Original Thinking with Christian J Ward

AI hands you more of whatever you walked in with.

I was a guest on Leadership Beyond Borders, the podcast hosted by Kimberli J. Lewis, with Stéphanie Vérilhac also on the episode. It airs around November 10.

We talked about what AI is doing to students, and I think it's the same thing behind the two camps I keep seeing in programming and engineering.

Kimberli knows I have three kids, from college to grammar school, and she asked me this.

"Do you think there's any kind of danger that this is going to make us lazy and think less? Or do you think that when you're using AI, you really do have to even think more?"

I think AI is the amplification of that which you already are.

For the very smart, high-agency individual, the person that wakes up with both urgency and agency, AI is making them better. It's making them more intelligent, with access to knowledge and access to action.

Then there's the student who woke up and decided it was going to be a video games and sit around the house day. They wanted the homework done, and they were going to cheat anyway. So they're going to cheat with AI too.

AI is an amplifier of the student, and generally of anyone.

In all fields and in all situations, you're going to see what looks like a normal distribution get pulled at both ends. It's going to pull people from the middle toward more of what they are.

If you're leaning toward resting on your laurels and skipping the hard work, it's going to pull you in that direction. If you're leaning toward starting a company at the age of 18 and taking over the world with your thoughts and ideas and ambitions, it's going to pull you to give that a shot.

Hopefully that's not too depressing for everybody, but that is absolutely what I am seeing.

There's a study that kind of confirms what I'm seeing. It's about a year old, but it makes sense. Bastani and colleagues ran a field experiment published in PNAS, the Proceedings of the National Academy of Sciences, with nearly 1,000 high school math students. Once access was taken away, the students who had used a ChatGPT-like version of GPT-4, often as a crutch for answers, scored 17% worse on the exam than students who never had it, and a tutor version that gave hints largely eliminated that harm.

We have never had a technology that adopted us, and that is what's happening. The AI is adopting us, which makes it infinitely easier to adopt it.

This is also the first technology that can explain how to best use itself.

I wish we could segment businesses on their savvy and their personal agency.

Gergely Orosz (@GergelyOrosz) wrote that he is starting to sense a divide in the tech industry. One group is "People thriving on autonomy who are reinventing how they work and what tools they use with AI". The other is used to being told what to do, and he sees them getting frustrated with the lack of guidance from management and more bitter about AI.

Aaron Levie (@levie) wrote that "It’s almost incomprehensible how wide the distance is between someone who’s working at the frontier with AI and basically everyone else right now." He also wrote that the two lines keep growing farther apart.

Peter Diamandis (@PeterDiamandis) wrote that the real AI divide will be "curious versus afraid."

Peter Steinberger (@steipete) was reacting to a Proton AI manifesto that Joseph Cox of 404 Media obtained. Steinberger wrote "Bullish on Andy Yen and Proton. He gets it." He called "the loss of identity some engineers feel as AI redefines our jobs" difficult and wrote "I sympathize", and he added that Yen "runs a business and might has limited patience when we now had two years to process".

I'm not sure the two camps are new.

I think the people embracing AI are the people who would embrace any new form of technology that could amplify them.

I also want to be optimistic here. At different times in my life I've had differing levels of urgency and agency. It's not an always on, always off thing.

I think what's fascinating about AI is the adoption cycle. With other technologies it was so long, let's say over a couple years or 10 years. With AI, the technology is changing so quickly that if tomorrow you decided you wanted higher agency and higher urgency in your career or your path, the AI you're going to adopt is so much better than the AI I've been trying to adopt for the last three years.

I think the positive news is the tool is there, and the tool gets sharper every single day. So when you do decide it's the right time for you, I think you're going to really be able to jumpstart something faster than ever before.

I understand the idea best when I can see it.

Anyone who has been on a call with me knows I stand at a dry erase board and diagram every call.

Some very astute people who worked with me for years once said the whiteboard is not for them, but for me. It helps me understand.

