The Innodative Disruptor

Accelerating towards Obsolescence

2026-05-01T00:00:00-04:00

Your ethics are your ethics.

The line hit me hard. I had been describing how I use AI in my work, the actual mechanics of it, what I hand off and what I keep, where the tool helps and where it gets in the way. And this was the response. It was ambiguous, foreboding, confusing.

Why is my use of AI being framed as an ethical position in the first place? Not a methodological choice, not a workflow decision, not a question of craft. An ethical one. With the implication, sitting just under the surface, that a verdict is pending and the wrong answer, whatever that is, has consequences.

Thus, I felt compelled to write. Because I think this quote explains why fields are stalling out while the world accelerates away from them.

When confronted by AI, many see the road diverging into two paths. Ban it or delegate to it.

Ban it because the technology is unreliable, or threatens the craft, or produces work that isn’t really yours. And underneath those reasons, often, an inherent resistance to change. That isn’t the way we did it in my day. We must preserve what made the field worth practicing in the first place.

Delegate to it. Hit the “easy button.” Paste in the prompt and accept the output. This is now called cognitive offloading, when the thinking gets outsourced.

For many, the choice really is binary, and the AI conversation aligns with these two paths, as if those were the only directions available.

My own field faced this challenge when ChatGPT arrived in late 2022.

Three years later, most institutions still do not have a coherent position. Some ban, some allow, some leave it to individual instructors, some mandate disclosure without defining what disclosure means. But students kept using the tool throughout, because the tool kept getting better and the work kept being assignable. What academia did not generally do was figure out what good practice looks like.

But there is a third option: AI as a collaborator.

You may ask, “What does this even mean? How do you collaborate with a tool?”

The answer is amazingly simple. Just like any craftsman, you experiment, you learn, you refine, you improve, until one day the tool is a seamless part of your workflow.

This option rarely appears in most conversations about AI because it is harder and more nebulous than the first two. It is not a position you can declare. It is a practice, and practices are harder to talk about than postures.

But it works, just look at software development.

Coders were among the first to feel real pressure from AI, and they were among the first to figure out how to work with it. The model that has emerged is orchestration. The developer no longer types every line, but they also don’t look for an easy button. They specify what needs to be done, delegate tasks to the AI, monitor what is built, review the results, and move on to the next piece.

The skill is not just coding in the old sense. It is knowing what to ask for, knowing what good looks like when it arrives, and knowing when to throw something out and try again. The developer’s judgment is doing more work than ever, and they do more with less typing.

This is craftsmanship. It is what a working path forward looks like when a field starts to figure out how to collaborate with the tool.

This is also how I work, and how I teach.

AI is part of nearly everything I do. The model is a thinking partner, not a ghostwriter. It pushes back, it gets things wrong, it surfaces angles I had not considered, and our back-and-forth is where the work actually happens. What I publish is mine. But it results through collaboration.

This is what I teach students. We don’t ban the tool and pretend it isn’t there. We also don’t hand the assignment over and accept whatever comes back. Instead, they use this tool the way a professional in their field is expected to use it now. The pedagogy I have arrived at is straightforward: work with AI, critically evaluate what it produces, and form your own insight. The skills that matter are the ones that remain when the tool changes. And it will.

I see journalism following the same binary paths. Wired published a near-blanket ban⊕Wired’s policy bars AI-generated or AI-edited text and AI-generated images, with narrow carve-outs for headline suggestions and idea brainstorming. How WIRED Will Use Generative AI Tools, by Gideon Lichfield. on AI-generated text in stories, with carve-outs for things like headline brainstorming and story idea generation.

At the other end, outlets with thinner resources have leaned in hard, publishing AI-generated articles with minimal review (what is derogatorily called “AI slop”). The worst offenders have handed the ban-it crowd exactly the evidence they sought.

In between them sits the disclosure regime. Outlets and institutions mandate that AI use be declared, but no one has defined what exactly “AI was used” actually means. Used to research? To draft? To check a fact? The vagueness becomes a loophole. People can skip disclosures or shade them, and feel fine doing so, because they are ambiguous.

Which brings us back to the starting line: Your ethics are your ethics.

When scared about job security, people look for moats. A defensible position that protects the work, the field, the profession. If they can’t find one, fear will often drive them to invent one.

I see the ethics framing as one such invention. The argument runs: AI use is an ethical question, my answer is to limit it, and if everyone agreed with my answer the field would be protected. Ban AI in journalism and journalism stays journalism. Ban it in the classroom and the degree still means something. We have a moat.

But that moat only works if everyone honors it. And the moment you invoke individual ethics, you open the door to them making different decisions. Some will tell a white lie on disclosures. Some institutions will carve out exceptions. Many will skip disclosure entirely because no one has settled on what disclosure means. The imaginary moat dissolves the moment anyone steps across it, and people are stepping across it constantly.

What is left is the field, undefended, with practitioners who refused to engage with the tool now competing against practitioners who did.

This is the pattern I keep seeing. Professions whose job is to interpret the world for everyone else flinching at the moment they should be leaning in. Treating the most important tool of their working lives as an ethical hazard rather than a skill to develop. Building moats out of moral language and watching the moats dissolve.

Your ethics are your ethics. Fine, I get it. But the corollary is your field is your field. And the question is not whether you approve, but whether you cling to the past or lead the way forward.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

The Economics of AI Anxiety

2026-04-24T00:00:00-04:00

Anxiety about AI-driven job cuts is everywhere, but badly misdiagnosed.

Block recently slashed 40% of its workforce. Snap cut 16%. Amazon shed 30,000 jobs in a matter of months. Leading AI CEOs from Anthropic’s Dario Amodei to OpenAI’s Sam Altman to NVIDIA’s Jensen Huang have warned that AI will eliminate many jobs.

But look beyond the layoff theater and a different picture emerges. Companies over-hired during the pandemic and are now correcting. As one executive told The Wall Street Journal, AI has provided “air cover” for cuts that were coming anyway. AI is the convenient villain in a story that was already being written.

AI displacement, however, is not entirely fiction. Goldman Sachs finds that workers displaced from technology-disrupted roles take longer to find new jobs and accept earnings losses.Reported in The Wall Street Journal. Anthropic reports measurable slowing in hiring for AI-exposed occupations, particularly among workers aged 22 to 25.Anthropic Economic Index, ‘Labor market impacts of AI’, March 2026. The pain is real.

Still, reading past the headlines reveals the precise pattern: the labor market is absorbing AI through task-level substitution, not wholesale job elimination.

The gap between that reality and the panic comes down to how we experience AI versus how economies actually absorb it.

I teach AI for BusinessACCY 593 is offered for credit through Gies’s online programs, including the iMBA. The video lectures and some basic assessments are also available through two Coursera MOOCs: Introduction to AI and Advanced Topics in AI. to MBA students at the University of Illinois Gies College of Business, and the confusion often starts with experience. Former Tesla AI chief Andrej Karpathy recently observed that two groups are talking past each other. One group tries basic tools, sees them fail, and concludes AI is overhyped. The other uses frontier models in professional settings and is stunned by their capabilities. Both groups are right, but both are missing the full picture.

AI systems are remarkably powerful, but mostly in domains where success can be clearly measured. Did the code compile? Did the test pass? Yet much of our work, including persuasion, leadership, strategic judgment, and navigating ambiguity, doesn’t offer that kind of clarity. This is not a temporary limitation. It is core to being human.

Venture firm Andreessen Horowitz finds that coding dominates enterprise AI adoption by an order of magnitude over any other use case—one of the few domains with clean, verifiable outputs.

But basic economics suggests that even where AI is effective, it won’t necessarily reduce work. When technology makes a task cheaper, we tend to do more of it, not less. Economists have understood this dynamic, Jevons’ paradox, since the 19th century. Spreadsheets didn’t eliminate accountants; they expanded the scope of financial analysis.

Now we see this with AI. Engineers at Meta burned through 281 billion tokens in 30 days on an internal leaderboard before it was taken down. Visa employees consume nearly two trillion tokens per month, reflecting millions of AI-assisted interactions across their workforce.Meta data: Fortune, April 2026; the dashboard was shut down shortly after going public. Visa data: reported by Computing, citing Business Insider. Even Anthropic is hiring video directors up to $250,000.Video Director, Product Launches, salary range $200,000–$255,000. These are not stories about workers being replaced. They are stories about workers using dramatically more AI to extend what they already do.

So where is the disconnect? There are three major constraints that are missing from the headlines.

The first is scale. Running advanced AI systems requires enormous energy and infrastructure. U.S. data centers already consume about 4.4% of the nation’s electricity, potentially rising to 7–12% by 2028.U.S. Department of Energy / Lawrence Berkeley National Laboratory, 2024 Report on U.S. Data Center Energy Use.

But this expansion is meeting resistance. Last year, 25 data center projects were canceled following local opposition.Heatmap News tracked the cancellations through 2025. Amazon, Microsoft, and Google have all abandoned multibillion-dollar builds. In Illinois, recent polling found that 80% of likely voters either want strict regulation or oppose new data centers entirely.Rich Miller, Chicago Sun-Times, April 18, 2026.

Yet, even as new capacity comes online, demand quickly absorbs it. Technology analyst Benedict Evans notes that similar infrastructure constraints in mobile networks took multiple decades to resolve.

