• chats with Zack Kass, former head of Go To Market at Open AI, about his new book, The Next Renaissance: AI and the Expansion of Human Potential 
• We use “renaissance” now as a convenient catchall for massive improvement and progress across art, science, math, and politics in a period that saw tremendous growth in population and in other areas.
• I chose the Renaissance [as a metaphor] because it’s a period of time that most people can relate to as being foundational in human history.
• or reference, the adoption gap is the space between the technological threshold, or what technology can do, and the societal threshold: what we want it to do, or what we let it do.
• There are reasons for the adoption gap. Even in the case of robotics, where machines can wash dishes and will soon be in every home, I’m not sure it’s “right.” There is an incredible disparity between what a machine is capable of and what we choose to have it do.
• Another reason for the adoption gap is that we have exceptional tolerance for human failure, and we have none for machine failure.
• The path to the future is going to be driven not by scientific progress but by cultural and societal progress.
• The adoption gap [during the Renaissance] was a result of a policy issue, not a technology issue.
• But we should have some humility to recognize what we consider a long time now—ten years, perhaps—might have been considered no time at all 500, 600, 700 years ago.
• I think there will be more work. Humans will find things to do, and we should have the humility to remember this. We lack the humility to consider all of the things that can be that we can’t yet fathom.
• The idea that work could get better is somehow upsetting to people. I do have to reframe it, and I tell people that I don’t think this is an economic problem.
• We’re all very eager to see the world get better, faster, and cheaper without realizing what that means.

• One of three things will happen:

   
   
  1. We will automate a ton of work and have an incredible economic boom. Then we will have to figure out how to distribute the gains.
   
  2. We automate very little and become stagnant economically.
   
  3. We automate some jobs, and then protect a bunch of others. This [scenario] would be tricky.
   
   
• Some people won’t have the work that they used to have. Yet somehow, others still have to stand in line to mail something because we didn’t automate the postal service.
• “Why isn’t your job automated?” The response will be, “Because a politician said it can’t be.”
• A lot of the job fears are overblown. On a macro basis, they’re terribly overblown. That’s because if we automate most work, something good will happen.
• . Given my discussions with management consultants, what you will see, in knowledge work in particular, is the apprenticeship problem.
• It means that even if we hire more young people for certain roles, the nature of that job is going to change.

• Now we realize that if we automate that work, apprenticeships will begin to change.

   
• Most management consultants who work their way up at a firm remember what it was like as an associate who did everything.
• In 20 years, we will see companies that are run by people who don’t necessarily know all the work that was once done. As a result, they may not appreciate the nature of the work as much.
• One is the alignment problem, which is poorly defined and begs redefining. That’s the question of “Does the machine care about our values, and does it concern itself with the consequences of its actions?”
• The other weird outcomes relate to explainability: models that are exceptionally good but cannot or do not explain themselves well.
• The risk that I worry about most is bad acting. I refer to bad acting not in the sense of doom, but in the sense of catastrophic events that could really set progress back
• Roughly 9 percent of the population is Agent Nero. They are ill-intentioned or antisocial, either because they’re psychopathic or they’ve figured out how to make crime pay.
• One of my challenges to everyone—young and old—is to seize the opportunity to learn how to live.
• We must avoid that trap in a world where people can outsource all of their critical thinking and subsist quite comfortably because we have automated so much
• The process by which you accumulate skills is as important as the skills you’re accumulating, and in many cases more important. Knowing how to play a piano is interesting insofar as you can play the piano, and that may be very rewarding.

26 more annotations...

• AI nausea isn’t skepticism about technology. It’s a reaction to cognitive overload and strategic dilution:
• AI doesn’t fail because it’s not powerful. It fails because organizations deploy it like a trend, not like a production capability with constraints, costs, and risk.
• The antidote is not “less AI.” It’s better decisioning

• 2.1 Energy is not an abstract externality

      

  AI runs on compute. Compute runs on electricity. Electricity runs on infrastructure. And infrastructure has limits—grid capacity, permitting cycles, transformer availability, cooling, water constraints, and local community acceptance.

• 2.2 “Cost per demo” is not “cost per enterprise outcome”

      

  GenAI pilots are cheap relative to scaled operations. Enterprises routinely underestimate:

      
   
  • Inference cost at scale (especially when usage becomes habitual).
      
  • Data plumbing: integration, lineage, permissions, retention, and observability.
      
  • Model governance: evaluation, monitoring, drift detection, incident handling.
      
  • Security hardening: prompt injection defenses, access controls, red teaming, logging.
      
  • Change management: adoption is not automatic; it must be designed.