Ever since AI started allowing drawings, I've been trying to get it to work.

I'll explain my current process at the end of this post.

In the early days, my Venn diagram from The AI Perception Gap.

About a year and a half later, the great-looking drawing of the store without signs out front, from Bigger Pie, Thinner Slice, Better Customers.

Drawing has always been how I've thought, in my notebooks and sketch pads. My starting point was the iPad and Pencil, with the app Paper.

I tried Xnote, the one that writes on regular paper that then syncs to my phone. It was good, but the pen battery really gave up quickly. Then I could just import a PDF, but then I just started taking a quick photo and uploading it myself.

I still often take a photo of my whiteboard with an idea or concept and give it to the AI, and it does a great job now.

If you're more of a visual thinker, I recommend trying it.

What helped was having the AI do everything as code instead of using image generation tools. Image generation was too volatile, and you couldn't make edits easily without breaking the whole thing.

For a very long time I had it code in SVG. Then I switched to HTML, and now I do a hybrid in between.

But I've found that Grok's image generation really does a good job of capturing both my style and some of the more complex concepts.

It has gotten very good at making revisions without deleting other things, and text and font types used to be so difficult.

I'm now experimenting with animating the drawings, which you may have seen on some of my posts. I mostly use Grok for image drawings and Opus 5.5 for the animations.

Right now, I record my voice talking about a topic or something that I want to cover, like drawing with AI, for example, and explain kind of everything that I've done. That then goes into any one of my AI bots, but mostly I've been using Grok at this point to manage the others from there.

Grok uses Opus 5.5 on high to transform my voice notes and ask me additional clarifying questions. Once I've done those, the blog post is mostly written and checked for grammar, style and everything else against a very large series of rules and stylistic options that are trained upon my writing.

From there, we go through and look for what links would be helpful, making sure we've gotten all the details right.

Then, I feed the whole thing back to Grok to start thinking about designs for drawings. Grok will show me two or three options. From there, I usually pick one and occasionally have Claude also take a shot at drawing.

I used to make sure they were all SVG diagrams instead of just images, but I've kind of moved away from that as Grok and ChatGPT's image gen have gotten so much better at understanding the concepts of a post.

For example, that collage of every image that I've posted on my website was Grok just going out and looking at all the files in my website archive, which it has access to, and creating that collage, which is pretty amazing.

So if you want to try this, my recommendation is either Grok or ChatGPT for your drawings, and then really work on your style.

What sort of style works for you?

I've been diagramming and drawing for a really long time, so I'm pretty particular about how it looks, but I think you'd be amazed at what it can do if you have a complex concept and you just talk to it in terms of laying out a diagram to help you communicate your meaning better.

I see my journey with these tools as a microcosm of the whole process of model improvement and adoption, seen through my own usage.

It has really helped me understand that things I dreamed of these tools doing two years ago are now complete afterthoughts.

Things worth reading this week, most of them about agents doing the choosing.

Read the full issue

If you're trying to get comfortable using AI, X and LinkedIn offer plenty of examples of somebody making something amazing in one shot. And yeah, the first shot can be pretty amazing.

I rarely stop there. I'm more like ten versions, based on what I'd like it to do once I see what it has done.

This animated explainer about my writing took nine versions. I used Claude Opus 5.5 on High effort, and I kept changing what I wanted the video to be about. Here's the finished video.

To make something like this, I would have needed an artist, a composer, a narrator, a writer and probably someone to work out the storyboard. Opus wrote and ran the code on my Mac mini, wrote the narration script and composed the music. Google's Gemini speech model supplied the voice.

That's still kind of insane.

Here's how I do it. I start with an outline of a deck or some research I've done, and I give it a theme. Then I just speak the entire thing. Let's say it runs ten or fifteen minutes.

I'm basically giving the presentation to myself. I record it either with Apple Notes, which offers free transcription, or with a tool called Monologue, which I really like.

