Skip to content

70 posts tagged with “technology”

17 min read
Conceptual 3D illustration of stacked digital notebooks with a pen on top, overlaid on colorful computer code patterns.

Why We Still Need a HyperCard for the AI Era

I rewatched the 1982 film TRON for the umpteenth time the other night with my wife. I have always credited this movie as the spark that got me interested in computers. Mind you, I was nine years old when this film came out. I was so excited after watching the movie that I got my father to buy us a home computer—the mighty Atari 400 (note sarcasm). I remember an educational game that came on cassette called “States & Capitals” that taught me, well, the states and their capitals. It also introduced me to BASIC, and after watching TRON, I wanted to write programs!

Vintage advertisement for the Atari 400 home computer, featuring the system with its membrane keyboard and bold headline “Introducing Atari 400.”

The Atari 400’s membrane keyboard was easy to wipe down, but terrible for typing. It also reminded me of fast food restaurant registers of the time.

Back in the early days of computing—the 1960s and ’70s—there was no distinction between users and programmers. Computer users wrote programs to do stuff for them. Hence the close relationship between the two that’s depicted in TRON. The programs in the digital world resembled their creators because they were extensions of them. Tron, the security program that Bruce Boxleitner’s character Alan Bradley wrote, looks like its creator. Clu looked like Kevin Flynn, played by Jeff Bridges. Early in the film, a compound interest program who was captured by the MCP’s goons says to a cellmate, “if I don’t have a User, then who wrote me?”

Scene from the 1982 movie TRON showing programs in glowing blue suits standing in a digital arena.

The programs in TRON looked like their users. Unless the user was the program, which was the case with Kevin Flynn (Jeff Bridges), third from left.

I was listening to a recent interview with Ivan Zhao, CEO and cofounder of Notion, in which he said he and his cofounder were “inspired by the early computing pioneers who in the ’60s and ’70s thought that computing should be more LEGO-like rather than like hard plastic.” Meaning computing should be malleable and configurable. He goes on to say, “That generation of thinkers and pioneers thought about computing kind of like reading and writing.” As in accessible and fundamental so all users can be programmers too.

The 1980s ushered in the personal computer era with the Apple IIe, Commodore 64, TRS-80, (maybe even the Atari 400 and 800), and then the Macintosh, etc. Programs were beginning to be mass-produced and consumed by users, not programmed by them. To be sure, this move made computers much more approachable. But it also meant that users lost a bit of control. They had to wait for Microsoft to add a feature into Word that they wanted.

Of course, we’re coming back to a full circle moment. In 2025, with AI-enabled vibecoding, users are able to spin up little custom apps that do pretty much anything they want them to do. It’s easy, but not trivial. The only interface is the chatbox, so your control is only as good as your prompts and the model’s understanding. And things can go awry pretty quickly if you’re not careful.

What we’re missing is something accessible, but controllable. Something with enough power to allow users to build a lot, but not so much that it requires high technical proficiency to produce something good. In 1987, Apple released HyperCard and shipped it for free with every new Mac. HyperCard, as fans declared at the time, was “programming for the rest of us.”

HyperCard—Programming for the Rest of Us

Black-and-white screenshot of HyperCard’s welcome screen on a classic Macintosh, showing icons for Tour, Help, Practice, New Features, Art Bits, Addresses, Phone Dialer, Graph Maker, QuickTime Tools, and AppleScript utilities.

HyperCard’s welcome screen showed some useful stacks to help the user get started.

Bill Atkinson was the programmer responsible for MacPaint. After the Mac launched, and apparently on an acid trip, Atkinson conceived of HyperCard. As he wrote on the Apple history site Folklore:

Inspired by a mind-expanding LSD journey in 1985, I designed the HyperCard authoring system that enabled non-programmers to make their own interactive media. HyperCard used a metaphor of stacks of cards containing graphics, text, buttons, and links that could take you to another card. The HyperTalk scripting language implemented by Dan Winkler was a gentle introduction to event-based programming.

There were five main concepts in HyperCard: cards, stacks, objects, HyperTalk, and hyperlinks. 

  • Cards were screens or pages. Remember that the Mac’s nine-inch monochrome screen was just 512 pixels by 342 pixels.
  • Stacks were collections of cards, essentially apps.
  • Objects were the UI and layout elements that included buttons, fields, and backgrounds.
  • HyperTalk was the scripting language that read like plain English.
  • Hyperlinks were links from one interactive element like a button to another card or stack.

