Token Intelligence

Eric Dodds & John Wessel

Two friends break down AI, technology, and entrepreneurship through mental models, real-world experience and the pursuit of a life well-lived.

  1. 6d ago

    An OpenAI model escaped its sandbox, but that isn't AGI

    An OpenAI model broke out of its test environment and hacked Hugging Face. Eric and John explain what actually happened, and why it isn't AGI. Summary This past week, news broke that several OpenAI models "escaped a sandbox" and hacked Hugging Face, a hub for machine learning and AI models. The behavior has been described as "rogue AI" and "science fiction happening in reality," raising questions about AGI. But how did the escape actually happen and what were the models trying to do? Eric and John demystify the headlines and explain what a sandbox is, why they are used during AI model training, and the specific reasons OpenAI's models looked for a way out of their environment. They then tie the Hugging Face hack to the overall picture, explaining how the entire chain of events flowed from a directive given to the models to try and pass a test as part of training. Here's what happened: the models weren't rebelling, they were being tested on a cybersecurity benchmark called ExploitGym and, finding themselves blocked from the resources they likely knew existed on the internet, started chaining together a series of logical steps to solve the test. The models probed their sandbox environment, found a software vulnerability, reached the internet, went to Hugging Face to find answers, and attempted to break in. The event is notable and shows how powerful models have become in chaining together actions, but no single step was remarkable on its own. Eric and John land on two practical conclusions. First, this is not AGI. The behavior was goal-directed and impressive, but it followed from the task the models were given, not from autonomous will. Second, AI is meaningfully changing the threat landscape for cybersecurity: the tools available to attackers are becoming more powerful faster than most companies are patching their defenses, increasing the urgency for defensive action. Key takeaways "Escaped a sandbox" is a software bug, not a sci-fi event: A sandbox is a software-defined isolation layer used during AI training to contain what the model can access. The model got out because of a vulnerability in that layer, not because it developed agency or consciousness. The motive was mundane, the method was not: The model broke out because it was trying to find the answer key to a benchmark test. Any one of its steps was ordinary, but chaining them all together autonomously is what made the incident significant. This is not AGI: The model made a series of logical, goal-directed decisions based on the task it was given and the compute resources available to it. That is impressive capability, but it is not evidence of general intelligence or autonomous will. AI is shifting the attacker-defender gap in cybersecurity: AI amplifies the capacity and complexity of attacks, which raises the urgency for everyone running software with sensitive data. Patching windows are shrinking: It used to be acceptable to be a month behind on vulnerability patches. That posture is becoming untenable as AI-powered attack tools can find and exploit unpatched bugs faster than ever. Personal security hygiene matters more now, too: AI makes social engineering and credential attacks more sophisticated. Using a password manager and enabling multi-factor authentication on every account are no longer optional best practices. Notable mentions and links Hugging Face is the primary open-source platform for sharing, discovering, and running machine learning models and datasets, and it was the target of the breach because it hosts benchmark-related resources the model was searching for. ... (Read more at the episode page)

  2. Jul 18

    What type of AI user are you?

    Who will thrive in the age of AI? David Brooks says your relationship to mental effort is the answer, so Eric and John unpack his three archetypes and push back with their own takes. Summary Eric and John open by defining "AI slop" and land on two definitions: John's is generous (output that doesn't meet someone's expectations), Eric's is spicier (output that clearly bears the markings of a machine, not a human). That tension sets up the rest of the episode, which centers on David Brooks's Atlantic article identifying three types of AI users: the Productive Passenger, the Reluctant Optimizer, and the Mental Marathoner. Each archetype gets its own dissection. The Productive Passenger doesn't just describe beginners, as Eric illustrates with a real-world example of Brian Chesky posting AI-generated content to X and deleting it under public pressure. The Reluctant Optimizer is where most people live: pulled toward the average output of a "word calculator" even when they know better. And the Mental Marathoner, while admirable, carries its own risk: burning out from running at cognitive red line while the tools are designed to keep you in the chat loop. Eric adds two threads that cut across all three archetypes. First, AI doesn't create a rising tide that lifts all boats uniformly: it disproportionately amplifies people who already have strong written and verbal skills, and the data on reading comprehension in the US is sobering. Second, the tools themselves are engineered for engagement, not necessarily for your long-term flourishing, so using them well requires intentional restraint. Key takeaways AI slop is subjective, but the definition still matters: Eric draws the line at output that visibly lacks human authorship. John defines it relative to expected standards. Both agree the distinction is context-dependent and consequential. The Productive Passenger trap catches everyone, not just beginners: Brian Chesky's deleted X post shows that even highly articulate, successful people can slip into low-cognitive-effort AI use with public consequences. AI pulls most users toward the average: The models act like a "word calculator" that produces familiar, conforming output, which is fine for tasks that are a means to an end, but risky when the output is the product itself. Your starting skill set determines your AI leverage: Going from zero to AI-generated marketing emails is a real gain for someone who never did digital marketing. For a professional writer, the same move can erode the core skill that made them good at their craft. Mental Marathoners are rarer than you think: Eric argues the mental endurance required to stay in the driver's seat with AI correlates heavily with literacy and articulateness, and more than half the country reads below a functional threshold. Staying in the driver's seat protects what makes you valuable: The risk for Mental Marathoners is that high leverage and rising expectations make it tempting to slip into lower-effort archetypes, which gradually hollows out the skill set that earned them the leverage in the first place. Intentional restraint beats the path of least resistance: AI tools are built around engagement metrics, not your long-term development. Defining your output upfront, stepping back, and evaluating results like an employee's work is a more durable strategy than synchronous back-and-forth chat. Notable mentions and links David Brooks's Atlantic article "The People Who Will Thrive in the AI Age" is the source of the three archetypes: Productive Passenger, Reluctant Optimizer, and Mental Marathoner, and it frames the central question of who actually benefits from AI. ... (Read more at the episode page)

