this power is coming to common “office work” involving Word docs, spreadsheets, file organization, data analysis, etc. in a way that’s accessible to people with typical office work skillsets.

It will be interesting to see how quickly this kind of tool is adopted and integrated into workflows, or is used to create entirely new workflows. As with most AI tool adoption, I expect there will be a steep curve at the bleeding edge, with a very few people creating outsized value and garnering outsized returns, while a long tail majority are still figuring out ChatGPT fundamentals.

A memo from Fiverr CEO Micha Kaufman has been making the rounds on social media. The first page has gotten the most attention from what I’ve seen, but the second and third pages are the most important.

Kaufman enumerates 7 actions you can take if the premise on the first page concerns you. Two thirds of this memo is about agency, and I think that’s a great place for our mindset to be when it comes to AI and the related changes to how we live and work.

Where does AI fit into your organization?

I’ve observed a tendency for people to dismiss, or at least back-burner, AI discussions and critical thought based on an implicit assumption that a particular type of work or task is too niche, too specialized, too physical, too [insert excuse] for AI to do.

I think this is a big mistake, and here’s the main reason why.

No matter what industry your business is in, no matter what you think makes your business money, I bet a huge part of what actually makes your business money are things that AI is already very good at (or at least good at helping with).

These are things like marketing, sales, risk management, and finance.

In short, AI is already terrific at amplifying one’s ability to do all of the basics that form the foundation of most successful businesses, regardless or niche or industry.

I say this from personal experience, and relate one related story below.

Using AI to help with financial analysis

As the chief executive of my small company I ask for a lot of help from many different sources. On financial matters, some of those advisors are AIs.

When the books close each month and our bookkeeper delivers monthly statements for my review, I will often delegate the first round of analysis and summary to my AI assistants. I similarly ask for their help with periodic cash flow projections and other financial analysis tasks.

In short order Claude, Gemini, or ChatGPT (sometimes all three) have prepared detailed summary reports for me to consider along with the statements. The reports include a breakdown of the numbers, relevant interactive visuals, and recommendations for indicators to watch or actions to take.

I usually do this with a simple “act like my CFO” type prompt. I’ll then follow up with questions and discussion as I would with a human finance team.

I, of course, review the results with a critical and skeptical eye, building trust over time where warranted and getting to understand limtations. In this way, my interaction with these tools is very much like my interaction with employees and other teammates.

These tools have not replaced anyone, they have allowed me to do more. They have enabled me to be a better executive and leader by putting more information at my fingertips, giving me more and better analysis for decision-making.

Delegation is a skill that takes deliberate study and practice, and results are highly correlated to the capability of the delegator. This applies as much with AI as with people.

Where AI fits in every business

Today, in May 2025, even if AI development stopped right now there are huge opportunities for businesses of all sizes, in all markets and niches, to leverage AI tools in ways similar to what I just described.

Obviously it makes sense to try to discern where things are moving and skate to where the puck will be, but there is also much that can be done right now and I think those present-day actions are often overlooked while we fret about the future.

Every business could benefit from a more efficient, faster, knowledgeable, and scalable back office foundation in core functions like sales and finance.

Whatever your role is in your business, I’m quite sure there are aspects of your job that AI can help you do better.

For less than you spend on coffee you can have access to these kind of tools today.

What are you doing to learn these tools and hone your superpowers?

This post’s format is inspired by 50 things I know by Sasha Chapin, who clearly knows a lot more than I do.

Sasha's 'Newsletter'

50 things I know

I know what makes people grow more reliably than anything else. It is: taking on a difficult project with some amount of public accountability. This can be large or small: a lecture series, a business, a blog, a house, a child, etc…

Read more

2 years ago · 1584 likes · 82 comments · Sasha Chapin

Finding myself at the apex of a phase of extreme optimism for AI development, I noticed over the last few days that I’d swung too far into blind enthusiasm, bordering even on idolatry.

One sign of this starry-eyed fanaticism was the sheer number of redundant AI apps and tools I was using (and paying for). ChatGPT, Claude, Grok, Gemini, and Perplexity… all with their own monthly subscription fees. Each had an icon on my phone’s home screen.

They all offer more or less the same features and capabilities, with only slight variations at the margins. Any one of them would have seemed magical just a few years ago—a groundbreaking tool of revolutionary power.

But five of them all at once just created friction and vibration—movement for movement’s sake rather than forward progress.