Once I have that, I take the transcript and the slides and give them to Opus 5.5 on High. I tell it to render a fully animated explanation that stops and pauses with the slide clicks, and to make sure those clicks will work with a clicker on stage.

It pretty much can do it.

That's not to say what you saw above didn't take hours of refinement. But it still would have taken literally dozens of additional hours to try to recreate those animations and that movement myself.

I can get something like 90% of the way there on the first shot. Then I have something to react to, and my requests start getting more specific.

The first storyboard mapped roughly 18 months of my writing. It showed the themes growing and connecting. I thought it was really good.

The pencil lines were too light, though. I wanted darker marks and the iceberg on screen from the beginning, with AI strategy above the waterline and data strategy underneath. The next storyboard built the video around that.

Then I changed the purpose of the whole thing. A map of my themes made sense to me because I'd written the posts. Someone arriving for the first time needed to know what I covered and what they'd get out of reading it.

So I asked for a primer. Explain the writing first, then move through the themes faster. That meant a new narration script and a different balance of what stayed on screen.

I also asked for a visible eraser, then asked for it to disappear. I wanted a calmer read, then found the result way too slow and downbeat. The model followed those instructions. I was working out my preferences by watching them happen.

The voice made this especially clear. The model could compare voices by measurement and use transcription to check that the words came through correctly. It couldn't listen to the result.

One of the voices it selected still wasn't one I liked. I asked for the earlier voice back at the faster pace.

There were technical fixes along the way too, including overlapping labels and a picture that looked blurry on my iPad. Those needed correcting. But the larger changes came from me deciding what the video should say and how it should feel.

As anyone who works with data will tell you, going from 90% accuracy to 99% is where the effort comes in. With this video, that last stretch meant getting closer to what I wanted, while I was still figuring that out myself.

I haven't tried the higher Opus effort settings yet. Maybe that would change what I get on the first attempt. I do know how much of this process came from my own changes of mind.

There's still an element of taste here, and I really don't want us to lose it. Being impressed with the first result leaves plenty of room to want a different voice, a faster pace or a better explanation of why someone should care enough to watch.

Below is the brief I could have given if I'd known what I wanted. It assumes the same source material, approved artwork, local tools and access to Google's Gemini voices. It includes decisions I made across those versions, so I can't promise it would produce the same video in one pass.

I've been using Grok Bot to edit and update my newsletter, send it out, help with social posting and keep up with the news. I now have nine bots doing different jobs, and I'm still finding things I want to try with them.

One I've just started trying is a travel bot. Before a trip, it briefs me on what's going on in the country or city I'm visiting and suggests things I might want to do while I'm there.

Another manages Codex and Claude on a machine. It isn't burning tokens itself, which I think is kind of great.

But the one I really love is Hagrid.

Yes, he's named after the Harry Potter character. We have three kids, and Hagrid watches everything to do with their school, sports and activities, including all the newsletters.

I've also given him secure access to the county school website so he can track assignments and grades. Every week, he looks at which classes the kids might be struggling in. He can then build extra practice material and test questions to help them with those subjects.

The biggest win, though, is the messaging. Between the schools, teachers, newsletters, sports teams and activities after school, there's a huge volume of information coming at us. A lot of it is badly written. Even when the information is all there, working out what we actually need to do takes some effort.

My son's swim meets are a good example.

There are events, heats, lanes and times to keep track of. The meet publishes four PDFs in different places, each with only part of what you need. One has the event list with estimated start times, another has every swimmer's prior times, and another has lane assignments. They're all in different formats.

So you have parents walking around the pool with 32 pages of highlighted paper, trying to figure out where their kid needs to be and when. It's easy to miss a heat while you're sorting through it all.

I asked Hagrid for a super simple screenshot I could text to my wife and son. He pulled everything we needed together into one image, and I texted it to them.

It even includes the practical stuff, like what to bring and the fact that there's no concession stand. The times are marked as estimates, and it notes that the heat sheet could change. I'm grateful for something that can take all those scattered details and make them easy for our family to use.