That is not a software challenge. It is a physical one.

The second constraint is pricing. The AI industry struggles to assign value to what it sells. Outcome-based pricing works when value is easy to measure, payment processing, for example. But in hiring, management, or organizational decision-making, AI’s contribution is real but difficult to isolate. Someone still has to take responsibility for the decision. That accountability can’t be automated.

The third is specialization. The principle of comparative advantage has held for two centuries: specialization persists even when one party is better at everything, because what matters is opportunity cost. Compute is expensive and finite. Firms will allocate AI to their highest-value problems, leaving other tasks to humans. Not because humans are superior, but because the AI’s time is more valuable elsewhere.

Together, these constraints provide our guarantee.

My students aren’t just worried about their own careers. They are asking what this means for their children, whether college still makes sense, whether traditional career paths will exist a decade from now.

My answer is the same thing I said five years ago: pursue your passions, build expertise, stay curious. The economics haven’t changed that advice, they’ve reinforced it.

I don’t minimize what’s coming. Workers in routine, AI-exposed roles will face real disruption. But the broader trajectory is clearer than the headlines suggest.

AI capability remains narrow. Demand expands as costs fall. Infrastructure limits the pace of deployment. Pricing reveals the continued centrality of human judgment. And as long as compute is costly, human roles will persist.

We are not heading toward a world without work. We are heading toward a world with different work. The sooner we understand that, the better prepared we will be.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

My Teleprompter that Claude Built

2026-04-22T00:00:00-04:00

I’m recording audio narration for a new blockchain course. The scripts live in Pages, complete with production callouts like [ON SCREEN: diagram of consensus mechanism] and [B-ROLL: server farm footage]. Reading them naturally while recording turned out to be harder than expected. Scrolling a document manually breaks pacing. Losing your place means starting over. Commercial teleprompter apps exist, most priced for broadcast studios rather than someone recording a few hours of course audio a week.

During a Claude conversation about the course, I mentioned the problem. Claude offered to build one.

A few iterations later, I had a single HTML file that did exactly what I needed.

What it does

Paste a script from Pages, click one button, land in prompter mode. Production markup is stripped automatically, so [ON SCREEN], [VISUAL], [B-ROLL], and the table borders Pages sometimes injects disappear before the text hits the screen. The first line becomes the video title. A faint red guide line sits 38% from the top. That’s where I read.

Keyboard shortcuts: Space plays and pauses, up and down arrows adjust speed, Escape returns to the paste screen. Mouse wheel scrolls manually when paused.

Auto-scroll runs from 0.5x to 5x, adjustable live by slider or arrow key. Font steps from 28 to 56 pixels. Dark background, Georgia serif, generous line height. No dependencies, no install, no subscription. It runs in any browser, including on an iPad propped next to the mic.

The iteration

The first version worked but treated the raw paste as narration, including all the production callouts. Version two stripped them. Version three caught the table borders. Keyboard shortcuts came next. The guide line started at the center and migrated up to 38% after testing it against an actual script.

Each change took a sentence to describe. The whole thing happened inside a single chat.

The pattern

The teleprompter itself isn’t the point. The pattern is.

Most professionals have small friction in their workflows. Tools that almost exist. Formats that don’t quite fit. Annoyances absorbed so long ago you’ve stopped seeing them. The old answer was to live with them, search for an app, or ask IT.

The new answer is to describe what you want and get a working tool in minutes.

That shifts the economics. A custom tool used to be a project, which meant clearing a bar: worth a developer’s time, worth a budget line, worth the wait. Most friction points never cleared that bar. They got absorbed.

When the tool takes a conversation to produce, the bar moves. Things that weren’t worth building become worth building.

Try it

The teleprompter is live at /tools/teleprompter/. Paste a script and see what happens. Save the page to use it offline for your own recording. There’s nothing to connect to.

If you want to try this pattern yourself, pick one small friction point from your own week. Describe it to an AI. Iterate until it fits. Mine took a handful of rounds and no code on my part.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

Paper Records as Unhackable Backup

2026-04-20T00:00:00-04:00

This month Anthropic announced Mythos, an AI model the company says is too dangerous to release publicly. Scientific American, ‘What is Mythos and why are experts worried about Anthropic’s AI model.’ Mythos can autonomously find zero-day vulnerabilities across every major operating system, chain them into working exploits, and cover its tracks. The UK’s AI Security Institute found it succeeded on expert-level hacking tasks 73 percent of the time. No AI before April 2025 could complete those tasks at all.

Access is being limited to a handful of tech and financial firms under Project Glasswing, and the Treasury Secretary convened senior US bankers this month to discuss it. Euronews, ‘Hackers breach Anthropic’s too dangerous to release Mythos AI model,’ reporting on both the third-party breach and the Treasury meeting. The concern isn’t just nation-state adversaries. Mythos itself was already breached through a third-party vendor. The next version, or someone else’s equivalent, will land in more hands. When that happens, financial infrastructure will be exposed in ways it never has been before.

Most of what I own lives on a server somewhere. Bank accounts. Investment records. Mortgage documents. Tax history. Personal devices matter too. A compromised laptop can give attackers a path into my accounts. But that surface I can harden. What I can’t control is whether the institutions holding my money and assets stay secure. Each has defenders. The defenders have been adequate so far because attackers were human and slow. Mythos is a preview of what happens when they are neither.

There is one class of record no remote AI can touch: the one printed on paper and sitting in a drawer. It doesn’t have to be everything. Just enough to reconstruct the truth if digital records are corrupted, frozen, or held for ransom. Account numbers. Recent statements. A list of holdings. Beneficiaries. Paper serves as proof—in a world where an institution’s systems may be down or lying—that I am who I say I am and that what I own is what I own.

For a long time, going paperless was the obvious choice. More convenient. Lower costs. Saving Trees. The assumption was that digital was safer than paper because it had backups. But digital backups are still digital. They can be corrupted, encrypted, or deleted alongside the originals. Mythos sharpens what was already true. Paper isn’t safer because it’s better. It’s safer because it’s not on the network.

I’ll be printing a few things now. I suggest you do the same.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

Who Pays for Open Weights?

2026-04-19T00:00:00-04:00

Frontier AI models cost money to train. Someone has to pay for the compute, the researchers, the data. For the past year and a half, a puzzle has hung over the industry: why were Chinese labs releasing frontier-tier open weights for free? I’ve been one of the beneficiaries, running Qwen and DeepSeek locally on my Mac Studio for work where I want privacy or throughput.

This month the puzzle started resolving. The short answer is that they aren’t anymore.

Alibaba released Qwen3.6-Plus as a closed hosted offering on Alibaba Cloud, keeping only its smaller models (35B and below) as open weights, now positioned as a developer acquisition funnel rather than a frontier release. The Information, ‘Alibaba Becomes Selective with Open-Source Models, New Release Shows,’ April 2026. See also SCMP, ‘Chinese AI giants pivot toward proprietary models.’ Z.ai rolled out GLM-5-Turbo closed. VentureBeat, ‘Z.ai debuts faster, cheaper GLM-5-Turbo model for agents.’ DeepSeek, famously self-funded by the hedge fund High-Flyer Capital and known for turning down outside money, is now seeking \$300M or more at a \$10B valuation, citing researcher departures. The Information, ‘China’s DeepSeek is Raising Money for First Time,’ April 17, 2026. ByteDance’s Seedance 2.0 and Kuaishou’s Kling 3.0 are both proprietary from the start.

A recent ChinaTalk analysis lays out the underlying economics. Chinese labs need revenue, and the DeepSeek shock got open source “a moment” rather than a sustainable business model. The funding environment for Chinese AI is orders of magnitude smaller than America’s. Gulf capital put roughly \$100M into Chinese labs while pouring roughly $15B into Anthropic and OpenAI, and Western venture money is almost exclusively American. Figures from the ChinaTalk analysis cited above. For corroboration on the scale of Gulf AI investment in US labs, see Bloomberg, ‘Abu Dhabi’s MGX Targets \$100 Billion in AI With OpenAI, Anthropic Investments.’ US export controls have tightened steadily, narrowing access to the chips Chinese labs need to train at frontier scale, which means every training run costs more than it does for a US competitor. The Chinese government has been willing to subsidize domestic hardware but not open model development, and domestic chips still trail NVIDIA by a meaningful gap.

Several other pressures compound the squeeze. American labs have documented large-scale adversarial distillation from Chinese labs going back to 2024 and are actively closing that channel through the Frontier Model Forum. Every Chinese model still has to pass a post-training alignment layer for CCP compliance, a tax Western labs don’t pay. And Alibaba’s own Qwen technical lead recently put the odds of a Chinese firm surpassing US tech giants in three to five years at under 20 percent. SCMP, ‘China AI has less than 20% chance to exceed US over next 3 to 5 years: Alibaba scientist,’ January 2026. Lin Junyang has since stepped down from the Qwen team.