I rationalized that I was testing, researching. That I was learning something and coming back with useful knowledge to share.

In hindsight I think a lot of that time was wasted and represented idle fiddling more than productive work.

It’s not the tools that are so important, but the applications and the impacts. The individual tool matters far less then how it is used.

I want to refocus my energy on the applications and impacts of AI more broadly, rather than getting lost in the weeds of each day’s new feature announcements.

So I’ve set some intentional constraints and cancelled subscriptions to all but ChatGPT. I still have access to Gemini through my Google Workspace plan as well.

It’s not a cost issue, per se, at least not in terms of dollars. It is an opportunity cost. Time and mental cycles wasted.

I don’t want to venerate AI tools (or any other tools) on their own. To me, tools are useful and important to the extent that they serve some greater project—to the degree that they help create, build, and explore.

Too many tools, even great ones, just get in the way. Sometimes it’s important to clear out the garage, keeping the essential and letting go of the rest.

Subscribe now

Gemini in Google Sheets

You can now type ‘=ai’ to access Google’s Gemini AI directly inside a spreadsheet cell in Google Sheets.

Read more here.

Claude adds deep research and new connections to Google apps

Anthropic has added a “deep research” functionality to its chatbot, Claude, as well as new connections to Google Workspace apps. Claude can now access data in Gmail and Google Calendar, along with Drive.

Deep research has been among the most impactful use cases I’ve found for these AI tools, so I’m happy to see it come to Claude.

Read more here.

Grok adds Studio and Workspaces

Grok now has canvas and project features in its app, similar to ChatGPT and Claude.

Canvases allow you to work on a document, code, or other generated content in a window alongside the main chat thread. Projects allow you to group related chat threads together, along with background documents, and set custom instructions that direct all chats in the project.

Grok calls these “studio” (canvas) and “workspace” (project).

I hope everyone is well. As ever, the tech and AI world is moving a breakneck pace and this past week has been packed with developments.

Read on for two high-leverage tools and a rundown on the biggest recent news and research that will impact your life as a knowledge worker in 2025.

An AI-Powered Graphic Design Team

It’s been a while since I’ve found a tool that excites me as much as Typeset.

Typeset is a web app for creating visual content like slide decks, e-books, social and media posts. It leverages AI to assist with various aspects of the design and layout, which:

1. saves a ton of time

2. empowers you to create compelling visuals without graphic design and layout skills

Typeset is sort of like if Canva and iA Presenter had a baby. I adore both of these tools as well, and still use them (especially Canva), but Typeset stands out for its ability to help you quickly create compelling visuals at scale.

In Typeset you simply type (or paste) in text content and the AI-powered tool immediately creates a bunch of layouts for you to review and work with. It regenerates the options after each change you make, adjusting things as you go.

Adding images from stock sites like Unsplash (or Getty on the Pro plan) is directly integrated into the app and super smooth. Exporting to various file formats and layout sizes is easy.

There is a lot more as well—I’ve been using it for about a week and have only scratched the surface of its capabilities. One example: I created the banner image at the top of this post in Typeset.

What I love most about Typeset is the sense of confidence and energy it creates in me. I’ve noticed that with Typeset in my toolkit ideas blossom instead of wither.

I find myself excited to create instead of overwhelmed at all the work of turning my idea into something tangible. Typeset makes creating these visuals fun.

It’s like having an enthusiastic, fast-working team of graphic designers on demand, ready to collaborate on creating your next project.

A Custom Podcast on Any Topic In Minutes

In 1940 Winston Churchill wrote to his war cabinet:

To do our work, we all have to read a mass of papers. Nearly all of them are far too long. This wastes time, while energy has to be spent in looking for the essential points.

This is as true in 2025 as it was in 1940—as knowledge workers we’re constantly trying to read and digest all kinds of documents.

Google just created an incredible tool to help with this problem—it’s called Audio Overviews and it lives in the Gemini chatbot app.

With Audio Overviews it’s as if you have a crack podcast team on call 24/7, including researchers, hosts, and producers.

Here’s how it works:

1. you upload a document (e.g. a long report, a scientific paper, annoying email chain, etc.)

2. wait a couple of minutes

3. Gemini creates custom podcast in which the co-hosts talk through the key points from the document

This is not AI text-to-voice transcription. The resulting audio is not simply a reading of the document. It is a bespoke discussion, including interpretation and breakdown of the document’s content.