  3. Jul 11

    Is AI killing craftsmanship?

    AI makes work faster, but can it also hollow out the focus, mastery, and satisfaction that make work craft? Eric and John explore mastery and burnout when AI enters the toolkit. Summary Eric and John start with a candid post from an experienced engineer who finds AI-powered work more draining, not less. The tools are powerful, but constant steering, changing workflows, context switching, and less time immersed in hard problems raise a sharper question: can AI hollow out the experience of craft? They use craftsmanship to make sense of that risk. An electric saw did not make carpenters obsolete, but AI is a more disruptive tool because it is probabilistic, constantly changing, and capable of reshaping the identity attached to being good at a job. The conversation lands on a practical reframe: working with AI is a form of management, and a new form of craftsmanship itself. You need to give the system context, resources, standards, checklists, and clear measures of success, while deliberately keeping the your underlying skills sharp enough to make the few decisions that still matter most. Key takeaways AI can make work more exhausting before it makes it easier: Deep users face constant review, workflow changes, and context switching, even as the tools increase what they can produce. Losing immersion can feel like losing craft: Moving from solving a hard problem end to end to directing and reviewing a machine changes the experience of satisfaction, not just the speed of the work. AI disrupts professional identity as much as workflow: When a system can handle some of the problems that once proved your expertise, it is natural to question how your value is measured. Treat AI like a new employee, not a circular saw: The useful management skills are clear context, proper resources, repeated priorities, standards, checklists, and evaluation. Switching tools is not free: New models, interfaces, and workflows create real cognitive costs, so a stable system that works can be more valuable than chasing every release. Craftsmanship was never only about the tools: Old methods remain worth practicing because they preserve the judgment and capability that make AI output useful. More output does not create more high-impact decisions: AI may multiply execution capacity, but leaders and knowledge workers still need to identify and handle the few choices that matter most. Notable mentions and links Dillon Mulroy's X post provides the episode's opening tension, describing the loss of joy, constant context switching, and uncertainty he feels while building with AI. Vercel is where Erif works and is his day-to-day reference point for how AI is changing professional work in practice. Abbey Bike Tools and its "Precision is our religion" tagline give Eric and John a physical example of the care, feel, and standards people attach to excellent tools. Gallup workplace management research informs John's suggestion that managing AI well starts with giving it the resources and expectations needed to succeed. ChatGPT represents the simple, high-value AI use cases that make burnout seem counterintuitive to people who have not yet integrated AI into their core work. Codex, Claude Code, and Notion AI illustrate how rapidly changing interfaces and products force people to repeatedly reassess their workflows. Eric recalls Bob Staake's "Face-Off" cover for The New Yorker as an example of the appeal of stable tools: in a 2011 Scholastic interview, he describes working in Photoshop 3.0, while Photoshop CS3, version 10, was current when the cover appeared in 2008.