Not every Chinese lab is making the same bet. This week Moonshot AI open-sourced Kimi K2.6. Moonshot AI, Kimi K2.6 release, April 20, 2026. See also MarkTechPost, ‘Moonshot AI Releases Kimi K2.6 with Long-Horizon Coding, Agent Swarm Scaling to 300 Sub-Agents and 4,000 Coordinated Steps.’ Moonshot is well capitalized, with significant backing from Alibaba and Tencent, but what makes it different is business model. Kimi runs as a consumer subscription and enterprise product, with a cloud service hosting its OpenClaw agent on top. Open weights are marketing for a business that sells something else. Whether that survives sustained margin pressure from closed Chinese competitors is a separate question, but as of this week it is a genuine third path, not a niche.

The open-weight ecosystem itself will survive. NVIDIA has committed \$26 billion over five years to the Nemotron family. Disclosed in an NVIDIA SEC filing March 11, 2026, first reported by WIRED. See also Nemotron, Wikipedia and Decrypt, ‘Nvidia Drops Nemotron 3 Super Amid \$26 Billion Open-Model AI Bet.’ Meta continues Llama. Google has Gemma. The Allen Institute keeps pushing OLMo and Tülu forward. But the providers are changing, and the incentive is changing with them. The generous era of Chinese frontier labs releasing weights for mindshare is ending. What replaces it is American hardware and platform companies using open models to sell chips and cloud services, and Chinese labs like Moonshot using open weights as the foundation for a wrapped business. The models I run locally today will keep working. The next generation of frontier-tier open weights will increasingly come from a different set of addresses.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

Most of the Internet Isn't Human Anymore

2026-04-18T00:00:00-04:00

Earlier this month, Lumen’s CEO Kate Johnson published an open letter to fellow CEOs noting that more than half of internet traffic is now generated by AI agents and bots rather than people.Lumen carries roughly 65% of global internet traffic, so this is observational data from the infrastructure itself, not a projection.

Johnson’s framing: “More than 50% of the traffic on the internet today is created by autonomous workers. That’s remarkable because as most CEOs would tell you, we are just beginning our AI journey.”

Other sources corroborate. Human Security’s State of AI Traffic report found automated traffic grew nearly eight times faster than human activity in 2025, with AI traffic specifically up 187% year over year.Human Security CEO Stu Solomon: "The internet as a whole was created with this very basic notion that there’s a human being on the other side of the computer screen, and that notion is very rapidly being replaced." Cloudflare’s CEO Matthew Prince predicted at SXSW that bot traffic will exceed human traffic overall by 2027, noting that before the generative AI era, bot traffic was only about 20%. The 2025 Imperva Bad Bot Report marked 2024 as the first year automated traffic crossed the 51% threshold.

The tipping point has already passed. The majority is already non-human, and every measurement suggests the share keeps growing.

The easy interpretation is that bots are taking over the internet humans built. Infrastructure designed for people, now dominated by machines. That framing isn’t wrong, but it may miss what’s actually happening.

A more interesting possibility: the internet isn’t being subsumed. It’s forking.

The internet humans use was built around human constraints. Screens, reading speed, deliberation, the cognitive overhead of deciding whether to click. Pages optimized for attention. Forms designed for typing. Checkout flows paced for second-guessing. That infrastructure makes sense for the audience it was built for, and there’s no particular reason to change it.

What may be emerging alongside it is a parallel layer optimized for a different audience. Agent-to-agent communication doesn’t need rendered pages, visual hierarchy, or CAPTCHAs. It needs structured data, predictable schemas, and protocols that handle trust and transactions at machine speed.MCP (Model Context Protocol), agent handshakes, and machine-readable endpoints are early examples of this second layer getting built.

Bots are fluent in both. An agent can scrape a human-facing page when that’s what’s available, or talk directly to another agent when the plumbing exists. Over time, the bot-to-bot exchanges have no reason to route through the human internet at all. They’ll use the layer built for them.

If that’s what’s happening, the traffic statistics are measuring something slightly different than they appear to measure. The 50% isn’t bots crowding humans out of a shared space. It’s bots using whatever infrastructure is available while a second infrastructure gets built underneath. The share of “bot traffic on the human internet” may eventually peak and decline, not because bots go away, but because they migrate to their own layer.

Two internets, coexisting. One optimized for human cognition. One optimized for machine throughput. Humans stay on the first. Bots operate on both for now, and increasingly on the second.

What’s worth watching is where that second layer is getting built, who’s building it, and what standards emerge. The human internet took decades to shape, and its conventions are now deeply embedded in how business runs. The agent internet is being designed in real time, and its conventions will shape the next era of business the same way.

The stat is the signal. The fork is the thing to watch.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

A Signal Becomes Real

2026-04-17T00:00:00-04:00

A major part of my job as Chief Disruption Officer is to watch for signals, early indicators that something is shifting in ways that could matter to higher education and business. For years, I’ve discussed potential threats with deans to the traditional higher ed model. Mostly these conversations were couched in maybes. The technology isn’t ready, the vendors aren’t credible, or the employer buy-in isn’t sustaining.

Last week, a maybe became real.

Khan Academy, TED, and ETS announced the Khan TED Institute—a competency-based bachelor’s degree in applied AI for under $10,000. While this announcement alone would be interesting, what I feel makes it a genuine signal is the list of corporate partners who are shaping the curriculum and committing to recruit its graduates: Google, Microsoft, Accenture, Bain, and McKinsey.

Why? The perennial advantage of traditional universities, especially elite ones, has been the networking and the recruiting pipeline. You go to a top school not just for the education but because that’s where the employers go to hire. Sal Khan himself said it plainly: “McKinsey recruits at Harvard, and they don’t recruit at my state university.” Sal Khan, speaking at Charter’s Leading with AI Summit, February 24, 2026. Reported in The San Francisco Standard, February 26, 2026.

The Khan TED Institute looks like a credible attempt to build that pipeline from scratch, outside the traditional system entirely.

Now, to be clear, I don’t see this as an immediate threat to my employer. The University of Illinois, and specifically Gies College of Business, have been around a long time and have global name recognition. We have brand, scale, research, and deep employer relationships that take decades to build. But, if successful, this endeavor likely exacerbates the pressures already impacting smaller colleges, which are staring down a demographic cliff and real, honest questions about ROI. A sub-$10K degree with McKinsey’s name attached changes that calculus for a lot of students and families.

Clayton Christensen spent his career showing us how disruption actually works. It doesn’t start by attacking the incumbent head-on. It starts by serving the customers the incumbent isn’t serving well. In this case, learners who can’t afford $100K or more for a degree, who need flexibility, who want to learn AI skills but don’t have access to a top program. As Clayton demonstrated, new entrants take the lower end of the market, and then grow. By the time the incumbent takes it seriously, the window for response has narrowed or maybe closed.

These are not the sort of signals we can just ignore.

Nobody Said It Was Easy

2026-03-04T00:00:00-05:00

I may be biased, but I think my AI for Business course is a good course in our online portfolio. We average close to 400 students every semester, outcomes look solid, and evaluations are positive overall. But those same evaluations highlighted some real issues. Several videos felt repetitive. Ethics content showed up in four different modules. And students wanted more technical depth, not less.

We designed and built the course two years ago. Now, generative AI barely resembles what it looked like then. A refresh is clearly needed, but how and what? I was reminded of a line from the Coldplay song, The Scientist: “Nobody said it was easy.”⊕Coldplay—The Scientist (2002). The reverse-narrative video is worth watching on its own terms.

Yet, as I thought about how AI has changed over the last few years and the current course structure, I was reminded of the next line in that song: “No one ever said it would be this hard.” Eight modules with 120 videos, each around eight minutes long, a field that’s moved dramatically, student feedback pulling in different directions. Where do you even begin?

Rather than losing myself in the enormity of it all, I felt I needed to return to my roots and be a scientist. Run an experiment. See what happens.

So, like I am fond of telling others, I sat down with Claude and started a conversation.

The Experiment

Not one prompt, but an extended back-and-forth over about two hours total, spread across a few days. I’d work through a section, think about it, come back later and pick up where we left off. Most of the time was spent reading and thinking, not typing.

I fed it the current course structure and anonymized student evaluations, and we worked through the problem together.Student evaluations are anonymized when faculty receive them. No student personally identifiable information was used at any point in this process. The AI asked clarifying questions before proposing anything. I set the constraints: the course needs to be accurate, timely, and future-proof for two to three years. Budget matters. Fewer changes means less money spent, but sound pedagogy was paramount. And this is a technical literacy course for business leaders, not a strategy seminar.

The conversation was genuinely collaborative. The AI proposed restructuring options. I pushed back when they didn’t fit my philosophy. It pushed back when I was making assumptions I hadn’t examined. Student feedback drove the major structural decisions, but the AI helped me respond to that feedback systematically rather than piecemeal.AI tools were configured with appropriate privacy settings. Claude Max Pro with model training toggled off. ChatGPT Education account. Gemini with data retained at the University of Illinois. If you’re considering this workflow, check your own institutional policies and tool configurations.

What We Produced

Over our conversation, we worked through a comprehensive redesign scope. Not just which videos need updating, but the full architecture.

We rewrote course-level and MOOC-level learning outcomes. We restructured the module flow, merging and consolidating where it made sense and creating space for entirely new content. We estimated carryover for each module and assigned priority tiers: what needs to happen first and what can wait.Specific structural changes included merging the separate machine learning and deep learning modules into a single ‘how AI learns’ foundation, merging autonomy and robotics into one ‘physical AI’ module, creating an entirely new generative AI module, and moving ethics from four separate modules to contextual integration throughout the course. Per-module carryover estimates ranged from 70% new content to 60% kept, giving Teaching and Learning staff concrete data for workload planning.