The voices are natural-sounding. I find the banter engaging. I actually enjoy listening.

Curious how this actually sounds?

Keep reading—I include an example below with the rundown of the recent Microsoft/Carnegie Mellon paper AI and critical thinking.

The Impact of AI on Critical Thinking

Should we be worried about skill atrophy when using AI?

This is certainly a question worth considering. One data point comes from a recent study by researchers at Carnegie Mellon University and Microsoft Research, entitled The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers.

I find the title of the paper somewhat misleading. The title seems to imply that AI use caused a reduction in critical thinking and confidence in survey participants. I don’t think the content of the paper supports this.

The main finding of the paper, which I find very interesting, is that one’s inherent level of self-confidence directly correlates to the level of critical thinking one applies to a task (vs. offloading the critical thinking to an AI).

In other words, when someone has a high degree of confidence in their ability to perform a task, they will apply more critical thinking than if they have lower self-confidence.

They may still engage an AI tool to help with the work, but they will apply critical thinking to the way they use the tool, how they interact with it, and how they verify the results.

The paper also discusses how the availability of AI tools can shift where humans apply their skills in the work process:

Specifically, for recall and comprehension, the focus shifts from information gathering to information verification. For application, the emphasis shifts from problem-solving to AI response integration. Lastly, for analysis, synthesis, and evaluation, effort shifts from task execution to task stewardship.

As an entrepreneur and business owner, this sounds a lot like the transition one goes through during one’s career—growing from more of a “doer” role to a role in which delegation, management, and leadership are the essential skills.

In other words, AI is enabling knowledge workers to move higher up the value chain—creating leverage that previously required managing a team of people.

Speaking of leverage, how about creating a custom podcast to discuss the content of this paper?

The old way:

1. hire a researcher to read the paper and create show notes

2. hire the “talent”—a couple of co-hosts—to present the findings

3. hire a producer to record, edit, and create the final product

4. wait several days at least for all of this to get done

The new way:

1. upload the paper to Gemini

2. click “Generate Audio Overview”.

3. wait 1-2 minutes

4. click the play button

Want an example? Here’s the 7-minute podcast episode Gemini created for me about this Microsoft/Carnegie Mellon paper.

Top AI News

Still reading? Cool! Here’s a quick summary of some recent AI news that will impact life as a knowledge worker.

• Google released Deep Research with Gemini 2.5 Pro for free.

• Google also announced Firebase Studio, a web-based “vibe coding” platform that I expect will help unlock all kinds of custom app development.

• Wordpress released an AI-powered site builder, aimed at helping businesses create better websites faster and easier.

• Google and author Lee Boonstra released a white paper on prompt engineering. While highly technical in parts, there are some gems for getting better results out of LLMs.

That’s it for this week!

I hope you found this helpful and I’d love to hear your feedback. Anything stand out as particularly great? Anything you could do without? Please let me know!

News I’m Following

The pace of news in AI continues to be blistering, with new releases and announcements from the major players on a daily basis. Here are some that caught my eye over the last couple of weeks:

• Microsoft has added agentic features to Copilot, such as being able to take actions on your behalf at sites like 1-800-Flowers.com and OpenTable.

• AI labs have released or announced several new models, including:

  • DeepSeek V3

  • Gemini 2.5 Pro

  • Llama 4

  • full o3 and o4-mini

• OpenAI released a significant upgrade to ChatGPT’s image generation capabilities

• Perplexity continues its momentum, shipping a bunch of new updates and features including improved source-checking and a simplified UI where the AI dynamically chooses the appropriate model for the task (ChatGPT could take a cue).

How I’m Getting Leverage

One of the most useful applications of AI tools for me has been in processing RFPs and developing proposals.

Writing proposals is a big part of my professional services business. Whether the RFP comes in the form of brief email or phone call, or a detailed formal document, we spend a lot of time reviewing, evaluating, and responding to these requests.

Responding to RFPs are what we call “non-billable” time—in other words it’s a cost of doing business rather than an activity that directly generates revenue.

Since launching my firm 7 years ago I’ve strived to compress the time spent on proposals to a minimum, looking for efficiencies that previous firms I’d worked for had missed.