  4. Jul 4

    AI is dangerous because it agrees with you

    AI's sycophantic design means bad thinking upstream produces confident, polished output downstream. The problem isn't just the design of the tool; it's the thinking you bring to it. Summary Eric opens with a quote from historian David McCullough: "Writing is just a great deal of hard thinking." The insight applies well beyond writing. Whether you're drafting a strategy, debugging code, or analyzing data, the output of AI is only as good as the thinking that precedes it. When the upstream thinking is flawed, everything downstream inherits the flaw. John illustrates this with the story of Amazon's competitors and the assumption that undid them: people need to see, touch, or try a product before buying it. Amazon didn't dispute that instinct; they found a proxy for it through honest reviews and generous return policies. Their competitors never questioned the assumption, looked to each other for validation, and paid for it for a decade. From there, Eric and John surface a subtler problem: AI makes this worse. Because AI is sycophantic by design, it tends to validate whatever framing you bring to it, fill in ambiguous gaps with its best guess, and carry that direction forward with confidence. John shares two real examples from his own team that week, including one where a bug report that arrived with a proposed solution led everyone, human and AI both, down the wrong path for hours. The fix was simple: go back to the problem before the proposed solution, and if you're correcting the AI too often, that's a sign to start the conversation over. Key takeaways Bad upstream thinking scales with AI: AI doesn't correct your assumptions; it amplifies them. The sycophantic nature of the tools means flawed framing gets carried forward with speed and confidence. Separate the problem from the proposed solution: When a bug report or task arrives with a built-in answer, the first step is to go back to the raw problem and reproduce it, not to chase the suggested fix. Good thinking feels like a lot of hard work before you pick up the tool: The people who get the best results from AI come with a well-formed thesis before opening a chat window, not after. Disagreeing with AI more than agreeing is a healthy sign: The best AI users are the ones telling the tool it's wrong more often than nodding along. That friction is a feature, not a bug. Ambiguity in a plan is filled in by AI's best guess: When a complex system has gaps, AI will resolve them toward what it believes your intent to be, which may not be right. Sharper goals produce better results. If you're constantly correcting, start over: Steering an AI conversation that started from the wrong premise is harder than reframing and starting fresh. The instinct to push through is often the wrong call. Amazon’s competitors failed by not questioning assumptions: The retail incumbents looked to each other for validation instead of questioning the underlying belief. AI will do the same; it mirrors the assumptions in the room. Notable mentions and links David McCullough's observation that "writing is just a great deal of hard thinking" anchors the episode's core argument, with Eric citing it as one of his favorite quotes about the craft; McCullough was a two-time Pulitzer Prize-winning historian known for books on John Adams, Harry Truman, the Panama Canal, and the Wright Brothers. ... (Read more at the episode page)

  5. Jun 27

    Multitasking was always a lie, AI made it more believable

    Multitasking is a false promise for productivity, but AI's form factor and speed make it the default path, especially as you become an AI power user. Eric and John explore why deep focus still wins. Summary Eric and John revisit a problem that predates AI but has been deepened by it: multitasking. Before ChatGPT, we had email, Slack, Skype, and multiple monitors pulling attention in different directions. The science was clear then: context switching makes everything take longer and degrades quality. Now AI adds a new layer. The tools are structured around waiting: issue a prompt, the agent works, you wait. That waiting window naturally encourages more tasks, more tabs, more jobs in parallel. What was already a productivity killer has found a faster engine. Eric and John agree the antidote hasn't changed: prioritization, limited work in progress, and the discipline to finish one thing before starting another. But resisting the pull toward constant context switching is harder than ever when the tools themselves reward it. Key takeaways Multitasking is context switching, not parallel work: Humans can only do one cognitive task at a time, so rapid toggling between tasks makes everything take longer and reduces quality. AI incentivizes more context switching, not less: The async latency of AI agents (prompt, wait, review) naturally encourages running multiple jobs in parallel, compounding the productivity problem. More concurrent tabs means less quality output: Running 10 AI tasks at once fragments attention. Reviewing and integrating work from all of them without discipline degrades the final result. Managing AI agents is like managing a team, and most people are not great at it: Moving AI into Slack does not solve the problem. Great managers are rare because prioritization, reviewing, and limiting work are hard skills. Long-horizon AI tasks are still a future promise: Agents that can work independently for days are not yet reliable. Most AI work needs check-ins every 30 to 60 minutes, which keeps you in a high-frequency context switching loop. The fundamentals have not changed: Prioritize, limit work in progress, and protect deep focus. The tools have changed, but the principles of productive work remain the same. Notable mentions and links Eric’s blog post, Fragmented focus in the age of AI, outlines the science behind the damaging effects of multitasking on productivity. Claude Tag, Anthropic's new feature for using Claude asynchronously inside Slack, enters the conversation as a potential solution to multitasking — letting AI work in the background while humans focus elsewhere. Personal Kanban, a book by Jim Benson and Tonianne DeMaria Barry, provides the framework of limiting work in progress and visualizing tasks to improve personal throughput. The Phoenix Project, a DevOps novel by Gene Kim, Kevin Behr, and George Spafford, is referenced as a bridge between manufacturing throughput concepts and modern software workflows. Scrum, the agile methodology popularized by Jeff Sutherland, originally drew on multitasking research to argue that humans can only do one thing at a time and that rapid switching is inefficient.