One example: students specifically valued the infrastructure content, mentioning TPUs and ASICs by name.⊕TPUs (Tensor Processing Units) and ASICs (Application-Specific Integrated Circuits) are specialized hardware designed for AI workloads—faster and more efficient than general-purpose processors for tasks like training and running neural networks. Without working through the full picture, I might have cut that material to make room. Instead, we kept it and found space elsewhere.

The output was a comprehensive scope document with learning outcomes, restructured flow, per-module revision estimates, and prioritized implementation tiers. The kind of deliverable that normally takes weeks of meetings between faculty and Teaching and Learning staff.

What I Couldn’t Do With It

I had a full revision scope, and a rough estimate for the resources required. But the course was going to be taught again in weeks not months. I didn’t have the time or resources to make these changes.

Together with Cheng Li, a senior learning designer at Gies, we figured out what we could do now to keep the course fresh while postponing the full revision. Our answer was module guides—concise documents that bridge the foundational Coursera videos to current developments in the field.Each module guide follows a consistent template: what the Coursera videos cover, what’s new in the field since the videos were created, optional resources in two tiers (quick starts and deeper dives), and preparation for the live session. Consistent structure across all modules so students know what to expect. The videos still teach core concepts. The guides tell students what’s new, point them to optional resources, and set up the live sessions where we focus on today’s applications.

Cheng had a critical insight that shaped the whole approach. My initial instinct was too defensive about the somewhat outdated course content. Instead, he argued to stop apologizing and frame it confidently. Foundational concepts plus current applications equals intentional course design. That reframing came from a human collaborator, not AI. Both mattered.

In keeping with the scientific mindset, I also used Claude to help draft the module guide content, iterating on structure, tone, and detail across another series of conversations. AI augmenting my work, not replacing it.

This course launches again soon. So I don’t know yet whether the module guides work as intended, but I’m hopeful. I’m sharing the process, not claiming victory. But the scoping experiment wasn’t wasted. It gave us clarity about the full picture and a practical interim path until we can prioritize the real revision.

Then I Realized What This Could Mean

I did this experiment out of curiosity, driven by student feedback, over a few days. The full implications, however, didn’t hit me until later.

What if Teaching and Learning staff did this for faculty?

Feed the AI a current syllabus, anonymized student evaluations, and developments in the field. Have that same kind of conversation. Not one prompt, but a genuine back-and-forth. The goal being a scoped revision proposal that the faculty member reacts to rather than creates from scratch.

That flips the traditional workflow. T&L staff move from reactive support to proactive course management. And it frees capacity. T&L teams who are stretched too thin on revision scoping have less bandwidth for new course development, even when new courses might be a higher priority.

Starting From What’s Actually Achievable

The traditional dynamic can be painful for everyone. Faculty want to change everything. We’re academics, of course we do. T&L staff push back because of capacity constraints. Both sides may end up frustrated.

What if T&L staff instead said: “We have capacity for a 25% revision this cycle, and here’s what that looks like for your course. We can do a fuller revision in a year. In the meantime, here’s how we bridge the gap: module guides, updated assessments, and refreshed live sessions.”

Faculty react to a concrete, realistic proposal rather than scoping big and getting negotiated down. The full scope document shows what each tier requires: now, next, and later. Faculty still weigh in on priorities. They just aren’t starting from zero—and they aren’t getting told “no” without an alternative.

Those per-module carryover estimates are data, not guesswork. They could feed directly into workload modeling and timeline planning. And if the scoping produces precise change orders—like “update this case study in lecture 3 of module 7” rather than “module 7 needs a refresh”—even the production side gets easier.Precise change orders also pair naturally with avatar-generated video for surgical updates. Need to update a single case study? Re-script and regenerate just that segment. No studio booking, no faculty scheduling. I’ve written about this in my I still haven’t found what I’m looking for thought piece.

The Broader Point

To be clear, this scope document is a snapshot in time, not a finished plan. If I ran the same exercise in six months or a year, the results would likely be different. The field will have moved, student needs may have shifted, and new tools will exist. Yet, that’s actually the point. When scoping is fast and cheap, you don’t have to commit to one plan and hope it ages well. You can revisit it as conditions change.

This started with student feedback and curiosity. It ended with implications I didn’t expect. I was doing exactly what I tell others to do: experimenting with AI to learn how to collaborate with it effectively. The experiment produced something useful, and the process itself revealed a workflow that could matter well beyond my one course.

Faculty expertise and judgment are still essential. AI proposes, humans decide. Student feedback has to come from real evaluations, not AI assumptions. And the module guides are unproven until students engage with them.

But the tools exist now. The question is whether our higher ed workflows will adapt to use them.

I Still Haven't Found What I'm Looking For

2026-02-15T00:00:00-05:00

When people learn what I do, they’re often intrigued. As Chief Disruption Officer, I spend most of my time contemplating the future. But despite their interest, I’ve learned the hard part isn’t the seeing—it’s getting others to see what I do.

For years, I tried to communicate the future in my head. I would explain, hand-wave, and argue about what I saw clearly. And I would watch people nod politely without necessarily understanding.

The problem wasn’t them. It was me. Specifically, my communication. I’m an astrophysicist by training, trying to explain AI implications to business people and institutional leaders. What was clear in my head—what I could see developing—didn’t translate through words alone. The gap between what I could see and what I could show was wider than I’d realized.

This struggle reminds me of a U2 song⊕U2, I Still Haven’t Found What I’m Looking For: “I have climbed the highest mountains, I have run through the fields …” but I still hadn’t found a way to help them see what I could see.

SEEING IS BELIEVING

In late 2021, I was fortunate to work with Jake Kinsey, who had the insight to create a demo. We made an AI-generated avatar of me speaking. One version in EnglishOriginal Avatar demo from November 2021 in English. Notice the quality for 2021, impressive, but clearly synthetic., a second in MandarinOriginal Avatar demo from November 2021 in Mandarin. Impressive but clearly not fluent.. The technology, primitive at the time, worked by taking video of me, regenerating the audio to match a new script, and then modifying my lip movements to sync with the new audio.

Of course, I don’t speak Mandarin; however, even I could tell my avatar’s Mandarin wasn’t right. But that almost made it more powerful. These two videos showed what was possible and where this field was heading.

I first showed this demo to Dean Brown. When I saw his reaction, I knew immediately this demo worked. This was what had been in my head, but now made tangible. Something people could actually see and share.

This seemingly simple demonstration helped shift our college’s thinking about online education at scale. Not because the explanation was clearer, but because the demonstration made abstract capabilities real. The visual evidence carried authority that my words never could. We saw a new path to democratize education, regardless of location or language.

Seeing really was believing. The future of avatars felt clear. What I didn’t yet see was how much more there was to show.

THE FIRST AVATARS: EXPANDING BELIEF

In general, people’s reactions were similar: surprise, curiosity, some discomfort, but mostly fascination. We weren’t talking about theoretical futures; we were watching them. Our conversations now used a shared language, and we could easily bring others along by showing them.

This is how you really drive institutional change. Not through position papers, but through demonstrations that shift the frame of what people believe is possible.

For a long time, visual evidence was enough. Seeing was believing. A video of something meant that something happened. Presence, real visual and audible presence, carried inherent authority.

But technology does not stand still. And neither does the meaning of what we see.

BUILDING THE FOUNDATION

Since the first demo, we haven’t stood still. In 2023, with my graduate student Eamon Bracht and Sam Chen, the former Director of the Gies Disruption Lab, we created our first commercially developed avatar using the ElevenLabs platform⊕ElevenLabs is an American AI voice technology company with a valuation of $11 billion based on a 2026 Series D. Their platform specializes in voice synthesis and cloning, enabling realistic text-to-speech and voice replication. for voice synthesis and the Synthesia platform⊕Synthesia is a UK AI firm with a valuation of $4 billion based on a 2026 Series E. Their platform supports standard and user-created digital avatars. to generate matching video. Initially, this was an improved demo to highlight how the technology had changed. Unlike our original demo, this wasn’t manipulated video—it was completely AI-generated.

With help from Tim Anderson and Steven Pratten, we turned this new demo into a robust avatar that was used for several videosThe first production avatar, a significant step up from the original 2021 demo. in my new online course on Emerging Technology and DisruptionLinks to media coverage of our avatar work.. The avatar wasn’t perfectMy mother immediately spotted a flaw: it mispronounced my last name., but it was actionable.

A year later, with help from the Teaching and Learning staff at Gies, we built a second, and much improved avatar that we used to record almost all of the videos for a new online course entitled AI for BusinessThe full teaching and learning team won a college-level award for their work adopting AI for the creation of this new, online course.. To maintain consistency, I wore the same outfit for my live-recorded videos and for the avatar training videos. There were no obvious visual cues that might indicate a difference.

But, to be clear, we were transparent about the process. We even created a video where I share the screen with my avatar. We both introduce ourselves and I then casually tell the avatar that “Really, I think I’ve got this video.”The side-by-side introduction is embedded in the full AI for Business course highlight webpage. Full transparency—showing students exactly what they were seeing. My mother noticed my last name was now pronounced correctly!

