Ethan Mollick on Fable 5.2: Jagged AGI

I love Wharton professor Ethan Mollick‘s One Useful Thing Substack. He wrote a brilliant post reviewing Fable 5.2. The marketing video that Fable made for itself with a 1 shot prompt is clever and not AI sloppy.

Ethan’s conclusion is almost poetic and drawing attention to his inherently human role: “But there are still gaps, and it is the reason my answer is that this is “almost” AGI. Most critically, the AI did not notice these problems itself. It did not see that the course was a slog or that it had tied itself in knots with its own rulebooks. This was not for lack of checking. Fable put critics, fact-checkers or judges on nearly every project, but those checks found errors and added fixes. None of them asked whether the work was too long, too repetitive, or no fun. That all fell on me, and if I hadn’t watched and taken action, the results would have often been less impressive (though not exactly a failure, either). Knowing when your own work is not good enough is a large part of what makes someone an expert, and it is the part Fable does least reliably. I would still not trust it to run a business on its own, even though its plan would have impressed me coming from a team of my MBA students.”

Every time someone sends me a long doc full of Claudespeak my heart sinks. But then, this is why we still need humans to be creative and poetic.

Watch the video and read Ethan’s full post here:

https://www.oneusefulthing.org/p/fable-52-and-jagged-agi

And put what you build with Fable on Spacefast.com 🙂 (sorry I could not resist the promo)

Meet Spacefast, Everything Your AI Needs to Bring Your Ideas to Life

About a year ago, everyone started humble-bragging about vibe-coding their own software. So you built an awesome app with Claude. Vibe-coding the app was easy. But getting it live meant figuring out GitHub and Vercel and Supabase and how to connect them all, and you gave up. So you tried Lovable. Much easier to get started. But you were already paying for Claude, and Claude already had your context.

If that sounds familiar, try Spacefast, the latest product in the Automattic family 🙂

Batuhan İçöz, one of our engineers, built it because he kept being asked by non-technical friends and family for help to get their stuff live. (More of his founder story in our official launch blog that I stole my headline from!)

Go to Spacefast.com, paste the prompt into the AI you already use, and tell it what you want to build. When you’re ready, ask it to launch it on Spacefast. A couple of clicks later, you have a shareable link.

If you’re old like me you remember what WordPress.com did for web publishing, suddenly anyone could launch a blog. Now you and your AI can do that with a deck (HTML is the new PowerPoint!), a website, or a full-stack app. Spacefast comes with hosting, domains, databases, login, privacy, security and all the important but boring stuff.

Spacefast is great for personal side projects – people have been building websites for their hockey team coaching, games and fun apps. Inside Automattic we’ve also been using it for business use cases. We’ve been putting dashboards on Spacefast, and getting our agents to update them from Slack and email. I’ve always found CRM systems overwhelming and annoying to use, now I track my partnership outreach in an app on Spacefast that ChatGPT updates and which our other systems can read.

Every marketing team I talk to these days is working on AEO to show up in Google and ChatGPT answers, and those teams are making and managing a lot more landing pages than they used to. Our marketing teams use Claude with brand guidelines we keep in GitHub to get pages live on Spacefast in hours instead of weeks. (Yes it’s not instant, our humans still check and edit the pages before they go live, we don’t want to publish AI slop.)

You can sign up for Spacefast for free today and try it. We’re also looking for enterprise beta customers to help us build the next set of features. If your marketing team is trying to launch hundreds or thousands of pages and keep them current as products and prices change, get in touch with me on my contact form or LinkedIn or X. I would love to hear how you do it today and what’s a pain.

PS: Spacefast works with Claude (including Claude Desktop, Cowork, and Claude Code), ChatGPT and ChatGPT Work, Codex, Muse, Instinct, Cursor, Grok, Grok Build, Grok Bot, Gemini CLI, GitHub Copilot (in VS Code, the app, and the CLI), Kimi, Devin Desktop (formerly Windsurf), Devin Cloud, Zed, OpenCode, Amp, Warp, Factory Droid, Cline, Continue, Raycast AI, Poke, Pi, Indent, Hermes, OpenClaw, and Earendil. You can connect other MCPs.

Does AI currently benefits the people making AI the most?

I’m no economist but it’s been interesting following the debates about whether and when AI will lead to productivity increases.