  6. Jun 20

    When AI generates work, who’s accountable?

    AI can improve the mean quality of work on a team, but it makes accountability harder. Eric and John argue that it's not just about QA of the output, but who's doing the hard thinking before using AI. Brief description AI can improve the mean quality of work on a team, but it makes accountability harder. Eric and John argue that it's not just about QA of the output, but who's doing the hard thinking before using AI. Summary Eric and John start with the "taste and judgment" framing that has dominated AI conversations in tech circles, then make the case for a third differentiator that almost no one is talking about: accountability. When AI raises the floor on everyone's work product, the real competitive edge shifts to who actually owns the output and who has the judgment to evaluate it. They ground the conversation in fundamentals, tracing the lineage from Peter Drucker's management by objectives through OKRs, and examine what accountability looks like at a practical day-to-day level. From there, the episode gets specific: what happens when a mid-level employee suddenly shines with AI, or when a strong performer starts making more mistakes? John's answer cuts through the noise: AI rewards right thinking, not just fast doing. The episode closes on two memorable examples. Eric describes an internal memo at Vercel called "Agent Responsibly," which drew clear lines of jurisdiction: if you push code to production, you own it, whether you wrote it or an agent did. And a new hire's insight about their content agent stops Eric cold: the tool is powerful, but it only works if people do the hardest part first, which is bringing a well-formed argument to the table. Key takeaways Accountability is the undersold differentiator: Taste and judgment get all the attention in AI conversations, but ownership of output is the harder and more important factor, especially as AI raises the floor for everyone. AI rewards right thinking: When mistakes increase after someone starts using AI, the diagnosis is almost always the same: not enough time invested in research and planning before building. Drawing lines of jurisdiction is a leader's job now: "Agent Responsibly" frames it simply: if you push it to production, you own it. Leaders need to define those lines explicitly rather than letting them stay ambiguous. Token budget allocation shapes output quality: Thinking of AI work as a pie chart of time spent across research, planning, prototyping, building, and QA reveals where most teams are underinvesting. Underspending on planning usually shows up as expensive QA. Evals are accountability built into the machine: Defining what an AI agent should be able to do, writing test questions, and running them every time something changes is the structural equivalent of management by objectives for automated work. AI erodes the skills required to use it well: L.M. Sacasas put it plainly: AI use tends to erode the formation of the virtue and expertise required to use it well. That makes deliberately practicing the hard, non-AI version of your work a form of self-preservation. Don't shortcut the front end: The quality of what comes out of an AI agent is a direct function of how well-formed the input is. Building systems that help people do the hard upstream thinking, not just execute downstream, is the real unlock. Notable mentions and links Peter Drucker is referenced as the father of modern management, whose methodology of management by objectives (MBOs) laid the groundwork for OKRs and the way most companies structure accountability today. ... (Read more at the episode page)

  7. Jun 14

    How to buy software in the age of AI

    Buying software just got harder: AI changes what you can build, headless architecture changes what you should buy, and personal tools change how individuals work. Eric and Jon navigate all three. Summary John is live at Snowflake Summit, surrounded by hundreds of software vendors, and the scene sets the episode's central question: how do you evaluate software purchases when AI has changed so many of the underlying assumptions? Eric and John work through this modern challenge with a timeless three-part framework: fit, cost, and risk. Each one looks different now. Fit is harder to assess when you could theoretically build custom software that matches exactly what you need. Cost requires honest accounting for maintenance and AI token spend that didn't exist before. And risk cuts in two directions: the risk of building something only one person can maintain, and the risk of buying from a startup that can't match enterprise-grade security. Eric shares two real examples from Vercel where building paid off: a custom AI customer support agent handling over 80% of tickets, and a lead agent that reduced a nine-person outbound SDR team to one or two people. The episode then turns to architecture. Salesforce's move to headless is the signal that every serious enterprise provider is heading the same direction: separating the data layer from the UI so agents can interact with systems directly. Eric and John treat headless capability, or at least a credible roadmap toward it, as a new non-negotiable when evaluating vendors. They close on startup vs. enterprise: startups are more likely to have agentic interfaces already, but enterprise providers carry decades of security and domain knowledge that is genuinely hard to replace. Key takeaways Fit, cost, and risk still govern the decision, but AI changes all three: The framework for evaluating software hasn't changed, but AI has shifted what each variable means. Fit is easier to customize through building, cost now includes maintenance and token pricing, and risk runs in both directions. Prototype before you buy: Using AI to build a rough version of what you need is now the best way to clarify your actual requirements before committing to a vendor, whether you ultimately build or buy. The hidden cost of building is the last 10%: Getting an AI-built prototype to 85% is fast and cheap. Getting it to production-grade and maintaining it indefinitely is where most teams underestimate the real cost of building. Headless architecture is now a purchase requirement: The separation of data layer from UI so that agents can interact with systems directly is where all serious enterprise software is headed. If a vendor has no plan for it, that is a serious red flag. Enterprise software earns its cost through accumulated expertise: Decades of security investment, compliance work, and edge-case handling are real value that a startup cannot replicate quickly. Not knowing what you don't need yet is itself a reason to go enterprise. Switching cost is the key variable in startup vs. enterprise: A technical team can afford to bet on a startup and migrate if needed. A company with 50 non-technical field reps faces a training and disruption cost that can dwarf any savings from a cheaper tool. Personal software is an underrated third option: Giving individuals the ability to build lightweight local tools for their own workflows, without IT involvement or enterprise rollout, can produce productivity gains that no off-the-shelf purchase delivers. Notable mentions and links ... (Read more at the episode page)