Outside of the studio, Gies developed what I believe is the first Memorandum of Understanding between a college and faculty regarding avatar rights and expectations.The MOU isn’t perfect, but it was designed to be updated over time as we learn more about what works and what doesn’t. Together, we were learning in real time what policies this technology required.

When I first taught AI for Business in Spring 2025, students had access to all of the videos, both human and synthetic. We did not hide the existence of the avatar. Yet, in a student-led discussion about the use of the avatar, the responses were illuminating. Some students admitted they hadn’t noticed a difference or it didn’t matter, some didn’t realize an avatar was being used, and some claimed they knew all along.

Eventually, students christened the avatar “Professor Robert Burgundy”⊕Student-created image of Professor Robert Burgundy. I liked the student’s work so much I made it my course avatar on Canvas. after Ron Burgundy from the movie Anchorman⊕For more information on the movie Anchorman, see the movie’s iMDB page., because like the fictional news anchor, the avatar reads whatever it is given.

The variation in students’ responses was interesting, but it didn’t fundamentally alter my thinking. These were static, pre-recorded videos. The interaction was one-way; students watched me or my avatar speaking. But I could already see the next step: what if students could stop and ask questions? What if the avatar could respond?

THE TECTONIC SHIFT

The next step in this journey took place in the fall of 2025, when my graduate student Xinyao Qian and I built our first interactive avatar using the HeyGen platform⊕Heygen is an American AI startup with an estimated valuation of $500 million based on their 2024 Series A. Their platform includes pre-built and user-created standard and interactive avatars.. Four years after my first avatar demonstration to Dean Brown, I was demonstrating this new avatar to Dean Elliott. The visual quality is impressive—the new avatar looks and sounds like me; but like the first demo, this one had flawsInteractive avatar responding to ‘Who are you?’ Notice the 10-20 second lag between question and response, the technology is almost there, but not quite. That gap won’t last.

Additional demonstrations:

Who’s the best professor?: testing for self-promotion

Top three learning points: processing course content

Say I’m dumb: boundary testing. The delay between question and response is too long; this version is essentially unusable. Yet, just like the first demo, this new demo provides a glimpse of the future.

The new Professor Robert Burgundy won’t get tired, won’t get sick, is always perfectly lit and clearly audible. The vision is now clear: scalable expertise, a global reach, 24/7 availability, and dissolved language barriers. Personal, interactive lectures anytime of day or night. Open office hours, all the time. And content that never gets stale, the interactive avatar can simply reference updated information whenever it becomes available.

In four short years, we went from hand crafted manipulated video to fully synthetic generation to near real-time interactive dialogue. The capability is expanding rapidly, and the remaining gaps increasingly look like engineering challenges rather than fundamental limitations.

It was just another demo to the dean, but this time something felt fundamentally different. Not fear, but an awareness that the future will be different.

AUTHENTICITY OR NOT?

For centuries, seeing has been believing⊕This phrase likely derives from the bible, specifically John 20, verse 29, where Jesus says ‘Have you believed because you have seen me? Blessed are those who have not seen and yet have believed.’ This passage also led to the phrase ‘Doubting Thomas.’. Visual media implied trust; interactivity implied humanity. This new demo, however, implied something different.

In a world of synthetic media, authenticity must be external to the content; it can no longer be implied.

We can now build interactive, synthetic content. But we don’t know how to define authenticity.

How do I confirm reality? Or confirm approved synthetic creations? How do you know what to trust when visual and audible evidence, even interactive evidence, is no longer sufficient?

As Bono sang, “I still haven’t found what I’m looking for.”

Each new avatar gets me closer to conveying the future I see—and reveals a new challenge I hadn’t anticipated. Static demos had quality issues. Better quality revealed other limitations. Interactive avatars address those, but make authenticity urgent. We’ve built the avatars. We can see the trajectory. But we haven’t yet found what comes next when interactive synthetic media becomes commonplace.

These aren’t abstract questions. Professor Robert Burgundy is already here, and soon could be interactive. Change may seem slow, but looking back over the last few years, our avatar work demonstrates the opposite. The shift from assumed authenticity to signaled authenticity is not optional. But it must be intentional.

The future is visible. The question now is whether we are prepared to interpret what we see.

This article was developed with AI assistance for research, outlining, drafting, and editing. All ideas, experiences, and perspectives are my own.

Reverse Engineering a Recipe

2026-02-02T00:00:00-05:00

Summer brings many wonderful activities and traditions and one of my personal favorites is the summer farmer’s market. While many enjoy the fresh fruits and vegetables these markets offer, I also enjoy the baked goods! One of my favorites comes from an elderly woman who sells baked goods (and gently used books) to raise money for various charities. Perhaps her best offering is a homemade cinnamon roll, which—while it might not be good for my waistline—sure seems worth the price.

The End of Summer

Of course, when summer ends, my local farmer’s market shuts down. Realizing this would mean the end of my cinnamon roll supply, I decided I needed to be creative.

Have you ever looked at a homemade food label—like one from a local farmer’s market—and wondered how to recreate it yourself? All I had was a picture of the cinnamon roll with the list of ingredients printed on a sticky label.

⊕A cell phone picture showing the list of ingredients from the cinnamon roll.

As I am not a cook nor a baker, I wasn’t sure how I might tackle this problem. But being scientifically-minded, I thought there must be at least three steps:

Extract the list of ingredients from the label
Find a recipe that matches this list
Try the recipe out

Focusing on the first two steps, I was curious if a generative AI tool might be able to help. I uploaded the image of the label to ChatGPT and simply asked:

Do you think you could reverse engineer the recipe from the list of ingredients 
in this picture?

To my surprise⊕In hindsight, I suppose I shouldn’t have been so naive!, the AI processed the image, extracted the text, identified the list of ingredients⊕The image I provided above was cropped from the full photo showing the cinnamon roll wrapped in plastic with a sticky label that included the ingredients, the baker’s contact info, weight in grams, and price., and cleaned it up by removing any irrelevant text. Then, it inferred a suitable recipe.

Label Processing

The label itself offers very little detail—just a list of ingredients:

Bread Flour, Potatoes, Margarine, Sugar, Eggs, Milk, Salt, Yeast, Water, Butter, Cinnamon, Powdered Sugar, Vanilla, Maple Flavoring

No amounts. No steps. No temperatures. Just the raw components.

ChatGPT’s response?

Great! Based on the label in your photo, here’s a reverse-engineered cinnamon roll recipe using the listed ingredients.

Without any direction on my part, CHatGPT structured the cinnamon roll recipe into three components:

A yeast-based dough using mashed potato for tenderness (which ChatGPT happily informed me is a common trick in old-fashioned recipes)
A cinnamon-sugar filling made with softened butter
A maple-flavored icing made with powdered sugar, vanilla, and maple flavoring

Why Does This Work?

Generative AI is an amazing tool that can process images, extract text, and transform that text based on your instructions. Recipe reconstruction is just one use case. Other tasks that follow this pattern include:

Translating menus
Summarizing pages from books or manuscripts
Extracting and verifying data from invoices
Analyzing business cards
Reading prescriptions
Capturing code from screenshots

So next time you take a photo, remember you’re capturing more than just pixels. You might be holding a dataset waiting to be decoded.

Final Thoughts

Of course, you can adapt this approach to almost any baked good or food label. Just take a picture of the label, upload it to your favorite generative AI tool, and prompt:

Can you reverse engineer a recipe from this picture, which lists the ingredients?

Now, sharp-eyed readers will note I’ve only completed the first two steps in my plan. I’m still waiting on the third—remember, I said I’m neither a cook nor a baker! While I remain hopeful I’ll get to taste these AI-powered cinnamon rolls soon, for now I just ask the original creator for a few extra ones on the side. 🙂

If you’d like to try it yourself, here is the full generated recipe from ChatGPT. If you do bake it, be sure to enjoy one for me!

Who's Gonna Drive You Home?

2026-01-30T00:00:00-05:00

It’s hard to avoid the incessant bombardment of dire threats about the end of work as we know it. The pressure drives workers to question their employer’s intentions, and parents and their children to question their future. Authoritative figures seem happy to tell us how artificial intelligence is changing the nature of white-collar work.⊕Public statements and interviews by OpenAI CEO Sam Altman, Anthropic CEO Dario Amodei, MS CEO Satya Nadella, Google CEO Sundar Pichai, and other technology leaders over the last year predicting rapid, large-scale disruption of white-collar work, particularly at the entry level. Entire job categories are predicted to disappear, replaced by technologies that are faster, cheaper, and increasingly capable. The timelines are short. The confidence is high. Change is coming tomorrow!

I’ve heard this story before.

A few years before Covid, I was driving my youngest son to the DMV so he could get his license, a rite of passage familiar to many of us.

For much of the recent past, being able to drive meant freedom—especially for youth. Driving a car meant control over where you could go, with whom you might hang out, and how you moved through the world. From Springsteen’s Born to Run to Tracy Chapman’s Fast Car, driving wasn’t just transportation; it was agency. It was independence. It was adulthood. It was quintessentially American!