Recently this has included using AI tools, primarily Claude Sonnet and its Artifacts feature, to cut down on non-billable legwork. It’s been great for:

• summarizing the RFP and its key points

• making lists of essential requirements to be included in an RFP response

• helping to evaluate go/no-go decisions

• drafting proposal language consistent with our standards

…and I know I am just scratching the surface here. Listening to a recent KP Reddy podcast on this topic I gathered several more ideas, like using AI to check a proposal draft against the RFP and identify any potential shortcomings.

The podcast also includes references to a couple of purpose-built tools for AI-powered proposal workflows like Joist AI and WorkOrb. I haven’t tried these but will report back if I do.

Trend I’m Watching

A lost of the focus on AI is at the bleeding edge, where companies are spending billions of dollars to buy and operate the latest hardware. Those top line numbers get a lot of attention.

What’s more interesting, though, is the pace at which the cost of any given level of performance is dropping like a rock.

In this chart from Ethan Mollick you can see that for comparable performance at the “cost frontier” (i.e. the most cost-efficient models) the cost has decreased 99.7% in less than two years.

Relatively few people can afford or justify the cost for access to top-end AI models. But as the cost to use AI tools that seemed like magic just a couple of years ago trends towards zero, the democratization of access to those capabilities is fascinating.

Could the cost frontier could end up being more impactful than the performance frontier?

Let’s dive in!

Tool I’m Using

I’ve been rapidly increasing my usage of “deep research” tools in the various AI chatbot apps. I first wrote about this last week, when I described using Open AI’s Deep Research to help identify potential clients.

I’ve found a new use case: conducting autonomous research studies to identify pain points felt by customers.

Here’s an exact prompt I used with Grok 3’s DeepSearch this week:

What are the biggest pain points that attorneys experience when dealing with experts in their litigation matters? Scour X postings on this topic as well as any published articles etc in trade publications. I want to know what drives attorneys nuts when it comes to dealing with their experts.

Grok was particularly helpful here because of its direct access to posts on X. The AI research agent surfaced a ton of enlightening content that can inform my marketing efforts and my work, all in a matter of minutes.

ChatGPT, Perplexity, and Grok all have useful “deep research” functionality built into their apps.

Content I’m Reading

I love this post by Andrew Ng about the power of AI to make human knowledge workers an order of magnitude more capable.

I wrote previously that I see tech-savvy people coordinating AI tools to move toward being 10x professionals — individuals who have 10 times the impact of the average person in their field. I am increasingly convinced that the best way for many people to accomplish this is not to be just consumers of AI applications, but to learn enough coding to use AI-assisted coding tools effectively.

This was the post that motivated me to vibe code my focused task app. It’s a reminder that we should think about AI not as a way to automate jobs—but rather a way to automate tasks and thereby augment jobs.

Subscribe now

Video I’m Watching

Boston Dynamics released a new video of their robot, Atlas, doing cartwheels and other complex movements with a high degree of human-like smoothness.

For reference, compare the video above to one released about a year ago. Pretty amazing progress.

News I’m Following

Anthropic’s Claude can now search the web. I love this because I find Claude to be the most useful model for “work” tasks like data analysis, parsing and interpreting long documents, and content creation. Adding the ability to get real time info from online is a nice bonus.

I still default to Perplexity for quick web searches and answers to questions, mostly using its Cmd-Shift-P keyboard shortcut.

Feeling scattered? Too many conflicting priorities pulling at your attention?

To-do lists and task managers make the problem worse. I know because I’ve tried nearly all of them. The only thing they did well was to create a laundry list of everything I once thought was important (and now just stresses me out).

Instead, let’s break things down. Serial, not parallel.

Try identifying just ONE thing that you want to accomplish today. Take the time to choose the ONE thing that will be most impactful and focus on only that until it is done. Then move on to the next.

I created a simple web app to help. Check out OneFocus here. It’s free and open source!

When I started this newsletter I thought I’d focus on specifically how tech and AI was being applied in the architecture, engineering, and construction space. As things have evolved, however, I’ve started writing more about how technology tools and developments affect “knowledge workers” more generally.

The great opportunity right now is for people who make a living applying their intellect and creativity to leverage those capabilities using these new tools.

The goal is to scale your strengths and offload the rest.

So I decided to do a little rebranding to better align with this focus. Let me know what you think—and if you love it, share it!

Share Leverage

Ok, on to the good stuff…

Tool I’m Using

I’m really excited about this one. For the last few days I’ve been using a slick app to help me stay focused on my most essential task each day.