Giving an assistant access to school grades raises a real privacy concern. I'm not dismissing it. There are plenty of doomers already, though, and the benefit here is real.

Research on the privacy paradox describes the gap between what people say about privacy and what they actually do. People keep using services they worry about when the benefit is worth it to them. Even many of the loudest privacy critics live on Instagram and aren't about to give it up. I understand the appeal.

There's an AI assistant called Instinct that I haven't tried yet, but everyone keeps telling me it's amazing. I'm curious to give it a try.

With Grok Bot, one thing I love is that the agents have a virtual machine in the cloud to work on. That solves the Mac mini problem, where people buy a Mac mini just to host agents and keep them running all the time. I still like running my Mac mini with other coding setups on the side, but I appreciate having the computer provided as part of this.

The other thing is how well the bots work together as a team. My coding agent talks to my chief of staff, which gets information from Hagrid and works with my travel agent to help plan my calendar. The handoffs work well. With nine bots, being able to have them talk to one another makes a real difference.

Tesla is also making it possible to talk to your Grok Bot while you drive. You can get a lot done that way, and talking to this same team from the car means the drive doesn't have to be wasted time.

Alex Finn shared a video of himself riding in his Cybertruck and talking to his agents after the Grok Bot for Tesla release. It gives you a look at what that feels like.

If you haven't tried Grok Bot, it might be worth giving it one of those jobs you keep having to piece together yourself. I've been pretty happy to hand over the swim meet PDFs.

Yesterday, I bookmarked Google's announcement on X. Publishers can now add a Preferred Sources button to their pages. Click it and you tell Google that you want to see more from that site.

As News from Google shared on X.

Google says more than 600,000 unique sources have already been selected. Preferred sites can appear more often in Top Stories, AI Overviews, and AI Mode. A related control coming to Discover lets people type which topics or links they want to see more or less of. The feed adjusts and remembers.

That gives Google a direct declaration of taste. I choose a source and say, in effect, I prefer this.

Then I go back to searching and clicking, and what I actually read may tell a different story.

Economists have studied this gap for nearly ninety years. Stated preference is what someone says they want. Revealed preference is what their choices show when something is at stake.

Paul Samuelson formalized revealed preference in 1938 and developed it again in 1948. His claim was that we can infer preference from choices rather than talk. A survey records what I say I would buy. The market records whether I hand over the money.

Revealed preference still needs care. John Beshears, James Choi, David Laibson, and Brigitte Madrian showed why a choice does not automatically reveal what is best for someone. The choice may be passive, complex, unfamiliar, heavily marketed, or separated from its consequences by time.

Digital privacy research gives us a more recent example. In an experiment by Susan Athey, Christian Catalini, and Catherine Tucker, stated concern about privacy gave way when small incentives or navigation costs entered the decision. Daniel Solove has argued that this gap often reflects cost and context rather than hypocrisy.

People share more with AI than they type into a single-line search, and they often do not realize they are adding context. Some of that context can become memory.

I have written before about why memory needs to forget and why AI memories need an expiration date. A memory can accurately preserve what I preferred last August and still be a poor guide to what I will choose today.

Once an assistant can follow more of the chain, it has to decide when a preference has gone stale, when behavior changed because of friction, and how much weight each signal deserves.

I have no problem with people stating their preferences. People are going to consume content the way they want, and a button that lets someone support a source they value is a reasonable product choice.

The filter bubble question has been with search and feeds for years. Accumulated preferences can reduce the number of unfamiliar sources that get a chance to appear. More memory and context can make the experience feel personal while carrying yesterday's preferences into tomorrow's answer.

Google is trying to understand demand for different content sources and how to cite them. Preferred Sources gives people a direct way to express that demand. The question I keep coming back to is how memory should respond when the declared preference and the later behavior diverge.