Torsten Schlok, Chief Economist at Apollo, a global asset manager, is pretty bearish in his latest analysis of where we are today. Quoting him: “Despite massive investments and widespread hype around artificial intelligence, the technology has delivered zero financial payoff to corporate America outside of tech. Revenue per worker is rising for the Mag 7 but is flat for the S&P 493, see the first chart below. Profit margins are rising for the Mag 7 but are flat for the S&P 493, see the second chart below. And there are still no signs of AI boosting productivity growth, see the third chart below. With no impact on the bottom line, mainstream corporate America is pulling back and starting to view token spending not as a source of growth but as a cost center that weighs on margins.”

The companies whose revenue per employee is going up the most are the magnificent 7 (Alphabet, Amazon, Apple, Meta Platforms, Microsoft, Nvidia and Tesla.

Anthropic”s S1 shows $6.6M per employee, so they are higher than Mag7 except NVIDIA. Nvidia (NVDA) brings in about $7.2 million per employee, by Yahoo Finance calculations. Meta (META) is at $3.6 million. Apple (AAPL), at $3.1 million, brings in less than half as much per worker as Anthropic. Amazon (AMZN), with more than 1.5 million workers, is at $0.5 million.

I’ve joined Automattic, the company behind WordPress.com, WooCommerce and Tumblr

📣 I will be working across marketing, product and partnerships to drive growth. (My marketing friends, let’s chat about co-marketing partnerships!)

After closing down my startup last year, I took my time exploring the AI space, including interesting gigs at the OpenAI Foundation, Notion and Distyl. Then I had a chat with Matt Mullenweg, Automattic’s founder, who asked me what I want to do next. I answered, “Join a mission-driven company, in a growth role that’s not boxed into a specific function. Bring the benefits of AI to everyone, including consumers and SMBs, not just enterprise.” He said, “Why don’t you come do that at Automattic? We need someone like that.” I said yes, with my head, heart and gut.

Here’s why I joined:

→ Mission. WordPress democratised access to publishing, Woo did the same for online stores. Open source, open web and customer freedom are at the heart of the company. I’ve already seen that play out in decisions, even when a ‘closed’ option would make more money.

→ Scale of the opportunity. 40% of the internet runs on WordPress. While the open source version requires technical skills, anyone can build and own their online presence at WordPress.com. Some of the world’s biggest sites run on WordPress VIP for enterprise. WooCommerce powers 4M+ stores. Navigating the AI transition we’re in matters for that whole user base, not just Automattic as a company.

→ People and culture. I have known Matt a long time and always loved his vision and willingness to cause chaos for the greater good, even when it’s the harder path. I’m excited to work with him. The Automatticians I’ve met so far are smart, kind and mission-driven, and there are plenty of people like me who don’t fit neatly into a box. The culture is more jazz band than corporate; I’ll write more about that in a future post.

→ Global. The company has been remote-first since long before COVID (2005!). Our teams are distributed in 81 countries, speaking over 100 languages. Harder to run than having everyone in one building in the Bay Area, but when I worked at Google across Europe, Middle East and Africa, I saw the benefits of a broad footprint and diverse team in preventing tunnel vision.

Me & Matt in Regent’s Park, London, Oct 2021
(with ‘Pineapple’ by British artist Rose Wylie)

My first weeks

I started the day before Matt was briefly ousted as CEO by the board. He came back 33 hours later, faster than Sam Altman! This made for an interesting first week…

The next week I got the chance to meet some of the Woo product, marketing, partnerships and finance leaders at an in-person meet-up. We were back to business, brainstorming ways to grow Woo. I felt right at home as the conversation moved from high-level partner strategy to optimising signup flow detail, with everyone being super action-oriented.

Last week I started my rotation in Automattic’s customer service team, which we call “Happiness Engineering”. The last time I built a WordPress site (this one!) was back in 2021, so I’ve been acting as a customer, playing with newer features like block-based editing and AI agents to build and edit. I’ve been shadowing customer calls with creators and small business owners, including a tutoring marketplace, a student newspaper, a flooring business with a whisky distillery sideline (two separate websites, in case you’re wondering). Watching people who apologise for “not being technical” struggle with our product, then light up when their website comes to life, was a reminder of why we need to make tech work for everyone. Many of these folks are not using AI in their daily lives yet, let alone for their business – living in the Bay Area bubble, it’s easy to forget that.

Shameless plug for our Spacefast, our newest product

Those of you who are in the AI bubble: try Spacefast, our newest product. Publish what you have built with your ChatGPT, Claude, Gemini or agent of choice in seconds, hosted on Automattic’s secure infrastructure. Websites, marketing landing pages, full stack apps – anything! Over 1,000 spaces have already been created. Feedback welcome.