I’ve tried every task management app under the sun and most excel at one thing: building a huge long list of things that once seemed important, but now seem like a burden. I spend hours managing the list instead of actually getting anything meaningful done.

Instead, I wanted an app that helped me focus. I wanted to start every day fresh and identify the ONE most important thing I needed to accomplish.

So I built a bespoke app to my exact specifications.

Using mostly Claude 3.7 Sonnet along with a couple of other AI tools I created a web app that does exactly what I want it to do and nothing more. It looks how I want it to look and has the features I want it to have.

I did all this in about an hour.

I have next to zero coding experience. I have never built an app before. This was all vibe coding.

You can check out the app for yourself here. Please let me know what you think!

Subscribe now

Prospecting with Deep Research

Deep Research is a term used in a few different AI chatbot apps for an agentic (i.e. somewhat independent) process where the AI goes off and does research on your behalf, then returns with a report.

This week I’ve been using OpenAI’s Deep Research to help with prospecting for potential new clients.

For example, one of the types of work I do in my building envelope consulting business is to provide expert witness services to attorneys.

In the past I might have used Google or another search engine to identify attorneys that handle construction claims litigation in a given market. This would involve a search query followed by clicking through several websites to find appropriate contacts.

Now I simply describe the type of client I’m looking for to Deep Research (often using superwhisper to give a ton of context and nuance quickly and easily) and ask for a nicely-formatted table of results. A few minutes later the AI returns a list of names and contact information.

That’s it for this week! Talk soon.

-Matt

Quick Note: If you find these newsletters interesting, helpful, or otherwise a good use of your time, it would mean a lot to me if you forwarded this on to just one person you think might also enjoy it. Thanks!! —Matt

Tool I’m Using

This week I discovered superwhisper. Superwhisper is an AI-powered voice-to-text app that works seamlessly across apps on Mac and iOS.

I’ve avoided experimenting with these kinds of apps before because I didn’t think I needed them. I figured my phone already had a voice-to-text feature if I wanted it.

But superwhisper is different.

First, it’s more accurate than the voice-to-text built into the OS. But more than that, you can set up various “modes” (like text, email, notes, etc.) that properly format the output for each situation. The AI first processes your voice into text, and then it can post-processes that text into a more useful final output.

Subscribe now

Video I’m Watching

In this video Barbara Jovanovic explains how she runs a “zero-employee” marketing agency powered by AI tools (including the above-mentioned superwhisper).

This video was immediately helpful for me in two ways:

1. ideas for powering up the existing internal marketing and content team in my business

2. improved my understanding of how to leverage AI tools in general, e.g. using ChatGPT to help craft effective prompts for… ChatGPT

News I’m Following

OpenAI is reportedly planning to release a series of “agents” ranging in price from $2,000 to $20,000 per month.

As a reminder, in this context “agent” means an AI-powered entity that can operate independently, without direct prompting for every step. The currently-available “Deep Research” tool in ChatGPT could be considered an early research agent, as it takes your initial prompt, goes off and follows various leads, then compiles an organized report.

Deep Research is currently available for $20 per month, which even one useful research report would easily justify.

What would an AI tool need to be able to do for you to consider shelling out $2k per month for access? What about $20k/month?

That’s it for this week! Be well.

Matt

Tool I’m Using

Claude + Artifacts is one of my favorite and most heavily-used AI tools right now. It’s been a big part of my workflow for several months, and I’m hopeful that the recent upgrade to the 3.7 model only adds to its power and usefulness. This video by Jeff Su does a great job of illustrating its capabilities.

How AI is Helping Me

• Slack’s daily recap, part of its AI suite, helped me catch up on a week’s worth of messages missed over vacation in a matter of minutes.

• I used Claude to help me review a proposed office lease agreement.

News I’m Following

Figure, an AI technology company building humanoid robots, has been determined that they’re so far ahead of schedule in their development that they’ve moved up their timeline for having humanoid robot assistants in the home by two years. They’re now planning to start experimenting with humanoid assistants in the home this year.

How long before robots like this can perform basic job site labor like cleanup and moving materials around?

Article I’m Reading

The End of Search, The Beginning of Research by Ethan Mollick. Ethan is a professor at the Wharton School of the University of Pennsylvania and his LinkedIn account is one of my favorite sources of well-reasoned AI analysis. His Substack, One Useful Thing, takes the analysis to the next level.