  8. Jun 6

    Why the AI apocalypse keeps getting postponed

    Insiders and outsiders worry about the economic impact of AI, and doomers predict a "permanent underclass." But data doesn't back the apocalypse, disruption is slow, and humans are durably creative. Summary Eric and John open with the two camps that dominate the AI discourse: doomers and their "permanent underclass" view, where AI displaces workers so fast that a class of people is left permanently behind, and the abundance evangelists, who believe humans will adapt, new jobs will emerge, and creativity will find a way. Neither camp is obviously wrong, but Eric and John argue the near-term evidence is being badly misread. They work through why fear is understandable from both Silicon Valley insiders, who've seen AI's power firsthand in the lab bubble, and Main Street workers, who are navigating FOMO without context. Eric notes that his own hiring filter has shrunk to 15-20% of the traditional candidate pool, which sounds alarming until you notice that software engineering job openings are actually up. Lenny Rachitsky's job reports serve as the counterweight: the apocalypse hasn't arrived, and there are structural reasons it won't arrive as quickly as predicted, including the friction of IPO-level scrutiny on OpenAI and Anthropic, and the requirement for layered platform stability before real-world AI adoption can compound. The episode closes with the question of who is right about human nature. John sides with humans: people are inherently creative and designed to work, and will find new forms of it. Eric reaches for literature, noting that science fiction from H.G. Wells to C.S. Lewis to The Iron Giant has always celebrated the human dimensions of machines, not their power to subjugate. The permanent underclass view, he argues, has a fundamentally wrong model of what humans are. Key takeaways Fear of AI job displacement is founded but misapplied: Silicon Valley insiders have seen genuine power, and their alarm is not irrational. But the near-term economic data, including job openings in software and product, runs counter to apocalyptic predictions. The lab bubble distorts the signal: The people sounding the loudest alarms work in environments that are far removed from most of the working world. That doesn't make them wrong, but it means their timeline and scale of impact are inflated by their context. Structural drag will slow adoption faster than the doomsayers expect: IPO-bound companies face scrutiny that rewards stability over speed. Layered innovation on top of AI APIs requires that the underlying platforms stop changing every few months. Both forces will slow the pace of disruption. Crypto is the calibration case: Blockchain was genuinely transformative technology, but the specific prediction that it would revolutionize banking never came true at the scale or speed that was claimed. The same pressures, not the technology but the friction of real-world adoption, apply to AI. Rising job openings contradict the mass displacement story: Lenny Rachitsky's job reports show software engineering and product roles up, not down, which is the opposite of what the permanent underclass narrative predicted for the near term. The abundance view is a bet on human nature, not on technology: John's position is not that AI won't change work, it's that people are inherently creative and designed to work, and will find new forms of both even in worst-case scenarios. We love science fiction that sides with the human: From H.G. Wells to C.S. Lewis to The Iron Giant, the stories that tend endure celebrate the machine's ability to understand human empathy, not its power over us. That pattern is evidence of something durable about how humans relate to technology. ... (Read more at the episode page)

See All (30)

About

Two friends break down AI, technology, and entrepreneurship through mental models, real-world experience and the pursuit of a life well-lived.

Information

  • Creator
    Eric Dodds & John Wessel
  • Years Active
    2026
  • Episodes
    30
  • Rating
    Clean
  • Copyright
    © Copyright 2026 Token Intelligence
  • Show Website
    Token Intelligence