#Automattic #WordPress #WooCommerce #Spacefast #newjob

Your kid’s school doesn’t matter

Auren Hoffman, an entrepreneur and investor who likes to be contrarian, wrote a Substack entitled ‘your kid’s school doesn’t matter’. Auren and I don’t agree on many things, but I wholeheartedly agree with his thesis here.

“if you are a decently well-to-do parent worried about which school your kid is going to attend, here are the four words you need to hear: it will not matter AT ALL. not kindergarten. not middle school. not high school. and definitely not college. there is one caveat. if you are poor, school matters. a great school can massively change the trajectory of a kid whose house does not have books on the shelves or a parent who finishes work after midnight. BUT if you are decently well-off (making over $100,000 a year), have a stable family, and read bedtime stories to your 3-year old — it will not matter.” [sic, Auren does not use caps]

In our family, we’ve been running an a/b test in our family based on our kids preferences. One kid is thriving in San Francisco public middle school, and largely has had great teachers with a couple of exceptions – even though many parents told me that public middle schools in SF are terrible and we can’t possibly send our kid to one. The other kid went to Eton College in England and is now skipping college to do an AI startup. Go figure! 🤷

Enjoy Auren’s full post here: https://auren.substack.com/p/your-kids-school-doesnt-matter

Image courtesy of Auren Hoffman (and probably ChatGPT)

And if you’re interested in the wonkier side of public education policy, read John Bailey’s essay in the Stanford Innovation Review: https://ssir.org/articles/entry/bailey-education-pluralism-trust

“The next era of public education will be judged less by the elegance of its ideas than by whether it responds, with humility and pragmatism, to the people it exists to serve.”

In defense of Rachel Whetstone

I just read the San Francisco Standard article about Rachel Whetstone. I am not sure why the SF Standard felt the need to write this piece just because her husband is running for CA governor, so I am not going to link to it.

I experienced Rachel as smart, fierce, brave, principled, and empathetic and personal when it mattered. I loved that she wore Converse and didn’t give a damn about conforming to female leaders being expected to dress up and be warm when their male counterparts weren’t, because it gave me permission to do the same.

I loved working with Rachel, and with Lorraine Twohill, her equally smart and fierce marketing counterpart who hired me into Google 20 years ago and changed my life. Let’s cheer for these powerhouses, not analyse their every move.

Resilience in the Age of AI | AI’s impact on young people

The OpenAI Foundation split off from OpenAI last year, and is now the biggest private foundation in the world. One of the Foundation’s focus areas is AI Resilience, led by OpenAI co-founder Wojciech Zaremba. Wojciech draws an analogy between AI and another general purpose technology: Fire.

Like fire, AI has many benefits but can also be dangerous. Over time, societies built resilience via many different layers, including fire-resistant materials, hydrant networks, professional fire departments, and building codes. We need a similar resilience ecosystem for AI, and Wojciech is funding it with OpenAI’s sky high share price. The foundation owns 26% of OpenAI stock, which makes them the biggest private foundation in the world, ahead of Novo Nordisk, Gates and Wellcome. They and sold an undisclosed sum in OpenAI’s fundraising round earlier this year to fund their first investments.

In March and April, I got the opportunity to work with Wojciech to help shape their funding strategy for AI’s impact on young people. The initial focus is to fund independent research to better understand where AI can support development, the risks it may introduce, and the contexts that shape those effects. AI has the potential to positively or negatively impact human connection, learning, and development – and without good measurement we won’t be able to steer it responsibly.I’m excited for the first funding announcements, hopefully they will be public soon!

Here’s an excerpt from OpenAI Foundation’s official announcement today that describes the work:

AI’s impact on young people

Young people are often the earliest adopters of new technologies, using them to learn, create, communicate, and explore the world. AI is no exception. But as these tools become an increasing part of young people’s daily lives, it is critical that we develop a stronger evidence base to understand its impacts.

Families, schools, policymakers, and community organizations are all grappling with questions about how and when young people engage with AI—including its impact on human connection, learning, and development. Our initial focus will be on advancing independent research to help guide those decisions—to better understand where AI can support development, the risks it may introduce, and the contexts that shape those effects.

These insights should drive broad safety standards and design principles that guide how any AI product is developed, how schools choose to deploy them, and if and how families decide to incorporate these technologies into their lives.

The Moonshot Podcast: Neurons and Networks

This episode of The Moonshot Podcast explores what the mind can teach us about technology—and what technology can teach us about the mind.

I recently went back to X, The Moonshot Factory to talk with my former boss and friend Astro Teller about Project Amber, the mental health measurement project I ran at X before I left to start Flourish Labs.