I've been discussing two parallel revolutions in AI: the rise of autonomous agents and the emergence of powerful Reasoners since OpenAI's o1 was launched. These two threads have finally converged into something really impressive - AI systems that can conduct research with the depth and nuance of human experts, but at machine speed.

That’s it! I hope you have a wonderful week.

Be well,

Matt

There’s an interesting dynamic at the moment where it’s not clear where the value will accrue in AI development.

Will it be at the model layer, by companies like OpenAI, Anthropic, Meta, Google, or DeepSeek?

Or will it be at the application layer, in tools like ChatGPT, Claude, Perplexity, Covve (an AI-powered personal CRM app I just found), or myriad others?

In some cases, like ChatGPT and Claude, the application is made by the same company that developed the underlying model.

In others, like Perplexity, the app developers use a variety of underlying models (created by the model builders) to power their apps.

Here’s a recent example of this dynamic playing out:

Just a couple of weeks ago we witnessed DeepSeek (a model builder) make waves with it’s R1 model, which surprised a lot of people due to its nearly-state-of-the-art capabilities, efficiency, and the fact that DeepSeek released it open source. With DeepSeek being based in China, some people were hesitant to use the DeepSeek-hosted version of R1 due to concerns about data security and privacy. Perplexity (an app developer) quickly addressed this by incorporating a U.S.-hosted version of R1 into its product within days.

It will be interesting to see if, over time, the value accrues more to consumer-facing tools (e.g. Perplexity) or the underlying models (e.g R1). Will the models themselves become a commodity? Or will the state-of-the-art stay sufficiently far ahead of the pack that it maintains its value?

2035 - Where are we going?

Now the other end of the spectrum… with all of this new technology and the pace of change, what will the impacts be on a large scale?

The people who presumably know the most about what capabilities are coming down the road had some interesting things to say about this topic over the last week.

First, Sam Altman (CEO of OpenAI) described his vision of the future like this (link, emphasis added):

Anyone in 2035 should be able to marshall [sic] the intellectual capacity equivalent to everyone in 2025; everyone should have access to unlimited genius to direct however they can imagine.

Apparently spell check hasn’t proliferated 100% yet, so perhaps take the prediction with a grain of salt, but can you imagine what the world could look like if Sam and his colleagues are successful and everyone has a team of Einsteins at their disposal?

Dario Amodei (CEO of Anthropic) is apparently concerned we’re not taking the issue seriously enough (link, emphasis added):

Time is short, and we must accelerate our actions to match accelerating AI progress. Possibly by 2026 or 2027 (and almost certainly no later than 2030), the capabilities of AI systems will be best thought of as akin to an entirely new state populated by highly intelligent people appearing on the global stage—a “country of geniuses in a datacenter”—with the profound economic, societal, and security implications that would bring.

Time will tell whether these companies, or others, actually create such powerful artificial intelligence. The more I learn and consider the topic, though, the more optimistic I get. What started as a fear reaction back in 2023 has become one of great curiosity and gratitude.

What a wild time to be alive. Of all periods in known history, having a front row seat to today’s current events seems to be a special gift.

The fear of the unknown is a reasonable, rational instinct that seems hard-coded to protect us from danger.

Looking back at actual history, though, shows us that technological improvement is not only what humans instinctively do, but overall it is responsible for a steady improvement in our safety, wellbeing, and capability.

Yes, there are bumps in the road. Yes, technology disrupts current ways of doing things and brings certain ways of life to an end.

But humans are incredibly adaptable. We have repeatedly used technology to create bountiful new ventures, and exciting new ways of interacting with life that were previously inconveivable.

What I’m coming around to is a recognition that, while I can see neither the road nor the destination clearly, I feel a growing confidence that we’re headed in the right direction.

It’s more instinct and intuition than rational analysis, but part of it is a clear-eyed look at the past and extrapolation from human history to date.

I’m excited to see how this all shakes out.

Subscribe now

So I’m going to try to revive this newsletter with these goals:

• curated AI and tech news for AEC (especially AE) firm leaders and innovators

• concise and high impact

• informal and easy to digest quick reads

• focusing on consistency and frequency

Basically, I want this newsletter to be a valuable asset as you try to navigate these fast-changing times.

Here’s some interesting stuff I’ve dug up lately…

Subscribe now

AI-Powered Generative Design Tools

AI-powered tools are being developed that take architectural drawings as input, process that information, and then output something new.