An assistant that remembers the conversation and follows the work can treat preference as a moving relationship among what I click, buy, pay for, and choose over time. That could produce a deeply personal experience. It could also turn a temporary pattern into a lasting assumption about me.

I want memory to learn without becoming a filter bubble of one.

Quick thought on today's news, principally the GitHub and Cursor thing.

On Monday, August 17, Cursor shipped Origin, its new code hosting platform. The early beta begins with repositories, pull requests, code browsing, and GitHub sync. It is built into Cursor, and it arrived on a day when code hosting already had everyone's attention.

As Cursor shared on X.

Since I started using Grok Bot, Cursor is looking more and more interesting as a tool to partner with. I wrote yesterday in Zero to Chief of Staff about how quickly Grok Bot gives an agent a computer and somewhere to work. Origin puts the codebase closer to the editor and the agents that are already working on it.

Cursor's Mirror tool is what I will add first. Mirror copies a GitHub repository into Origin and keeps the copy updated as the GitHub repository changes.

GitHub stays the source of truth. Pushes to the synced repository still go to GitHub, and pull requests can sync in both directions.

That gives me a calm way to experiment. GitHub keeps doing the job it already does, while Cursor gets an Origin copy close to its code browsing, pull requests, and agents.

The timing belongs on a longer historical line. The Trojan Horse worked after the Greeks seemed to leave, Washington crossed the Delaware on Christmas night before Trenton, and Origin appeared while code hosting was on everyone's mind.

Matt Palmer carried the timing joke in his quote-post on X.

I am going to add Mirror and see how Origin feels alongside the tools I already use. GitHub keeps its current role. Cursor gets closer to the repository, the agents, and the work.

After the last few days with Grok Bot, that sounds like a partnership worth exploring.

I keep thinking about Grok Bot and how easy it is to set up. The use of the virtual machine is incredible.

On Tuesday, August 11, SpaceXAI released Grok Bot in early beta. The virtual machine is part of the product from the start, so you are not assembling a computer around the agent before you can use it. Krista Letz, who leads Grok Bot go-to-market at SpaceXAI, was one of the people sharing the early launch.

Then on Wednesday, Anthropic put Cowork in the Chrome side panel, and on Friday, Cursor joined SpaceX.

That was one week. They were different releases, but each one shortened the distance between asking an AI for help and giving it somewhere to work.

OpenClaw was the hardest. Hermes was easier, but still complex.

When I built my personal AI agent on OpenClaw in February, the setup took the full long weekend. The result was an always-on agent connected to my notes, calendar, voice memos, and task system. Getting there required a Mac Mini, careful migration, and a lot of mistakes.

With Grok Bot, the computer is already there. The agent can begin from a virtual machine instead of asking you to build one first.

You go from zero to an incredibly helpful assistant, what they call your chief of staff, and that is probably one of the most interesting things I've found about it.

Yes, they can be expensive, and it adds up. But the people who looked at OpenClaw or Hermes and saw a technical project can now begin with a setup screen and a computer that is ready for the work.

I wrote last year that conversational interfaces break Jakob's Law because conversation does not ask us to learn another menu.

An affordance is what a thing invites you to do. A handle invites a pull. A chat box invites a sentence. Cowork invites you to hand the desk a job and walk away. Grok Bot invites you to start from zero and have a computer.

Claude has chat, Cowork, and remote, while ChatGPT has chat, Voice, and remote. These are the same underlying platforms. The affordance of each approach is really different.

Chat keeps the work inside a conversation, Cowork begins with a job and a result you want back, and remote lets the session continue while you leave the desk. Grok Bot begins by giving the agent a place to work.

ChatGPT Voice stands on its own as still one of the most dynamic things.

Alex Finn wrote that he spent a four-hour hike through the northern California redwoods talking to ChatGPT Voice through his AirPods, and got more work done during the hike than he usually completed in eight hours at his desk.

I recognize the feeling. In March, I wrote about the features that untethered me from my desk while I was walking through my town and speaking a post into my phone.