Amber was a team of neuroscientists, hardware engineers and ML people who designed an EEG device that could detect and measure mental health conditions like anxiety and depression. I like to think of Project Amber as my first mental health startup…. I learnt so much about the business of mental health that was invaluable for Flourish Labs.

The podcast also spotlights Valkyrie, a never-before-shared X project that explored new ways to build energy-efficient AI models by studying the formation of biological brains – led by Sarah Laszlo, PhD who was also one of the cofounders of Project Amber.

Watch the full podcast on YouTube:

Or listen to it on Apple Podcasts, Spotify, YouTube Music, or wherever you listen to your podcasts.

https://x.company/moonshotpodcast/


In this podcast, Astro explores what the mind can teach us about technology—and what technology can teach us about the mind. He first sits down with Sarah Laszlo from Valkyrie, a never-before-shared project that explored new ways to build energy-efficient AI models by studying the formation of biological brains like the fruit fly’s. He then chats with Obi Felten from Project Amber, a mental health moonshot that aimed to design an affordable EEG device that could detect and measure conditions like anxiety and depression.

0:00 – Intro
01:19 – Inside Project Valkyrie with Sarah Laszlo
03:30 – Mapping a fruit fly’s brain in 3D
05:21 – Project Valkyrie’s moonshot
06:44 – How biological brains could inspire foundation models
09:25 – The Valkyrie team’s first explorations
11:20 – Translating a fruit fly’s brain into a flying model for Wing
14:11 – Reservoir Computing explained simply
15:47 – From insect brains to agriculture
17:36 – Why Valkyrie hit a roadblock
20:37 – How Sarah wound down Valkyrie
22:32 – Up close with a 3D-printed fruit fly neuron
24:40 – Bringing “unbridled joy” to moonshot-taking
27:58 – Inside Project Amber with Obi Felten
29:10 – Why mental health needs a moonshot
32:11 – The “What-If”: Can you actually measure depression?
34:49 – The challenges of mental health measurement
36:17 – Other Amber explorations
39:39 – A look at Amber’s original prototypes
46:50 – The “Kill Criteria.” Knowing when to walk away
50:33 – Sharing Amber’s findings with the world
52:56 – Flourish Labs: Obi’s next chapter
54:45 – Obi’s early memories from X
57:44 – Final thoughts

Ethan Mollick: “People like AI when they use it themselves; they don’t like AI writ large”

Ethan Mollick, associate professor of management at Wharton, is one of my favourite thinkers on AI impact on society, including work and education. He recently did an interview with Tom Rachman, an AI Policy Writer at Google DeepMind. The questions they discussed really resonated with me.

How can companies get their employees to use artificial intelligence when human intelligence remains sharp enough to know that this risks replacing jobs?

How should education revise itself for the ever-revising technological world that students emerge into?

And how to understand the love/hate relationship so many people have with AI?


I think about these questions all the time – in discussion with my kids, friends and colleagues, when I give talks on AI.

I agree with Ethan when he says: “This is a leadership problem. The incentives have to be aligned properly. Currently, it’s, ‘I’m going to automate your jobs away’ or ‘I’m not going to share with you any of the gains the company gets.’ People are exquisitely tuned to rewards. So it’s about leaders articulating a vision of what the world looks like with AI for employees. ‘What should I expect to do? How are people rewarded for doing the right thing? If they automate 90 percent of my job, what happens to me?’ Without those answers, everything else is secondary.”

Nobody has the answers right now, but it’s our job to figure it out with our teams.

Another question is how we teach entry level workers expertise, which traditionally was taught by doing. If using AI well requires taste, how can you acquire that taste if your job is outsourced to AI?

Ethan talks about people needing to be t-shaped – broad and adaptable, and deep on one area. I am a t-shaped person (generalist with product depth) and I often felt I didn’t really fit in, because companies tend to organise functionally and prioritise subject matter expertise.

I also love being a ‘super IC’ with a small team, while my coworkers built empires. At some point to get promoted to VP at Google your role had to have hundreds of reports. I bet they don’t have that rule anymore 🙂

So personally I am excited that the old conventions of management are getting blown up. But I realise that many people find it scary because the existing system suited them.

Read the full interview here: https://www.aipolicyperspectives.com/p/q-and-a-with-ethan-mollick

And Ethans’s Substack here: https://www.oneusefulthing.org/

Image credits: AI Policy Perspectives

The Swarm Org: Managing Humans and Agents

AI is profoundly changing the way we work with each model release. My tasks that Claude struggled with a few months ago are now done in Cowork in a few minutes. The new Claude model 4.6 is insanely good. Bigger context window, ability find and retrieve info in this context window, ability to have several agents collaborate massively changes the game again. It speed things up, and needs less human intervention.