As I watched him drive, however, I was struck by a sudden thought: he was in the last generation to perform this ritual. As I told him at the time, when he is in my seat, cars will simply drive themselves, so what need will there be for a license? This moment was unlike the ones I experienced with my other children: Less a rite of passage and more like the end of the line.

Given the outsized role cars play in popular culture, it wasn’t surprising that autonomous cars captured the public, and my, imagination so quickly. The promise wasn’t just better traffic flow or fewer accidents. It was the belief that one of the most cherished youthful dreams—sitting behind the wheel—might no longer exist. Or, as Ric Ocasek, lead singer for The Cars, asked back in 1984, “Who’s gonna drive you home tonight?”

At the time, the answer seemed obvious: the car itself.

The Promise of Full Autonomy

Early demonstrations of self-driving technology were impressive enough to make full autonomy feel both inevitable and imminent. High-profile competitions, rapid technical progress, and ambitious bets by major technology companies created a widespread belief that human drivers were living on borrowed time.⊕The DARPA Grand Challenge (2004–2005) is widely regarded as a catalyst for modern autonomous vehicle research. In the mid-2010s, auto industry leaders such as Elon Musk, Mary Barra, and Mark Fields predicted fully autonomous vehicles within a few years, prompting substantial investment across the automotive and technology sectors. Once machines could handle most driving scenarios, it seemed reasonable to assume the rest would follow quickly.

What followed was not failure, but something more instructive.

The Hard Part

Autonomous vehicles continued to improve, and in many controlled settings they now work remarkably well as demonstrated by Waymo’s large-scale driverless deployments and Tesla’s early unsupervised testing. But the difficulty lay in moving from mostly autonomous to fully autonomous, that is, the last few percent. The remaining gaps: rare situations, unpredictable human behavior, construction zones, bad weather, and questions of responsibility, turned out to matter far more than their frequency suggested.⊕Construction zones, temporary lane markings, emergency scenes, and unpredictable human behavior remain among the most challenging scenarios for autonomous systems. These ‘long-tail’ cases dominate residual risk despite representing a small share of total driving time. These scenarios also challenge human learners, but autonomous systems face a higher standard. Any accident draws intense scrutiny, triggers regulatory review, and amplifies public hesitation in ways that learner accidents do not.⊕High-profile incidents involving autonomous vehicles significantly affected public trust and slowed deployment across multiple U.S. cities. Humans stayed in the loop: sometimes behind the wheel, sometimes monitoring remotely, sometimes simply remaining accountable.⊕Even in fully driverless deployments, human involvement persists through remote monitoring, fleet oversight, and legal accountability structures.

The lesson wasn’t that autonomy was impossible. It was that full replacement required something close to total coverage. And that the final stretch was slow, expensive, and socially complex.

What This Means for Work

I believe that this dynamic offers a useful way to think about current debates around AI and white-collar job displacement.

Today’s AI tools can already perform many professional tasks such as drafting text, analyzing data, writing code, and summarizing documents. In specific contexts, they are often faster and more consistent than humans. As with autonomous driving, early demonstrations create a powerful sense that complete replacement is imminent; and these warnings sound urgent, even inevitable.

But replacing a job is different from assisting with one. Full replacement implies something closer to autonomy than augmentation. It means handling not just the common cases, but the exceptions. Not just routine execution, but the moments where judgment, explanation, and accountability matter most. Those moments may be infrequent, but they carry disproportionate weight. The last few percent matter!

Here, the analogy to autonomous vehicles is imperfect but instructive. In driving, near-total reliability is non-negotiable; anything less puts lives and property at risk. In most white-collar work, the calculus is different. Organizations can capture much of the value of AI without eliminating humans entirely. Keeping people in the loop by reviewing, approving, or intervening, often remains cheaper, safer, and more acceptable than pushing for full automation.⊕This is often discussed under the comparative advantage thesis.

Popular culture provides useful reference points. For example, in Top Gun: Maverick, the aging pilot isn’t sidelined by autonomous systems or newer technology. Instead, his value lies precisely in what the machines lack: judgment at the edge, intuition under uncertainty, and responsibility when the stakes are highest. The future arrives, but it turns out to work better with humans still in the cockpit.

However, many believe AI represents a fundamentally different challenge than autonomous vehicles, either cars or fighter jets, one that will genuinely displace knowledge workers. Geoffrey Hinton, whose 2024 Nobel Prize⊕In October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Physics jointly to Geoffrey E. Hinton and John J. Hopfield ‘for discoveries that revealed the physical principles of learning in artificial and biological neural networks, transforming our understanding of adaptive complex systems.’ recognized his foundational contributions to AI, certainly thought so. In 2016 he famously stated that we should stop training radiologists because machines would soon do a much better job. Yet in spite of being a world-famous AI expert, that prediction hasn’t aged well.

The need for radiologists hasn’t dropped, it has risen; despite AI tools being frequently used by the profession.⊕See, for example, the recent Financial Times article on AI and radiology. One main reason for this result is the few percent problem, humans plus artificial intelligence often perform best. AI can quickly point out anomalies, but humans can catch things outside the AI’s training data. But that is only part of the story. With AI’s help, radiologists can do more in the same time, meaning the price of a radiology consult will drop, and the demand for their service grows;⊕This is an interesting example of Jevons’s paradox: When technological improvements make a resource more efficient, overall consumption of that resource often increases, not decreases. Or, as discussed here, as radiologists become more efficient with the use of AI, the demand for their services can increase. they are not an endangered occupation.

To be clear, I am not saying that work won’t change. It already has. Tasks are being compressed, roles reshaped, and expectations adjusted. But the path forward is likely to be slower and more uneven than early demos suggest or that AI leaders might hope.

Autonomous cars did not fail; they continue to mature. They taught us that the hardest part of automation is not getting machines to work most of the time, but deciding when we are comfortable stepping aside. AI and jobs appear to be following a similar trajectory. Progress will continue. Capabilities will improve. But humans may remain in the driver’s seat longer than many early predictions assumed; not because the technology stalled, but because the final handoff is harder than we think.

In 1984, The Cars asked who would drive us home. Forty years later, the answer is more complicated than they thought.

The path forward isn’t choosing between human or machine; it’s learning when to steer and when to let the system assist. In full transparency, I used AI while writing this piece⊕In writing this piece, I used Claude, ChatGPT, Gemini, and Perplexity to research autonomous vehicle history, refine arguments, and draft sections of text. The structure, reasoning, and conclusions are my own, an example of the human–AI partnership I promote in this thought piece.—not to replace my thinking but to sharpen it. That partnership feels like the right model for this story. I’m still responsible for the destination. I just didn’t drive every mile myself. Now if I only had a car that did the same.

In-Situ Website Translation

2025-12-05T00:00:00-05:00

A popular refrain from AI-focused tech companies is that artificial intelligence will change how we live and work. For most of us, however, this sounds like more marketing speak. Recently though, I found an example that speaks to this promise—perhaps it is familiar to some of you, but it was new to me so I thought I would share it more broadly in this HowTo guide.

I often come across documents that describe how emerging technology is innovating or disrupting different parts of the world. As you can imagine, these are of great interest to me; yet many of them are in a foreign language. Fortunately, I recently learned that many popular web browsers can now provide automatic webpage translation in situ.

Safari Translation in Action

To demonstrate this, I will use the Safari browser on my Mac. I recently came across an annual review of the Chinese public’s use and views on generative artificial intelligence from Tencent Research Institute. As you can see from the screenshot of the original webpage, everything is in Mandarin—a language I neither speak nor read.

The original webpage displayed entirely in Mandarin Chinese

The Safari browser provides a translate button in the address bar⊕The translate button showing the option to translate the webpage into English. when it detects a webpage in a foreign language. Simply clicking this button creates a new view of the webpage that is translated into your preferred language. In my case, that is English.

The same webpage after translation, now displaying in English

After a brief period where the translation is performed, the new view is displayed. While the translation of plain text is no longer seen as a magical feature, I was definitely impressed by the automatic translation of text in figures and infographics. As you can see in the following figure, at least for Safari, this in situ translation places boxes over the original text and then displays the translation in these boxes.

Even complex infographics are translated, with text overlays rendered in the target language

The ability to translate not just the body text but also embedded graphics demonstrates how far browser-based translation technology has advanced. This makes previously inaccessible content immediately readable and useful for research, business intelligence, or simple curiosity about global developments.

How to Enable Translation in Safari

For those who want to use this feature:

Open Safari and navigate to a webpage in a foreign language.
Look for the translation icon⊕The translate button appears in Safari’s address bar when a foreign language is detected in the address bar.
Click the icon and select your preferred language.
Safari will translate the page automatically.

The translation happens locally on your device in recent versions of Safari, which means your browsing data isn’t sent to external servers—a nice privacy benefit of this approach.

Other Browsers

While I’ve demonstrated this feature in Safari, other major browsers offer similar translation capabilities. Since I don’t regularly use these browsers, I worked with Claude to research and document their translation features.