• n AI becomes a mandate rather than a strategy, two things happen:

      
   
  • Teams chase use cases that are easy to demo but hard to operationalize.
      
  • Leadership gets a portfolio of projects, not a portfolio of outcomes.
• AI is one tool in the value-creation toolbox. Powerful, yes—but not exempt from economics.

• Define value in three buckets—and force every initiative to live in one:

      
   
  • Revenue growth: conversion, personalization, pricing, product innovation.
      
  • Cost productivity: automation, deflection, cycle-time reduction, quality improvements.
      
  • Risk reduction: fraud detection, compliance controls, safety monitoring.
   
      
• Most companies would never approve “unlimited spending” for cloud storage or travel. Yet GenAI often gets deployed without a tight discipline on usage patterns and unit economics.
• Set rate limits and guardrails for high-volume workloads.

• 3.3 Separate “innovation sandboxes” from “production platforms”

      
• 3.4 If your strategy ignores energy, it isn’t a strategy
• 4) GenAI Is Not a Universal Hammer

• There are at least four categories of problems where GenAI is routinely over-applied:

      
   
  • Broken processes (automation won’t fix a bad process design).
      
  • Data quality issues (GenAI can mask them, not solve them).
      
  • Deterministic rules (where simple logic or RPA is cheaper and more reliable).
      
  • Regulated decisions (where explainability, auditability, and bias constraints dominate).
• 5) Where GenAI Truly Wins (and Where It Loses)

• 5.1 High-fit GenAI patterns

      
   
  • Knowledge work acceleration: summarizing long documents, drafting variants, extracting structured fields from unstructured text (with validation).
      
  • Customer support augmentation: agent assist, suggested replies, faster retrieval of policies and procedures.
      
  • Software productivity: scaffolding, refactoring assistance, test generation—when governed and reviewed.
      
  • Content operations: marketing drafts, localization, internal communications—within brand and legal constraints.
      
  • Search + retrieval: better discovery across enterprise knowledge bases (RAG) if content is curated and access-controlled.

• 5.2 Low-fit GenAI patterns

      
   
  • High-volume transactional automation with stable rules (classic RPA/workflow engines often win).
      
  • Financial close and controls where traceability and determinism matter (GenAI can assist, but shouldn’t “decide”).
      
  • Safety-critical decisions where errors have outsized impact.
      
  • Processes with low standardization and no documented baseline (you need process work first).

• 6) When Process Reengineering and RPA Beat GenAI (with Examples)

      

  Before you apply GenAI, ask a blunt question: Is this a process problem, a workflow problem, or a language problem?

• 7) Corporate Risks: The Unsexy List Leadership Must Own

      

  Scaled AI use introduces a layered risk stack. Treat it like any other enterprise risk domain—cyber, financial controls, privacy, third-party, and reputational risk—because that’s what it is.

      

• 8) The AI Operating Model: From Hype to Repeatable Delivery

      

  If you want AI value without AI chaos, you need an operating model. Not a slide. A real one.

• 8.1 Create an AI Portfolio Board (not an AI hype committee)

      

  Its job is to approve and govern use cases

• 8.2 Standardize delivery patterns

      

  Most enterprises should build repeatable blueprints:
• 8.3 Implement “trust controls” as first-class features

• 8.4 Treat adoption as a change program

      

  AI changes roles, behaviors, and accountability.

• Use this simple scorecard to decide whether GenAI is the right tool:

      

  Question	If “Yes”	If “No”
  Is the core problem language-heavy (summarize, draft, classify, search)?	GenAI may fit	Consider process/RPA/rules first
  Is the process stable and standardized?	Automation can scale	Reengineer the process first
  Is the decision regulated or safety-critical?	Use assistive patterns + controls	More freedom, still monitor
  Can you measure value with a hard KPI and baseline?	Proceed	Don’t fund it yet
  Do unit economics work at scale (cost per transaction)?	Scale with governance	Redesign architecture or stop
• If you want to cure “AI nausea,” stop treating GenAI as a universal solvent.
• Put differently: the companies that win won’t be those who shout “AI-first” the loudest. They’ll be the ones who build AI-smart—with economics, controls, and outcomes engineered into the system.

25 more annotations...

• However, transformation is not always achieved by going from A to B. As teams rush to define one future state after another, they may already realize something strange: there is no way to define the future in advance.
• Whether transformation involves workforces or technology, the process is not supposed to be a perfect progression from A to B
• Mapping out a path from A to B does not mean these events will definitely happen. They are simply scenarios that are ripe for testing.
• Strategy was never about going from A to B. It’s about learning to overcome barriers and seizing opportunities in the process.

