Language of AI

  1. Show the AI the spelling you want, not the list you ban

    British spellings kept turning up in my personal notes repo, written by my AI assistants. I added a checker script that flagged them by name, and the problem seemed to get worse. I fixed 26 by hand in scripts and tests and assumed it was a bad day.

    When I counted, 73 of roughly 84 instances sat in files about the rule, the checker included. This reminded me that a model is prone to emulate the words around it, so a banned list in the file it’s editing can work as a list of candidates to use.

    I’m now showing it the spellings I want and keeping the banned list out of sight. If you steer an AI with examples of what not to do, what happens when you show it what to do?

    A drawing studio split down the middle. On the left, a figure at an easel sketches a lopsided teacup in a cluttered room whose walls are covered with pinned sketches of that same teacup, each crossed out in red. On the right, the same figure in a calm, bare room copies one clean teacup sketch pinned at eye level, and the drawing matches it.

    Also on LinkedIn · Reply there

  2. Valuation vs. capability: the Anthropic IPO

    When a company goes public, investors put a price on what already exists plus what they hope it becomes. For Anthropic, reportedly going public this fall, what already exists is an offering (models and harness) that anyone can try for a modest fee.

    Think of a home appraisal. If it drops after the market turns, the roof still keeps the rain out. I’m fairly sure the IPO price won’t change what I can do with Claude code on Monday morning either.

    So don’t let a lofty IPO or the stock’s performance influence your decision to learn to work with this technology. Keep the valuation and what the capability can do for your day to day distinct in your mind…unless you think AI and LLMs will disappear soon.

    What’s a moment with AI when you felt like you’d just done something that wasn’t possible for you before?

    A similar post last year: The AI bubble: valuation vs. potential

    A lit house at night, a person working at a laptop inside, under storm clouds crossed by a glowing stock line that spikes and falls, with a blank price tag hanging from the sky.

    Also on LinkedIn · Reply there

  3. Role convergence doesn’t mean role elimination

    Role convergence doesn’t mean role elimination. I think the boundaries and expectations around our contributions are evolving.

    I’m not an engineer, and I’ve been using agents to do things like understand codebases and prototype interactive concepts and interfaces. I’ve seen this accelerate dialogue and give early conversations a more meaningful grounding.

    As a more specific example, outside work, I’ve set up a personal MCP server based on an open source pattern. That’s something I can only do with a capable coding agent.

    The perspectives here have me thinking about how teams adapt while specialist expertise continues to matter.

    Where are you seeing this convergence, and how are expectations changing?

    Role convergence AI is broadening what people can contribute across product, design, and engineering
    • The investor Marc Andreessen Co-founder
      Andreessen Horowitz
      “The job persists longer than the individual tasks.”
      “You become a super relevant specialist in the combination of the domains.” 1
    • The builder Boris Cherny Creator & Head of Claude Code
      Anthropic

      On his team, a designer can implement interface changes directly, without asking an engineer to make each adjustment. 2

    • The operator Sal Khan Founder & CEO
      Khan Academy

      Khan Academy lets designers and PMs propose code changes, primarily for prototypes and proofs of concept. 3

    • The product coach Marty Cagan Founder, Partner – Product
      Silicon Valley Product Group
      “Most serious products will need a range of skillsets, even with very talented people using the best tools.” 4
    • The researcher Ethan Mollick Professor, The Wharton School
      University of Pennsylvania

      In a P&G experiment with 791 professionals, AI helped participants work beyond their usual functional focus. 5

    • The product leader Nate B Jones AI analyst, former Head of Product
      Amazon Prime Video
      “The top of the [PM’s] funnel used to be words […] but now it’s got working tools and dashboards and automations and agents.” 6
    1. Andreessen, Lenny’s Podcast, Jan 2026
    2. Cherny, Bloomberg Odd Lots, Jul 2026
    3. Khan, Silicon Valley Girl, Jul 2026
    4. Cagan, “The Era of the Product Creator,” May 2025
    5. Dell’Acqua et al., “The Cybernetic Teammate,” Jun 2026
    6. Jones, “Product Management After Software Becomes Cheap,” May 2026

    Evidence 2, 3 and 5 is paraphrased. 1, 4 and 6 are quotations.

    Also on LinkedIn · Reply there

  4. You’re becoming a new type of leader

    You’re becoming a new type of leader, even if you’re an individual contributor. With AI, I increasingly think the job can look less like playing one instrument and more like conducting.