• One example is Higharc, which “took a sales brochure image of a two-story plan and converted it into a fully detailed, structurally integrated 3D model within minutes” during a recent demonstration.

• Another is Genia which generates structural designs from architectural floor plans. The tool evaluates multiple options for the structural system using finite element analysis to optimize the design.

• Firms are also using more general tools to enhance and speed up their processes. For example, Zaha Hadid Architects reports “productivity gains of up to 50% in mid-stage building design preparation, facilitated by tools such as Stable Diffusion and AI-integrated platforms like Rhino.ai”.

Homegrown AI Assistants

Larger and smaller firms alike are leveraging AI to build “assistants” that help parse vast amounts of data quickly.

• “Balfour Beatty is in the process of developing StoaOne, a large language model generative AI assistant that will help employees mine… ‘untold billions of data points.’” (Construction Dive)

• Skanska is building a tool “which will help employees at the company plumb the firm’s collective expertise on projects and problems.” (Construction Dive)

• A consultancy in my own field, Building Science Corporation, has built an AI-powered building science chatbot.

That’s it for now! If you have any questions or want more details please drop a comment below. I’d also love to hear if there are any areas you’re particularly interested in learning more about.

I started out in November 2022 understanding AI at a very superficial level as something that helps power social media algorithms and Google search results. Then ChatGPT happened and I started paying attention for real. Here in April 2023, I’m honestly not quite sure what to make of it, but I’m confident it’s way bigger than something that designs TikTok and YouTube feeds (which would be an important enough impact to warrant more attention than it gets).

I started this newsletter in part to give me a reason to keep up with AI technology as it relates to the AEC industry, where my professional services firm lives. That mission quickly morphed into one of tracking, and learning about, AI tech more broadly as I became fascinated by near daily news about one mind-blowing development after another.

Come to find out there are some very smart and well-funded folks out there dead set on creating computer intelligence that rivals, and eventually far surpasses, that of humans.

Will they succeed? What will it mean if they do?

These are questions that I’ve thought a lot about over the last few months. Too much, truth be told.

If I had to guess I’d say we’re heading for some wild times. If 2030 looks as much like today as today looks like 2016 I will be surprised.

But I could certainly be wrong. I could be caught up in the hype. The recent developments could turn out to be a blip. For one reason or a hundred all of this technology could fail to result in the society-shaking changes that many expect.

The bottom line is it exceeds my faculties to predict where this technology will go and the impacts it will have. Yet it’s very difficult to resist the allure of trying to figure it out in some misguided attempt to be prepared for an unknowable future.

Never let the future disturb you. You will meet it, if you have to, with the same weapons of reason which today arm you against the present. —Marcus Aurelius, Meditations 7.6

This is all to say I think I’m best served backing away from following the frenetic pace of AI development so closely. It’s not good for me—I need some space. I can’t see the big picture, can’t focus on the broad strokes that will actually impact my day-to-day.

It’s all out of my control anyway, of course. OpenAI, Microsoft, Google, Anthropic, et al. are going to do what they’re going to do, and the world will react. We’ll make the best of whatever the playing field is at the time.

I’m expecting there will still be some stuff worth writing about on the topic of AI in general, and as it relates to AEC in particular. I want to do my best to filter out the noise and report on the more powerful big picture trends and actual applications that are demonstrated to be impacting the world today—and spend less time on the latest updates, predictions, and speculation.

There are plenty of places to go for that kind of futurism and headline news. I’ll link to some YouTube channels at the end of the article that you may want to check out if that’s what you’re after.

I’m planning to write a bit less often—monthly, more or less. Please subscribe or share if that seems like something you’d like to see.

Subscribe now

I’d love to hear your questions, if you have any, and I’ll do my best to answer them in the next post.

Be well.

-Matt

AI YouTube

Links are not endorsements. I’ve found these channels to be interesting. Proceed at your own risk.

• All About AI

• AI Explained

• Till Musshoff

This week I’m sharing a few robotics projects currently in development that I’ve learned about recently. Some are more directly applicable to the AEC industry than others, but I think all serve to paint a picture of the current state of the technology.

Subscribe now

Boston Dynamics

Many are familiar with the work of Boston Dynamics as a result of their compelling videos. This video from 2 months ago demonstrates the ability of their robot, Atlas, to interact with objects on a construction site including a bag of tools and pipe staging.