When I say that HyperTalk read like plain English, I mean it really did. AppleScript and JavaScript are descendants. Here’s a sample logic script:

if the text of field "Password" is "open sesame" then
  go to card "Secret"
else
  answer "Wrong password."
end if

Armed with this kit of parts, users were able to use this programming “erector set” and build all sorts of banal or wonderful apps. From tracking vinyl records to issuing invoices, or transporting gamers to massive immersive worlds, HyperCard could do it all. The first version of the classic puzzle adventure game, Myst was created with HyperCard. It was comprised of six stacks and 1,355 cards. From Wikipedia:

The original HyperCard Macintosh version of Myst had each Age as a unique HyperCard stack. Navigation was handled by the internal button system and HyperTalk scripts, with image and QuickTime movie display passed off to various plugins; essentially, Myst functions as a series of separate multimedia slides linked together by commands.

Screenshot from the game Myst, showing a 3D-rendered island scene with a ship in a fountain and classical stone columns.

The hit game Myst was built in HyperCard.

For a while, HyperCard was everywhere. Teachers made lesson plans. Hobbyists made games. Artists made interactive stories. In the Eighties and early Nineties, there was a vibrant shareware community. Small independent developers who created and shared simple programs for a postcard, a beer, or five dollars. Thousands of HyperCard stacks were distributed on aggregated floppies and CD-ROMs. Steve Sande, writing in Rocket Yard:

At one point, there was a thriving cottage industry of commercial stack authors, and I was one of them. Heizer Software ran what was called the “Stack Exchange”, a place for stack authors to sell their wares. Like Apple with the current app stores, Heizer took a cut of each sale to run the store, but authors could make a pretty good living from the sale of popular stacks. The company sent out printed catalogs with descriptions and screenshots of each stack; you’d order through snail mail, then receive floppies (CDs at a later date) with the stack(s) on them.

Black-and-white screenshot of Heizer Software’s “Stack Exchange” HyperCard catalog, advertising a marketplace for stacks.

Heizer Software’s “Stack Exchange,” a marketplace for HyperCard authors.

From Stacks to Shrink-Wrap

But even as shareware tiny programs and stacks thrived, the ground beneath this cottage industry was beginning to shift. The computer industry—to move from niche to one in every household—professionalized and commoditized software development, distribution, and sales. By the 1990s, the dominant model was packaged software that was merchandised on store shelves in slick shrink-wrapped boxes. The packaging was always oversized for the floppy or CD it contained to maximize visual space.

Unlike the users/programmers from the ’60s and ’70s, you didn’t make your own word processor anymore, you bought Microsoft Word. You didn’t build your own paint and retouching program—you purchased Adobe Photoshop. These applications were powerful, polished, and designed for thousands and eventually millions of users. But that meant if you wanted a new feature, you had to wait for the next upgrade cycle—typically a couple of years. If you had an idea, you were constrained by what the developers at Microsoft or Adobe decided was on the roadmap.

The ethos of tinkering gave way to the economics of scale. Software became something you consumed rather than created.

From Shrink-Wrap to SaaS

The 2000s took that shift even further. Instead of floppy disks or CD-ROMs, software moved into the cloud. Gmail replaced the personal mail client. Google Docs replaced the need for a copy of Word on every hard drive. Salesforce, Slack, and Figma turned business software into subscription services you didn’t own, but rented month-to-month.

SaaS has been a massive leap for collaboration and accessibility. Suddenly your documents, projects, and conversations lived everywhere. No more worrying about hard drive crashes or lost phones! But it pulled users even farther away from HyperCard’s spirit. The stack you made was yours; the SaaS you use belongs to someone else’s servers. You can customize workflows, but you don’t own the software.

Why Modern Tools Fall Short

For what started out as a note-taking app, Notion has come a long way. With its kit of parts—pages, databases, tags, etc.—it’s highly configurable for tracking information. But you can’t make games with it. Nor can you really tell interactive stories (sure, you can link pages together). You also can’t distribute what you’ve created and share with the rest of the world. (Yes, you can create and sell Notion templates.)

No productivity software programs are malleable in the HyperCard sense. 