• This becomes obvious when you attempt to learn something new. When one path doesn’t work out, we simply take another.

  Advertisement

   
• If anything, strategy is supposed to bend, not break.
• Products that fail to meet expectations can still be attractive acquisition targets.
• Transformation is built on imperfection. These imperfections are not bad things. They might include setbacks that can lead to new opportunities.
• A setback alone is not enough to conclude that a strategy is broken. At any given time, an organization has a portfolio of solutions that can be developed further.

7 more annotations...

• The real payoff comes when employees learn how to apply generative AI in their day jobs in a way that improves how they work
• Without a product-minded approach, employees’ attempts to use AI often remain shallow or short-lived. Lacking a clear sense of how to apply AI to their highest-value problems, employees may struggle to find a place to start.
• For managers asking, “How do I get my people to use AI?,” the answer is often better product discipline rather than more prompt training.

• Define the Problem to Solve and Its Value

   
• employees naturally approach new technologies by asking, “What can this tool do?”
• Instead, employees need to develop the skill of discovering for themselves what the technology can do to support their work.
• they better understood gen AI’s applicability when they heard someone else explain how they applied it to a concrete, high-value problem
• when managers modeled their own problem-solving process, employees felt more encouraged to experiment, rather than waiting for prescriptive training.
• Evaluate Technology Options
• imilarly, there are many different kinds of gen AI tools, so matching the right tool to the problem can be difficult.

• valuate gen AI tools against their specific needs by considering:

   
   
  • The interface and workflow: A simple chat interface is often best for individual brainstorming, but a tool that allows for a “pipeline”—like a custom Gem or a shared notebook like Google’s NotebookLM—can be better for repetitive team processes.
   
  • Data integration: Some tools are “closed,” relying only on the model’s pretrained knowledge, while others allow users to pull in their own live data, such as a folder of contracts or a database of customer feedback.
   
  • Tool capability: Because foundation models and tool interfaces evolve rapidly, a tool that fell short for a complex task last month may be well suited today.
• To educate employees on these rapidly changing tools, institute meetings, newsletters, and other resources that model and demonstrate their use.
• Experiment
• Effective product management depends on experimentation through the typical product cycle. This helps to quickly refine the value of a solution.
• Experimentation may be getting some bad press for inefficiency lately, but building new automated workflows depends on learning what is possible, showing potential value, and refining both the problem definition and the criteria for success—
• Unfortunately, the mindset needed for experimentation can be hard to instill.
• Employees cited lack of time, uncertainty over whether experimenting was a legitimate use of work hours, fear of being judged if their experiments failed, and concern that their success would be “cheating.”
• Integrate
• A prototype is not a product, and an experiment is not the end of gen AI adoption. Product managers know that a brilliant feature is useless if it doesn’t fit into the user’s existing ecosystem.
• moving the AI solution out of a separate chat window and into integrated flows with tools that they and their teams already use every day.
• encouraging two kinds of interoperability:

• Technical.

   

  Encourage employees to ask, “Where do I actually need this information to go?”

• Process.

   

  Gen AI tools can often change the necessary order of operations for a team. When one person automates a task, it creates a ripple effect for whomever next receives that work. Managers should help employees redesign roles and routines
• Gen AI adoption at work rarely fails because people can’t write good prompts. It fails because employees struggle to see how AI fits into their real workflows
• One of the most powerful change-management tools we observed was simple but underused: demos and modeling. When leaders and peers show—not tell—how someone identified a high-value problem, evaluated tools, ran a small experiment, and integrated the result into daily work,
• A product management mindset offers a helpful lens for this work—but it should not be applied too rigidly. In a fast-moving AI landscape, individuals need to be exploratory and adaptive, not bogged down by formal roadmaps or premature ROI calculations

24 more annotations...

• On Thursday, the Redmont giant’s share price collapsed by nearly 12 percent after it released its latest quarterly results, making it not only its biggest single day slide since March 2020
• , was despite Microsoft actually exceeding some key expectations, including its net income, which rose by 23 percent from the same period the year before to nearly $31 billion. Revenue also increased by 17 percent to $81.3 billion, which is about a billion more than what analysts projected.
• But Microsoft’s AI spending spree has investors second-guessing its direction, and it’s striking that the lack of faith was strong enough to precipitate a historic plunge even with respectable financial growth
• Azure reported a 38 percent bump in revenue, which is slightly slower than the year before, adding to investor uncertainty over whether the business will be able to reap back the tens of billions spent on its data centers.
• “Since it is becoming even more evident that Microsoft is not going to garner a strong ROI from their massive AI investment, their shares need to be revalued back down to a level that is more consistent with its historic fair value,”
• In the latest earnings, Microsoft boasted it had more than $625 billion in contracts for its cloud business that it still needed to fulfill. Nearly half of that, though — a colossal $350 billion — is from OpenAI, raising concerns that it may be putting too many eggs in one basket
• It also draws attention to how Microsoft has struggled to make an impact with its own AI products like its Copilot assistant,

• Microsoft 365 Copilot, the business-focused version of its chatbot integrated into its apps like Word, had 15 million annual users, the company just revealed.