It is still not necessarily connected the way you can with Grok Bot to lots of other tools and systems, in a simple manner.

Voice lets you talk while you walk, drive, or sit in a cafe. With Grok Bot, you can hand that conversation to a computer with tools already around it.

One feels like a thought partner you can take anywhere. The other feels closer to setting up a member of staff.

In June, I wrote that I do not want to be a user anymore. A user works inside whatever screens a software company decided to ship. A principal sets the intent, gives an agent somewhere to act, and keeps the judgment about whether the work is right.

As the setup gets easier, people get to that second experience sooner. They do not have to begin by learning OpenClaw, configuring Hermes, or building a personal server. They can begin with chat, Voice, Cowork, remote, or a virtual machine that is already waiting.

I just see this migration. As more and more capabilities are just easy to set up, more and more people are going to start discovering how good Claude, ChatGPT, and Grok Bot actually are.

A person can now click through a setup, give the bot a job, and find out what a chief of staff feels like before they know what OpenClaw is.

I am writing this post by talking to Hermes, an agent I set up on my own machine. I talk, it writes.

The way I am writing it is the argument.

No app. No dashboard. No screen a product team built two years ago, guessing what I would need today.

There is a lot of debate right now about whether MCP is the right level. Whether customers and businesses can realistically wire up their own connections, since plenty of them are not at the level of coding up their own things.

That barrier comes down quickly as people get more exposure. It was never really the issue.

Where this heads is that businesses build AI that leverages multiple MCPs. Their own, their vendors', their partners'. The plumbing is already reaching normal users too. Claude artifacts can now call MCP connectors, so a small app can fetch information and take actions for each viewer on demand.

When every company can call the same connectors, the connectors stop separating you.

The data surface is still the real advantage for most companies.

If I am a marketer selling to roofing companies, I have to understand how roofers actually work. They are in the field all day, doing estimates and repairs. A roofer halfway up a ladder is never going to deal with an MCP.

But there is an entire process where that same roofer would love to talk to an AI.

I watch the same divide in my meetings. The people in the room who have set up an OpenClaw or a Hermes get it immediately. The people who have not are still staring at the setup, and the setup is a big barrier to understanding how easy all of this is once it is running.

We have been watching this adoption timeline for a couple of years.

People built their own dashboards, ones they were comfortable with. Then they learned what most dashboards and most software eventually teach you. They are far too rigid for how fast data is changing. And mixing and matching data across sources makes the classic dashboard incomplete the day it ships.

I wrote about software dying last fall, and in April I argued the API is the UI now. This is the same line, extended one more step.

Although people still need apps, they probably don't.

People who have tried Claude or Hermes start to see a real opportunity around conversation here. I do not think everyone else has seen it yet.

Now tie that to what happened this week in open source.

Thinking Machines launched Inkling. Full weights, multimodal reasoning, and the ability to fine-tune it into something specific. SpaceXAI open-sourced Grok Build. Muse Spark is out. Sriram Krishnan counted the reasons in one thread. Near-SOTA performance with clear training lineage. Organizations that want control over how their data gets used. Token costs ballooning without a clear line to revenue.

Jason Calacanis rolled it into one phrase, "the AI sovereignty stack." Open source enthusiasm, a dramatic drop in the cost of intelligence, and personal hardware capable of running models locally.

I would add one more reason. Companies are going to lean toward open models created, trained, tuned, and provided out of countries aligned with their own, or with a similar international relationship. Model choice is turning into a question of where you sit and who you trust.

So picture the split. Frontier models for orchestration and customer service, where you want the best available. Open models on your own hardware for everything close to the core, trained to focus on your own things, and they save you real money because the hardware is yours.

I think that is what starts to happen, quickly.

The roofer is never going to open your dashboard.

He is going to ask, and an AI is going to answer, through whatever mix of models and connectors his suppliers wired together while he was on the roof.

Whether that answer is any good depends on the data it stands on. That part has not changed, and I do not think it will.