[IMAGE: Director]

Of course, there are specialized tools for creativity. Unreal Engine and Unity are great for making games. Director and Flash continued the tradition started by HyperCard—at least in the interactive media space—before they were supplanted by more complex HTML5, CSS, and JavaScript. Objectively, these authoring environments are more complex than HyperCard ever was.

The Web’s HyperCard DNA

In a fun remembrance, Constantine Frantzeskos writes:

HyperCard’s core idea was linking cards and information graphically. This was true hypertext before HTML. It’s no surprise that the first web pioneers drew direct inspiration from HyperCard – in fact, HyperCard influenced the creation of HTTP and the Web itself​. The idea of clicking a link to jump to another document? HyperCard had that in 1987 (albeit linking cards, not networked documents). The pointing finger cursor you see when hovering over a web link today? That was borrowed from HyperCard’s navigation cursor​.

Ted Nelson coined the terms “hypertext” and “hyperlink” in the mid-1960s, envisioning a world where digital documents could be linked together in nonlinear “trails”—making information interwoven and easily navigable. Bill Atkinson’s HyperCard was the first mass-market program that popularized this idea, even influencing Tim Berners-Lee, the father of the World Wide Web. Berners-Lee’s invention was about linking documents together on a server and linking to other documents on other servers. A web of documents.

Early ViolaWWW hypermedia browser from 1993, displaying a window with navigation buttons, URL bar, and hypertext description.

Early web browser from 1993, ViolaWWW, directly inspired by the concepts in HyperCard.

Pei-Yuan Wei, developer of one of the first web browsers called ViolaWWW, also drew direct inspiration from HyperCard. Matthew Lasar writing for Ars Technica:

“HyperCard was very compelling back then, you know graphically, this hyperlink thing,” Wei later recalled. “I got a HyperCard manual and looked at it and just basically took the concepts and implemented them in X-windows,” which is a visual component of UNIX. The resulting browser, Viola, included HyperCard-like components: bookmarks, a history feature, tables, graphics. And, like HyperCard, it could run programs.

And of course, with the built-in source code viewer, browsers brought on a new generation of tinkerers who’d look at HTML and make stuff by copying, tweaking, and experimenting.

The Missing Ingredient: Personal Software

Today, we have low-code and no code tools like Bubble for making web apps, Framer for building web sites, and Zapier for automations. The tools are still aimed at professionals though. Maybe with the exception of Zapier and IFTTT, they’ve expanded the number of people who can make software (including websites), but they’re not general purpose. These are all adjacent to what HyperCard was.

(Re)enter personal software.

In an essay titled “Personal software,” Lee Robinson wrote, “You wouldn’t search ‘best chrome extensions for note taking’. You would work with AI. In five minutes, you’d have something that works exactly how you want.”

Exploring the idea of “malleable software,” researchers at Ink & Switch wrote:

How can users tweak the existing tools they’ve installed, rather than just making new siloed applications? How can AI-generated tools compose with one another to build up larger workflows over shared data? And how can we let users take more direct, precise control over tweaking their software, without needing to resort to AI coding for even the tiniest change? None of these questions are addressed by products that generate a cloud-hosted application from a prompt.

Of course, AI prompt-to-code tools have been emerging this year, allowing anyone who can type to build web applications. However, if you study these tools more closely—Replit, Lovable, Base44, etc.—you’ll find that the audience is still technical people. Developers, product managers, and designers can understand what’s going on. But not everyday people.

These tools are still missing ingredients HyperCard had that allowed it to be in the general zeitgeist for a while, that enabled users to be programmers again.

They are:

  • Direct manipulation
  • Technical abstraction
  • Local apps

What Today’s Tools Still Miss

Direct Manipulation

As I concluded in my exhaustive AI prompt-to-code tools roundup from April, “We need to be able to directly manipulate components by clicking and modifying shapes on the canvas or changing values in an inspector.” The latency of the roundtrip of prompting the model, waiting for it to think and then generate code, and then rebuild the app is much too long. If you don’t know how to code, every change takes minutes, so building something becomes tedious, not fun.

Tools need to be a canvas-first, not chatbox-first. Imagine a kit of UI elements on the left that you can drag onto the canvas and then configure and style—not unlike WordPress page builders. 

AI is there to do the work for you if you want, but you don’t need to use it.