Watch Nate B. Jones’ take on Opus 4.6 below, which inspired me to write this post. He is mostly talking about coding in the first part, then how this flows through to non-technical users like me who can suddenly do tasks that used to take us days in hours or minutes, and how it will change orgs.

Our teams are becoming a mix of humans and agents, working in swarms.

To get the most out of AI, knowing what you want, being clear on your intent and communicating it clearly is crucial. Which truthfully is not that different from successfully managing human teams, it’s just that many managers suck at this. The difference is that now we’re managing humans and agents in parallel.

Most commentators focus on the fact that agents work 24 hours, are endlessly patient, don’t need a pay raise or a promotion. I’m more interested in what this shift will do to the way we work and lead, and to organisational design.

Great judgement and the ability to assess output matters more than execution. The ability to interpret metrics matters more than doing the analysis.

Job titles matter less. Hierarchy matters less. It will be replaced by swarms of people and agents working cross-functionally regardless of where they sit in the org, rather than people working up and down the reporting chain within a function. Process matters less. Endless email and Slack threads coordinating tasks will be greatly reduced.

Team size as a leverage is turned upside down – smaller teams that use agents will beat bigger teams that don’t. What matters is how good are your people at managing agents that work for them, and what skills they need so you can keep the human:agent ratio small. This is why the best AI native companies have higher revenues per employee, compared to traditional companies or the even the best Saas companies. The best AI-native companies, ones that have truly mastered agent orchestration like Cursor ($3M/employee), are achieving 6x the efficiency of the best SaaS companies like Salesforce and Notion ($500K/employee). (A counter-example: Google’s already at $2M per employee without really trying – and they own the infrastructure that powers AI. I’m long on Google.)

Deeply understanding what users and customers want and building great products matters. The best way to do that is to co-create with them – helping customers figure out how to orchestrate agents for their user cases. Domain expertise matters, especially in areas like healthcare that are heavily regulated. Things that seemed impossible suddenly become possible, at much faster speeds.

So what does this mean for you, me, all of us?

Personally, I welcome that change. I’ve always struggled with hierarchy, process, set ways of doing things. I’ve always loved working across functions, with people who didn’t report to me. I’ve always loved working with small teams, while my peers built large empires to get promoted. (At some point Google had a requirement for VP scope that it would involve managing 1000 people – how antiquated that seems now…) My natural management style is hands-off – I describe the outcome I want and ask my team to run towards it, with little direction. I’ve always been frustrated by slow speed – that’s one reason I’m leaving healthcare and going back to tech. And I’ve always tried to do things that seemed impossible. I’ve often failed because the system was just too hard to beat – including with my healthcare startup. Maybe I’ll fail less in an AI + human world 🙂

If you’re scared, you are not alone – I’m scared too. Change at this pace is bewildering. These changes to hierarchy, process, team size will be painful for millions of middle managers and individual contributors whose jobs were built around those structures. This transition will be rough. Not everyone will win. That’s a hard truth we can’t ignore, and we haven’t even begun to figure out how to ensure that people are not left behind, and that as many people as possible benefit from the way AI will change the world of work.

But worrying about what we can’t control won’t help. What we can control is whether we start learning now or get left behind. My coding friends have been using Codex and Claude Code for ages, and I’ve been jealous. Claude Cowork was a breakthrough for me – suddenly I could deploy Claude’s prowess that was built for coding to my everyday tasks like writing a company overview for investors, making a financial model, making sure I don’t miss emails from people I care about now that my inbox is full of AI slop, posting stuff on Facebook marketplace, monitoring job websites for interesting things that pop up. You can also try Dreamer.com, an agent framework for consumers just launched by my friends David Singleton and Hugo Barra (yes, this is a blatant plug for them).

If you’re a parent, help your kids become discerning users of AI, help them to build judgement. Kids are much better at saying what they want, so maybe they can teach you a thing or two about how to write a good prompt!

If you are a people leader, your most important task right now is to help your team members through this transition. Nothing else matters. And that requires that you change your personal working style – you have to get hands on with AI in your work yourself, not wait for your team to figure it out. Experiment. Give AI tasks you thought were impossible for AI to do, not just toy tasks. If they fail, wait for the next model release and try again, don’t assume AI will never be able to do this.

The skills you personally build and that your teams and your kids build in the next few months will pay off for years to come.