Ameca

Ameca is a robot made by Engineered Arts, a UK company. It is capable of some very human-like facial expressions.

Figure

Figure is a newer company with an explicit goal to “eliminate the need for unsafe and undesirable jobs”. Per its master plan, Figure intends to:

…develop general purpose humanoids that make a positive impact on humanity and create a better life for future generations. These robots can eliminate the need for unsafe and undesirable jobs — ultimately allowing us to live happier, more purposeful lives.

Disney

Finally, on a lighter note, Disney introduced a lovable rollerblading robot “specifically designed to create an emotional connection with people”.

Subscribe now

AI News Updates

• Microsoft fleshed out its ideas on implementing AI in its Office tools.

• Text-to-video AI models are a thing.

• Google’s Bard (its answer to ChatGPT) is available to the public (via waitlist); reviewers have been mostly disappointed in its capabilities.

• Researchers at Stanford figured out how to train a pretty capable language model (like ChatGPT) for less than $600.

Adding to this complexity is the fact that our documentation takes multiple forms. Some is mostly text-based (reports, specifications) and some is more visual (photos, drawings, presentations).

There’s a word for this concept: multimodality.

Per ChatGPT…

The term "multimodality" refers to the presence or use of multiple modes or channels to communicate or process information. In various contexts, this can involve the combination of different forms of communication, such as text, images, audio, video, and gestures. Multimodality allows for richer, more diverse, and more effective ways to convey or understand complex ideas, as it enables individuals to access and interpret information through multiple sensory channels.

In the context of AI, multimodality refers to the capability of an AI system to process, analyze, and synthesize these various forms of communication and content.

Many AI systems currently focus on one modality. For example, AI damage detectors (e.g. T2D2) “look” at images and use pattern recognition to identify issues like cracks and spalls. ChatGPT (for now—see below) works with text, but can’t accept images.

You can see how acquiring multimodal capabilities will be key to AI systems advancing to the point of being able to pull off something like I described in this earlier post.

Subscribe now

Which brings me to this week’s news…

• OpenAI released GPT-4, a new multimodal model that can work with about 8x the text volume of the previous ChatGPT model and understand the content of images.

• Anthropic, founded by folks that left OpenAI, released its own chatbot named Claude.

• Microsoft confirmed that Bing Chat runs on OpenAI’s GPT-4.

• Google announced some forthcoming generative AI tools for Workspace.

• LinkedIn is adding AI tools to help write profile content.

• Robin AI has released an AI contract editor.

• Meta’s LLM was leaked to the public and sparked a debate about open vs. closed source for AI.

…there’s more, of course, but those are some highlights. The field is moving super fast, with major updates every few days (or hours). Stay tuned!

Subscribe now

These two conferences are the premier annual gatherings for the building envelope consulting and roofing industries, respectively.

AI was everywhere at both.

Presenters in educational sessions highlighted several applications of AI they’re seeing and using in daily practice. These included:

• determining drone flight paths

• visual damage detection and quantification

I learned that Walter P. Moore is developing an in-house visual AI damage detection model. A similar service I’ve written about previously (T2D2) was originally developed inside Thornton Tomasetti.

On the trade show floor the most common AI application on display was AI damage detection. I could hardly walk 10 feet at IRE without bumping into a vendor selling this service to residential roofers.

I saw one vendor (Nearmap) that was using AI to identify various features on aerial imagery like solar panels, tree overhang, etc.

While vendors of AI services were generally bullish on the capabilities of their products (as one might expect), I was struck by the dismissive tone taken by many educational session presenters.

Most mentions of AI were quickly followed up by something like “…of course, this will never be a substitute for hands-on/human/what we’ve always done…”

Obviously no one knows what the future will hold, certainly not me. But especially given the recent trend I would caution against being so dismissive.

Honestly it felt a bit like folks were trying to reassure themselves that everything will be fine and there’s no need for scary thoughts like considering how AI could dramatically disrupt our business.

On the other hand, AI discussions were more prevalent at these two conferences than what I’ve been encountering in recent “real life” day-to-day conversations. There seemed to be a higher-than-average awareness along the attendees, vendors, and speakers.

I’m very curious to see how the discussion has evolved by the time of the 2024 editions of these events a year from now.

Subscribe now

Recent AI News

• DuckDuckGo and Brave have incorporated AI generative text into their search results.

• Slack is introducing an AI bot based on OpenAI’s ChatGPT.