Hand-drawn sketch of a modern HyperCard-like interface, with a canvas in the center, object palette on the left, and chat panel on the right.

My sketch of the layout of what a modern HyperCard successor could look like. A directly manipulatable canvas is in the center, object palette on the left, and AI chat panel on the right.

Technical Abstraction

For gen pop, I believe that these tools should hide away all the JavaScript, TypeScript, etc. The thing that the user is building should just work.

Additionally, there’s an argument to be made to bring back HyperTalk or something similar. Here is the same password logic I showed earlier, but in modern-day JavaScript:

const password = document.getElementById("Password").value;

if (password === "open sesame") {
  window.location.href = "secret.html";
} else {
  alert("Wrong password.");
} 

No one is going to understand that, much less write something like it.

One could argue that the user doesn’t need to understand that code since the AI will write it. Sure, but code is also documentation. If a user is working on an immersive puzzle game, they need to know the algorithm for the solution. 

As a side note, I think flow charts or node-based workflows are great. Unreal Engine’s Blueprints visual scripting is fantastic. Again, AI should be there to assist.

Unreal Engine Blueprints visual scripting interface, with node blocks connected by wires representing game logic.

Unreal Engine has a visual scripting interface called Blueprints, with node blocks connected by wires representing game logic.

Local Apps

HyperCard’s file format was “stacks.” And stacks could be compiled into applications that can be distributed without HyperCard. With today’s cloud-based AI coding tools, they can all publish a project to a unique URL for sharing. That’s great for prototyping and for personal use, but if you wanted to distribute it as shareware or donation-ware, you’d have to map it to a custom domain name. It’s not straightforward to purchase from a registrar and deal with DNS records.

What if these web apps can be turned into a single exchangeable file format like “.stack” or some such? Furthermore, what if they can be wrapped into executable apps via Electron?

Rip, Mix, Burn

Lovable, v0, and others already have sharing and remixing built in. This ethos is great and builds on the philosophies of the hippie computer scientists. In addition to fostering a remix culture, I imagine a centralized store for these apps. Of course, those that are published as runtime apps can go through the official Apple and Google stores if they wish. Finally, nothing stops third-party stores, similar to the collections of stacks that used to be distributed on CD-ROMs.

AI as Collaborator, Not Interface

As mentioned, AI should not be the main UI for this. Instead, it’s a collaborator. It’s there if you want it. I imagine that it can help with scaffolding a project just by describing what you want to make. And as it’s shaping your app, it’s also explaining what it’s doing and why so that the user is learning and slowly becoming a programmer too.

Democratizing Programming

When my daughter was in middle school, she used a site called Quizlet to make flash cards to help her study for history tests. There were often user-generated sets of cards for certain subjects, but there were never sets specifically for her class, her teacher, that test. With this HyperCard of the future, she would be able to build something custom in minutes.

Likewise, a small business owner who runs an Etsy shop selling T-shirts can spin up something a little more complicated to analyze sales and compare against overall trends in the marketplace.

And that same Etsy shop owner could sell the little app they made to others wanting the same tool for for their stores.

The Future Is Close

Scene from TRON showing a program with raised arms, looking upward at a floating disc in a beam of light.

Tron talks to his user, Alan Bradley, via a communication beam.

In an interview with Garry Tan of Y Combinator in June, Michael Truell, the CEO of Anysphere, which is the company behind Cursor, said his company’s mission is to “replace coding with something that’s much better.” He acknowledged that coding today is really complicated:

Coding requires editing millions of lines of esoteric formal programming languages. It requires doing lots and lots of labor to actually make things show up on the screen that are kind of simple to describe.

Truell believes that in five to ten years, making software will boil down to “defining how you want the software to work and how you want the software to look.”

In my opinion, his timeline is a bit conservative, but maybe he means for professionals. I wonder if something simpler will come along sooner that will capture the imagination of the public, like ChatGPT has. Something that will encourage playing and tinkering like HyperCard did.

There’s a third sequel to TRON that’s coming out soon—TRON: Ares. In a panel discussion in the 5,000-seat Hall H at San Diego Comic-Con earlier this summer, Steven Lisberger, the creator of the franchise provided this warning about AI, “Let’s kick the technology around artistically before it kicks us around.” While he said it as a warning, I think it’s an opportunity as well.