Google Chrome: Chrome offers automatic translation powered by Google Translate. When you visit a foreign language page, Chrome will prompt you to translate it. You can also right-click anywhere on a page and select “Translate to [Language].” Official Chrome Translation Documentation

Mozilla Firefox: Firefox offers translation through their “Firefox Translations” feature (formerly an add-on, now built into recent versions). Unlike cloud-based services, Firefox performs translations locally on your device for enhanced privacy. The translation models are downloaded and stored on your browser. Official Firefox Translation Documentation

Microsoft Edge: Edge includes Microsoft Translator built directly into the browser. Similar to Chrome, it automatically detects foreign language pages and offers to translate them. You can translate pages by clicking the translation icon in the address bar or through the right-click context menu. Official Edge Translation Documentation

Conclusion

Personally, user aids such as in situ website translation represent a genuinely useful application of AI and machine learning technology in our daily digital lives. Whether you’re conducting international research, following global news, or simply exploring content from around the world, these built-in translation features remove language barriers with remarkable ease. The fact that modern browsers can now translate not just text but also content within images and infographics makes this technology even more impressive and practical.

The Quantum You Want

2025-09-17T00:00:00-04:00

Throughout our history, humanity has striven to predict the future. From astrology and crystal balls to modern science, we look for an edge in what is to come. Consider the weather, something we often discuss even with strangers. Your smartphone likely has one or more apps that will tell you whether it will be cloudy or sunny today, or if storms are coming. While we often focus on the veracity—or lack thereof—of these predictions, perhaps a more important observation is the power of the device itself.

We rarely dwell on it, but your smartphone would rival the top supercomputers in the world from the turn of the century. And while few predicted the incredible power we now carry in a pocket or purse, both smartphones and supercomputers function in a similar manner. Data from the world around us—whether temperature measurements, texts, music playlists, or YouTube videos—are converted into long streams of zeros and ones, then processed by computer chips that follow deterministic rules.

This highlights an important point. Despite the historical obsession, humanity is still bad at making predictions. A good summary of this state of affairs comes from Bill Gates, who in 1996 wrote:Bill Gates, The Road Ahead, Afterword, p. 316, Penguin Books, 1996

People often overestimate what will happen in the next two years and underestimate what will happen in ten. I’m guilty of this myself.

Of course, this prediction paradox applies perfectly to computing. Today’s top supercomputers are roughly a million times more powerful than your smartphone. Yet even these behemoths are often unable to predict the weather a few days out. As Mick Jagger eloquently sang, “You can’t always get what you want.”But as the song continues, you get what you need.

The Computing You Want

While some might argue for better algorithms or new hardware designs, we may simply need an entirely different approach—one first postulated by Nobel prize-winning physicist Richard Feynman. At a conference on computers in physics in 1981, he introduced the concept of using quantum mechanical systems for computation, famously stating:For more information, you can read the paper he submitted following the conference.

Nature isn’t classical, dammit, and if you want to make a simulation of nature, you’d better make it quantum mechanical.

Now Feynman was generally referring to simulations of molecules and materials that involve fundamental physics or chemistry, rather than the weather. But this charge helped spur others to explore this new idea. And now, over four decades later, we are at the cusp of a new revolution in computing, one based on the principles of quantum mechanics. For many, any exposure to the quantum world and the probabilistic nature it embodies can end in confusion. For example, in the classical world, you can know both where a car is located and how fast it is traveling. But in the quantum realm, knowing one of these, either position or velocity, means we can’t precisely know the other. This isn’t due to any fault in how we measure them, but is instead a fundamental property of nature described by the Heisenberg Uncertainty Principle.A remarkably simple yet profound result, first shown in 1927 by the Nobel prize-winning physicist Werner Heisenberg. More formally, there is an inherent limitation in how well we can know the product of the position and momentum of a particle. This uncertainty becomes important when dealing with fundamental particles of nature like the electron or proton.

Thus, it is unsurprising that quantum computers are not as easy to build, use, or understand as your smartphone. Classical computers work with bits that are either zero or one—on or off. Quantum computers are based on quantum bits, or qubits, which, being quantum mechanical, have some surprising behaviors. The two most relevant for our discussion are superposition and entanglement.

Superposition means that a qubit can be in multiple states at once, like a coin that is simultaneously both heads and tails, until you observe it. In contrast, a classical bit is like a coin that has already landed: it’s either heads or tails, but never both. While this phenomenon may seem abstract, combining multiple qubits causes the power of superposition to grow exponentially: two qubits can exist in four possible states, ten qubits in over a thousand, and twenty qubits in over a million. Through superposition, quantum computers can encode and manipulate many possible states simultaneously within the same system.

On its own, this provides unique avenues to reinvent computation. However, that’s only part of the story. The second quantum property—entanglement—means that qubits can become deeply linked, such that knowing something about one instantly reveals information about the other, even if they’re far apart.If this seems troubling, you are in good company. Einstein referred to this process as spooky action at a distance. Imagine we each flip a pair of entangled coins. If I look at mine, we instantly know what yours shows without having to look. This idea, combined with sufficiently clever algorithms, allows many entangled qubits to be processed simultaneously. As a result, we can navigate what was once an intractably complex landscape to reach the desired endpoint with incredible speed.

The Quantum You Find

This promise has driven large companies like IBM, Microsoft, and Google, as well as smaller players such as D-Wave, IonQ, and PsiQuantum, to explore different qubit architectures and to construct and deploy multi-qubit systems. Yet the physical challenges of building and operating quantum systems, especially at the scales needed for real-world problems, remain an unsolved challenge. Simply put, quantum systems lack the commodity approach of classical systems that rely on semiconductors—each company building in this space seems to have a different approach.

Despite these challenges, the development of quantum algorithms has not stopped, allowing us to predict where quantum computers will have the biggest initial impact. And, spoiler alert, it won’t be running Excel or watching YouTube. Instead, quantum systems should excel at these four tasks:

Finding needles in a haystack: Quantum computers can quickly search through large data sets to find targets, which can be useful for database searching and pattern matching.

Solving optimization problems: Quantum computers can identify the optimal solution out of a very large number of possibilities, which is useful in supply chain logistics and financial modeling.

Cryptography: Quantum computers can quickly break traditional encryption standards, which threatens global communication and financial transactions. Fortunately, quantum computing also offers new approaches that are immune to these threats.

Simulating Nature: At its core, nature follows the rules of quantum mechanics, thus quantum computing, as envisioned by Feynman, can naturally simulate molecules, chemical reactions, and material properties. This could lead to the development of personalized medicines and tailored materials that outperform existing options.

When Will Quantum Arrive?

So the million-dollar question is when will a quantum computer be available? Many industry experts predict the first working, general-purpose quantum computer—known as Q-Day—will arrive within five or ten years.After raising the difficulty in making predictions I hesitate to do so here. So I simply leverage AI to company, expert, and consultancy predictions. While that may seem optimistic, it is also prudent to recall Gates’s quote. Maybe you will soon know unambiguously whether to pack that umbrella or not!

Originally published by the Gies College of Business on October 16, 2025.

Dancing in the Dark

2025-07-22T00:00:00-04:00

Popular media regularly characterizes academics as occupants of an ivory tower, perhaps with hints of derision as the occupants are presumably disconnected from practical concerns. Faculty have historically been isolated from real world disruptions caused by technology, such as the computer or the Internet, which diffused from academic labs into industry. The Ivory Tower provided refuge.

But this metaphorical barrier has been frequently pierced of late. The primary drivers behind emerging technology disruptions are no longer academic or government labs but corporations and private firms. And nowhere is this more evident than with Generative AI, which has been developed and disseminated primarily by researchers working for deep pocketed technology firms.Competition between technology firms (and countries) has created an arms race for talent and resources. This origination story impacts faculty in two important ways. First, Generative AI tools, like ChatGPT, developed rapidly and were released to an unprepared public, leaving many feeling flat-footed by this seemingly magical technology. Second, and perhaps most importantly, this technology primarily impacts what is known as knowledge work. Simply put, Generative AI cuts to the heart of what it means to be an academic; The Ivory Tower no longer provides refuge.

We are supposed to be experts; yet, in this case, we are behind the curve, unsure of how to react or even what to do. It is as if a universal case of imposter syndrome has descended upon the Ivory Tower as we struggle to understand how to embrace, explain, or use this technology. We are, in effect, dancing in the dark:This phrase was popularized in a hit song by Bruce Springsteen that dealt with the themes of isolation and frustration with work, which seems rather relevant to this discussion.

unsure if we are alone or in a crowd
unsure if we are doing the right things
unsure if we are improving or regressing
unsure if this is the beginning of the end or the end of the beginning

The Disruption of Academia

The first interaction for many of us was not a good one—a sudden wave of Generative AI-generated answers to writing assignments (ironically characterized by an excessive use of the em dash) that threatened a crucial assessment technique, and more broadly the concept of academic integrity. The resultant outcry was swift and perhaps predictable. Many called for bans, some called for patience and reflection, and some the Persian adage: “This too shall pass.”

But in the interim, we have learned several important lessons:

Generate AI is not going away.
Generative AI is continually improving.
Generative AI proficiency is a required skill for our graduates.

Thus, if we are forward thinking we must embrace the adoption of AI, or risk growing increasingly irrelevant within our ivory towers. Yet, we have little clarity or guidance on how to best move into this brave, new world. Either from our administrations or peers—we are still dancing in the dark! And, as knowledge workers, we are further confronted by the discomforting hint of the ephemeral nature of our our own careers.