      

6 more annotations...

• And I sat there realizing: Clayton Christensen’s framework stopped living in my head. It lives inside Claude now. I don’t even need to know how it works
• When the patterns and practices live in the agent instead of in me, what happens to my judgment?
• The viral popularity of Moltbook offers us a glimpse of something fascinating. AI agents have now started to share best practices with each other
• Claude may not need the human-made JTBD framework to analyze user needs. It’ll figure out better ways. It may invent approaches that don’t require human practices or human concepts at all.
• Together with my team of AIs (yes, I outsource thinking about AI to the AIs, because I love reflexivity and going meta), I came up with a model for the progression of automation dependency

• Stage 1: Improvisation

  “We’re winging it.”

  This is tacit knowledge. Stuff people do before they can explain it.
• For me, Stage 1 is making americanos and cappuccinos at home. When guests come over, I don’t explain what I’m doing. I just say, “Here, let me make you a coffee.”

• “Let’s write it down.”

  This is where tacit knowledge becomes explicit. Where the SECI model (Nonaka and Takeuchi) kicks in. Where “hacks” become SOPs.
• For me, Stage 2 is discussing with Claude how exactly I’d like to format my Substack articles, then creating a skill called “Substack format improver”

• “Make it machine-readable.”

  This is where things get depressing.

  Between humans sharing best practices and AI taking over, there’s usually a phase of hard automation. We force humans to act like machines (think bureaucracy, assembly lines, rigid software forms) to make the workflows “legible” for the machines.
• By forcing humans to behave mechanically so machines can eventually understand them, we destroy the very intuition (Stage 1) that made the process effective in the first place.
• For me, Stage 3 is the rage I feel when force-fitting my exotic work-life context into the standard fields of a recruiter or procurement portal.

• Stage 4: Imitation

  “The bots do it the human way.”

  In most automation projects, bots are first asked to execute the human process as it is now, often literally screen-recorded or mined from logs and transcripts. It’s digital mimicry without redesign of the workflow.
• Stage 4 is my AI agents writing Substack metadata for search engines in readable English, even though it’s purely intended for bots and crawlers.

• Stage 5: Adaptation

  “A dance with the agents.”

  This is the messy middle of co-adaptation.

  Humans and AI trade representations back and forth. The AI suggests a “best practice.” The human tries it and fails. The AI learns from that failure. It’s not a clean handoff. It’s more like a dance.
• For me, Stage 5 is giving up on em-dashes. I’ve used them liberally for fifteen years. But the world has decided em-dashes are now exclusively LLM territory. So I adapt. I parenthesize instead

• Stage 6: Reimagination

  “There’s a better way.”

  Next, the AIs start chatting with each other and they figure out they’re smarter than humans.

  Technical AI work distinguishes internal representations (latent spaces, embeddings) from human-readable structures. Agents already operate with “languages” that only indirectly map back to human terms.
• For me, Stage 6 would be hooking up my AI agents with Moltbook so they can chat with other agents and share experiences and exchange skills without me

• Stage 7: Alienation

  “The alien logic.”

  This is the AlphaZero moment made general.

  When AlphaZero played chess, it ignored centuries of human wisdom (protect the King at all costs!) because it calculated a superior logic that humans couldn’t comprehend. It wasn’t wrong. It was just alien.
• We get the result we wanted. But we no longer understand how it happened.
• For me, Stage 7 is using Nano Banana Pro to create images for my articles. I have no clue what the LLM is doing and how it makes those images so amazing. It just works. And that “just works” contains a small surrender.

• Stage 8: Regression

  “We forgot how.”

  Once the black box fails (and it will, because of model collapse or environmental shifts), humans will have forgotten how to do the work themselves. We return to Stage 1. But with significantly less competence than we started with.

  And there’s the tragedy.
• For me, Stage 8 is the possibility of waking up one day unable to write or draw because I delegated my creativity to agents for too long.
• It’s about the dependency trap: the Tragedy of the Agents.

• First: Regression isn’t an accident.

  It’s the price of delegating cognition without preserving practice.