    An investment professional can move from editing slide wording to delegating an initial financial model. A small-business owner can have an agent use a cloud browser to log into siloed vendor portals, reconcile operational and receivables data, and build a dashboard. A product manager can ask an agent to update the interface and submit code for review.

    The agents are the musicians and their tools are the instruments. The goal and constraints are the score. Give them the wrong instrument, key, or tempo and the ensemble may produce polished work that misses what the audience needed, while the conductor is still accountable for what reaches them.

    How much of your work now feels less like playing the instrument and more like conducting the ensemble?

    A conductor on stage leads a human string quartet on the left and a glowing green orchestra of translucent figures on the right, with an audience in the foreground.

    Also on LinkedIn · Reply there

  5. Your employer’s AI setup is shaping what you think AI can do

    Your employer’s AI setup is shaping what you think AI can do.

    Different people I talk to informally have very different Claude or ChatGPT capabilities available to them at work. One (non-engineering) investment professional shared how they have ChatGPT Work and Codex. Someone else in product mgmt in financial services said they have Claude chat with a couple of connectors enabled and the ability to make personal skills, but no Claude CoWork and no Claude Code, seemingly provided only to those with traditional engineering job titles.

    That makes it easy to draw conclusions from your own day-to-day experience that aren’t true for someone else, somewhere else.

    In my experience, having a paid personal subscription for side projects outside of work has really helped broaden and accelerate my understanding of what’s possible, especially as the capabilities keep changing.

    What differences have you noticed when casually chatting with people who work somewhere else?

    Also on LinkedIn · Reply there

  6. No API, no MCP, no problem

    No api, no mcp, no problem!

    Now in 2026, a lot of software already has an interface AI can use: the browser.

    With ChatGPT Work, I can start a task from the ChatGPT app on my iPhone and go about my day. The agent uses its own browser on a computer in the cloud, rather than driving one on my laptop.

    For a side project, I’ve been using this for data extraction from within an outdated portal site that would otherwise be manual work I’d do, but I think the use cases are much broader. A lot becomes possible when software built for humans becomes accessible to an agent through the browser.

    What would you try?

    Phone screenshot of a ChatGPT agent reporting that the ICC Evaluation Service publishes a searchable directory and PDF reports but no public API or bulk export, and that it is moving to the live directory, with a thumbnail of the Reports Directory page.
    Phone screenshot of the same agent searching for deck fastener reports, with a thumbnail of the ICC-ES Find Reports filter panel.
    Phone screenshot of the agent narrowing results to fastener reports and opening the PDF for listing report ESL-1284.

    Also on LinkedIn · Reply there

  7. “Code” in the name doesn’t mean developers only

    Next time you’re tempted to assume “code” in the name means developers only, remember that many people, including non-engineers, are finding a lot of value using Codex, and/or Claude Code, as a harness for non-software tasks.

    I use Claude Code for some personal projects, without an engineering background or task.

    What have you used these tools for?

    Successful coding tasks by occupation Share of coding sessions (%)
    • Software & math 34 / 60
    • Management 37 / 55
    • Legal 33 / 57
    • Business & Finance 29 / 54
    • Healthcare 28 / 56
    • Arts, Design & Media 28 / 53
    • Sales 28 / 51
    • Architecture & Engineering 27 / 54
    • Education 27 / 52
    • Sciences 27 / 54
    Show as a table
    OccupationVerified (%)Judged (%)
    Software & math3460
    Management3755
    Legal3357
    Business & Finance2954
    Healthcare2856
    Arts, Design & Media2853
    Sales2851
    Architecture & Engineering2754
    Education2752
    Sciences2754

    Data: Anthropic (claudeai on Instagram). Redrawn for this site.

    Also on LinkedIn · Reply there

  8. The AI bubble: valuation vs. potential

    From an equity valuation standpoint, yes I think we are probably in an AI bubble. But if and when it bursts, I think it will be more of a reflection on aggressive business valuation assumptions than the technology’s actual transformative potential.

    In other words, as we often hear with emerging tech., we tend to overestimate the impact in the short-term, and underestimate the long-term impact.

    So don’t let any change in the valuations of certain companies change your mind on the importance of learning how AI actually works, and how to use it effectively, one day at a time.

    Shared: Are we in an AI bubble? What happens if it bursts , The Wall Street Journal

    Also on LinkedIn · Reply there