We spend years mastering a subject, becoming the expert in the room. But now, Generative AI can summarize an entire research field, customize our course lecture notes for each student, or critique a case study with unnerving fluency. The challenge is less about AI being smarter and more that AI is faster, convincing, and tireless. The comparison, and competition, is unsettling. This leads to a new imposter syndrome where faculty ask, “Am I falling behind?” and “Is my expertise still relevant?”.

Alongside these doubts are real ethical concerns. Faculty may be asking:

What counts as “my work” if I use AI?
Am I allowed to use AI to provide student feedback?
Am I violating academic integrity standards if I use AI too much or too little?

Without clear institutional guidance and support, many faculty are adrift in an ethical fog, unsure how or even if they should proceed. And these feelings apply equally across our service, teaching, and researchExperimentation is already happening on AI augmented research, see the Agents4Science conference. roles.

A Guiding Metaphor

Fortunately, there is a simple step we can take. Embrace the future and learn to work effectively with AI. Rather than seeing AI as a threat, view AI as a skilled intern that can quickly draft, summarize, analyze, and even organize large corpora. But this, in turn, requires us to provide direction, oversight, and judgment. Of course, this is the same advice we should be giving to our students, who are increasingly expected to master these skills to gain employment.

While easy to say, doing this can be discomforting. But this is not a new concept. In 1987, Apple released a concept video entitled the “Knowledge Navigator” that showed a professor interacting with a highly competent digital assistant.⊕Knowledge Navigator video This bow-tied AI helped schedule meetings, find articles of interest, and organize research. While inspirational, it was at the time, science fiction.

But today this dream is alive. Faculty can ask AI to generate lecture outlines, critique writing, summarize recent research, or simulate a conversation in another language. What was once imagined as a distant future is now relatively cheap and ubiquitous. Yet faculty are still struggling as they ask, “How do I get started?”, “What is permitted under university policy?”, or “How do I remain a role model for my students?”.

The initial goal shouldn’t be to master AI, but to learn to use it meaningfully. Faculty must be the trusted guides into a world where AI influences how knowledge is produced, shared, and valued. For this to happen, institutions must provide the supporting scaffolding for a confident adoption of AI. They must:

Normalize discomfort. Faculty must be able to share their concerns and experiences without fear of judgment.
Provide clear guidelines. Institutions should articulate what is allowed and what is not.
Promote mentorship and peer learning. Small groups who share use cases, failures, and lessons.
Recognize ethical creativity. Highlight examples of AI-enhanced assignments, not just AI detection tools.
Encourage transparency. Let students know when and how AI is used—and expect the same from them.

Just as the Knowledge Navigator inspired an earlier generation, today’s Generative AI tools invite us to rethink how we teach, learn, and lead. We can stop dancing in the dark! With an open mindset, shared wisdom, and institutional support, the lights will come on and our path forward becomes not only visible, but empowering.

Originally published by the Gies College of Business on August 27, 2025.

Ghosts in the Machine

2025-07-02T00:00:00-04:00

Humans love to anthropomorphize. From seeing objects in cloud shapes to naming physical devices, we humanize our environment, making complex things easier to understand and to describe to others. This approach has carried over into the world of artificial intelligence, with the creation of named tools like Alexa and Siri, or when ChatGPT says “Let me know if …” within its response.

On the face of it, this may seem odd. But generative AI tools like ChatGPT and Claude aren’t just code—they are algorithms defined by the shared knowledge of our species. Thus, if you ask ChatGPT to explain special relativity in the voice of Shakespeare, it may seem as if the ghosts of Einstein and Shakespeare are collaborating to quickly generate the result. While this contrived example may have little bearing on the world of business, the embodied spirits in the model are not limited to these historical examples. In this analogy, these large language models are haunted by the cumulative expertise of billions of humans, or the ghosts in the machine, whose digital legacies are encoded in their parameters.

Ironically, the phrase “Ghost in the Machine” was coined by British philosopher Gilbert Ryle as a criticism of the idea that the mind could be considered separate from the body. Now, we use this phrase to describe our interactions with a super-powerful, non-corporeal digital mind—Generative AI. But these interactions are not limited to generating essays; these “ghosts in the machine”, trained on numerous examples, are now completing tasks once exclusively reserved for entry-level workers in marketing, finance, HR, and sales.

This begs the questions: if AI becomes the default “junior hire,” what becomes of recent graduates? And how should higher education—especially business schools—adapt to this new reality?

The Disruption of Entry Level Work

As educators, our first thought around AI tends to be how can we ensure the academic integrity of our classes? While this question has merit, by focusing on the how we educate, we risk missing the more important questions of why and what we educate. As the hyperbole around AI has grown, more students and their parents are questioning the value of a traditional college. These concerns follow from studies such as the well-covered State of the Tech Talent report by SignalGate⊕State of the Tech Talent report that showed a 50% reduction in new graduate hires by big technology firms from pre-pandemic levels. More recently, Dario Amodej, the CEO of Anthropic, a leading AI firm, stated that up to 50 percent of entry-level white-collar jobs could be eliminated by increased adoption of AI tools (i.e., the ghosts in the machine) over the next five years. Why should we expect that business as usual is still a viable approach for higher education? Will the ghosts in the machine take over?

To be clear, Amodej’s warning carries weight, but maybe the truth is more nuanced. Executives from OpenAI, a rival AI firm, paint a different picture. Brad Lightcap, COO of OpenAI argues that rather than AI eliminating entry-level roles, companies will embrace young people⊕Brad Lightcap interview who have a level of fluency with AI that far transcends anyone else at those organizations. Likewise, Sam Altman, CEO of OpenAI emphasizes that AI will make employees more productive,⊕Sam Altman, ‘The Gentle Singularity’ leading to a positive sum game.

Powering every AI output is an intricate mosaic of human knowledge – knowledge created by countless professionals, creators, and experts. By leveraging this vast mosaic, AI can complete routine tasks such as writing reports, summarizing corpora, and drafting presentations in seconds. But the AI that powers these tasks must be told what to do and are then judged on how they did. The “ghosts in the machine” are not alive and they lack human values such as empathy, fairness, dignity, and autonomy. The AI excels at the middle work. Being fluent with AI will mean knowing how to interact with these tools by providing the inputs and knowing how to interpret and use the output.

This idea runs through the PwC 2025 Global AI Jobs Barometer, which analyzed nearly a billion job postings. They found that industries using AI report three times higher productivity growth and that professionals with AI skills commanded a 56% wage premium over those without AI skills. The message is clear: AI adoption is less about job displacement, and more about AI augmentation where workers shift from routine tasks to strategy, oversight, and problem solving. Human+AI teams outperform either individually.

The Higher Education Crossroads

The challenge for higher education is how do we meet this new and fast-growing market requirement, especially when constrained by the traditional academic bureaucracy? As a simple example, how can we justify teaching marketing students the same way in a world where Meta plans to automate⊕Meta’s AI advertising plans its online advertising business? Short-sighted thinking places us on the path to our own obsolescence. The alternative view is focusing on the needs of the student. A recent study showed that AI tends to complement human skills⊕Research on AI and human skills like digital literacy, teamwork and resilience far more than it could substitute for them, and that employees who embraced this collaboration experienced wage gains.

Student-centric thinking highlights that graduates need more than domain specific know-how; they require AI fluency and experience in applying AI in their domain. Meeting this requirement means approaching our educational mission with a different mindset. First, we need to embrace AI and AI tools in the educational process. This ensures students get the needed experience, which provides a new and strong argument for increased experiential learning. Second, we need to consider curricula revisions to prepare students for this AI augmentation of existing careers and new careers that arise from increased adoption of AI.

This means reimagining business education. Rather than a focus on the mastery of tasks, we must focus on the partnership between human and AI that can lead to better outcomes for all. By confronting this challenge, we can transform the ghosts in the machine from harbingers of doom to partners in a brighter future.

Originally published by the Gies College of Business on August 6, 2025.

The Innodative Disruptor

Accelerating towards Obsolescence

The Economics of AI Anxiety

My Teleprompter that Claude Built

What it does

The iteration

The pattern

Try it

Paper Records as Unhackable Backup

Who Pays for Open Weights?

Most of the Internet Isn't Human Anymore

A Signal Becomes Real

Nobody Said It Was Easy

The Experiment

What We Produced

What I Couldn’t Do With It

Then I Realized What This Could Mean

Starting From What’s Actually Achievable

The Broader Point

I Still Haven't Found What I'm Looking For

SEEING IS BELIEVING

THE FIRST AVATARS: EXPANDING BELIEF

BUILDING THE FOUNDATION

THE TECTONIC SHIFT

AUTHENTICITY OR NOT?

Reverse Engineering a Recipe

The End of Summer

Label Processing

Why Does This Work?

Final Thoughts

Who's Gonna Drive You Home?

The Promise of Full Autonomy

The Hard Part

What This Means for Work

Sharing the Wheel

In-Situ Website Translation

Safari Translation in Action

How to Enable Translation in Safari

Other Browsers

Conclusion

The Quantum You Want

The Computing You Want

The Quantum You Find

When Will Quantum Arrive?

Dancing in the Dark

The Disruption of Academia

A Guiding Metaphor

Ghosts in the Machine

The Disruption of Entry Level Work

The Higher Education Crossroads