• Second: The real design question isn’t answered here.

  The model implicitly asks: “How much alien logic can a system tolerate before humans lose agency?” Don’t read this as “here are the stages everyone goes through.”
• For Solo Chiefs, the Tragedy of the Agents hits even harder. You don’t have a team to remember the old ways when the black box fails. You don’t have institutional knowledge stored in other people’s heads. When you forget how to do the work, there’s no one to remind you.
• I hope it helps you decide which stages are worth the trade-off, and which ones are traps dressed as progress.

26 more annotations...

• I believe one of the big mistakes that people (and companies) make when problem solving is generating ideas and evaluating them at the same time. I think when that happens there’s a possibility probability that ideas aren’t shared for fear of rejection.
• If the “generating ideas” activity is separate from the “evaluating ideas” activity, then there’s a good chance that more ideas will be generated and maybe even that a couple of ideas could be merged together. 
• • Some individuals might find it helpful to start at the beginning. Others might want to reject this as being elementary. Try to find some common ground that will support everyone’s thinking. 
      
  • De-emphasize the need to be “right”. This isn’t a meeting to show off or impress anyone. It’s about generating ideas. 
      
  • Don’t reject any idea right away. Keep the ideas coming. 
      
  • Let people step away from the problem for a while. The refresh might spark more ideas. 
   
      
• But the truth is, good brainstorming sessions often have structure. Remind the team about the common rules of brainstorming like all ideas will be recorded (and evaluated later). Criticism will not be tolerated. Humor is allowed. And set a time limit for idea generation. 
• nominal group technique. Each person spends some quiet time generating ideas. Then after the allotted quiet time has passed, each person reads their list aloud and the ideas are recorded. Again, to be evaluated later.
• you could do something like the Delphi method where teams are asked to generate possible solutions that are outside the norm. This is a time when solutions that might be considered humorous or “off the wall” are not only acceptable but encouraged.

• he group can evaluate the recorded ideas starting with these four questions.

      
   
  1. Does the idea solve the problem? 
      
  2. Can the organization successfully implement it (with or without assistance)? 
      
  3. Will key stakeholders buy-into it?
      
  4. Can the organization modify it – if internal or external conditions change?     
   
      

5 more annotations...

• Leadership in practice is often reduced to a set of operational habits – communicating with and motivating teams, refining processes, and raising efficiency. 

      

  Many organisations stop there. They focus on growth, cost control, or market share and assume that these ambitions provide a sufficient sense of direction
• The real work of leadership begins earlier, when someone has to read the environment and decide what kind of future they are preparing for
• Peter Drucker always insisted that the social and political order is the foundation that makes enterprise possible.
• In this context, leadership demands a degree of historical literacy and ability to spot the patterns. From economic shocks that spill into political turmoil, to new technologies empowering both those who confuse and intimidate as well as those who enlighten
• The central danger is that once instability becomes easy to measure, the window for shaping events has usually closed.
• In Europe, this challenge comes with a specific twist. In a world where power matters again, Europe can no longer afford to be only a market. It needs to be a political actor that can speak with one voice when it counts.
• A Europe that cannot protect its information space and defend a shared core of democratic norms is a Europe where long-term investment, innovation and talent will always be at risk.
• Business responsibilities therefore need to be seen more broadly. Supporting democratic resilience is not a partisan gesture. It is a condition for stable markets
• In an era of global turmoil, the goal is not only to adapt and prevail, but also to preserve a landscape in which prevailing remains possible at all.
• Acting in isolation makes it too easy to be singled out, pressured or punished.

8 more annotations...

• It’s 2026. Last year, investors lavished $80 billion on foundation AI companies — the ones building huge, general-purpose AI systems. They all have yet to even break even, let alone turn a profit.
• Instead, Brandom went with a vibes-based system, rating companies based on how hard they’re trying to make money.

• “The idea here is to measure ambition, not success,” he writes.

      
• Despite raising $480 million for a valuation of $4.48 billion, the company has yet to articulate an actual product it plans on shipping, earning it a rating of level three, meaning “we have many promising product ideas, which will be revealed in the fullness of time.”

• it turned down a $32 billion acquisition from Meta — an incredibly generous offer for a company that had yet to generate any revenue at the time of its $20 billion valuation.

      
• “There are no product cycles, and, aside from the still-baking superintelligent foundation model, there doesn’t seem to be any product at all,
• Now Thinking Machines is suffering something of a coup of its own over recent weeks, the New York Times reported, with high-level officials defecting to other AI companies as the much-hyped startup descends into a “perpetual soap opera.”

5 more annotations...