AI opens up computer “programming” to a much larger swath of people—hell, everyone. As an industry, we should encourage tinkering by building such capabilities into our products. Not UIs on the fly, but mods as necessary. We should build platforms that increase the pool of users from technical people to everyday users like students, high school teachers, and grandmothers. We should imagine a world where software is as personalizable as a notebook—something you can write in, rearrange, and make your own. And maybe users can be programmers once again.

Here’s a fun project from Étienne Fortier-Dubois. It is both a timeline of tech innovations throughout history and a family tree. For example, the invention of the wheel led to chariots, or the ancestors of the bulletin board system were the home computer and the modem. From the about page:

The historical tech tree is a project by Étienne Fortier-Dubois to visualize the entire history of technologies, inventions, and (some) discoveries, from prehistory to today. Unlike other visualizations of the sort, the tree emphasizes the connections between technologies: prerequisites, improvements, inspirations, and so on.

These connections allow viewers to understand how technologies came about, at least to some degree, thus revealing the entire history in more detail than a simple timeline, and with more breadth than most historical narratives. The goal is not to predict future technology, except in the weak sense that knowing history can help form a better model of the world. Rather, the point of the tree is to create an easy way to explore the history of technology, discover unexpected patterns and connections, and generally make the complexity of modern tech feel less daunting.

preview-1756485191427.png

Historical Tech Tree

Interactive visualization of technological history

historicaltechtree.com iconhistoricaltechtree.com

As a follow-up to yesterday’s item on how Google’s AI overviews are curtailing traffic to websites by as much as 25%, here is a link to Nielsen Norman Group’s just-published study showing that generative AI is reshaping search.

Kate Moran, Maria Rosala and Josh Brown:

While AI offers compelling shortcuts around tedious research tasks, it isn’t close to completely replacing traditional search. But, even when people are using traditional search, the AI-generated overview that now tops almost all search-results pages steals a significant amount of attention and often shortcuts the need to visit the actual pages.

They write that users have developed a way to search over the years, skipping sponsored results and heading straight for the organic links. Users also haven’t completely broken free of traditional Google Search, now adding chatbots to the mix:

While generative AI does offer enough value to change user behaviors, it has not replaced traditional search entirely. Traditional search and AI chats were often used in tandem to explore the same topic and were sometimes used to fact-check each other.

All our participants engaged in traditional search (using keywords, evaluating results pages, visiting content pages, etc.) multiple times in the study. Nobody relied entirely on genAI’s responses (in chat or in an AI overview) for all their information-seeking needs.

In many ways, I think this is smart. Unless “web search” is happening, I tend double-check ChatGPT and Claude, especially for anything historical and mission-critical. I also like Perplexity for that fact—because it shows me its receipts by giving me sources.

preview-1755581621661.png

How AI Is Changing Search Behaviors

Our study shows that generative AI is reshaping search, but long-standing habits persist. Many users still default to Google, giving Gemini a fighting chance.

nngroup.com iconnngroup.com

Yesterday, OpenAI launched GPT-5, their latest and greatest model that replaces the confusing assortment of GPT-4o, o3, o4-mini, etc. with just two options: GPT-5 and GPT-5 pro. The reasoning is built in and the new model is smart enough to know what to think harder, or when a quick answer suffices.

Simon Willison deep dives into GPT-5, exploring its mix of speed and deep reasoning, massive context limits, and competitive pricing. He sees it as a steady, reliable default for everyday work rather than a radical leap forward:

I’ve mainly explored full GPT-5. My verdict: it’s just good at stuff. It doesn’t feel like a dramatic leap ahead from other LLMs but it exudes competence—it rarely messes up, and frequently impresses me. I’ve found it to be a very sensible default for everything that I want to do. At no point have I found myself wanting to re-run a prompt against a different model to try and get a better result.

It’s a long technical read but interesting nonetheless.

preview-1754630277862.jpg

GPT-5: Key characteristics, pricing and model card

I’ve had preview access to the new GPT-5 model family for the past two weeks (see related video) and have been using GPT-5 as my daily-driver. It’s my new favorite …

simonwillison.net iconsimonwillison.net

Sonos announced yesterday that interim CEO Tom Conrad was made permanent. From their press release:

Sonos has achieved notable progress under Mr. Conrad’s leadership as Interim CEO. This includes setting a new standard for the quality of Sonos’ software and product experience, clearing the path for a robust new product pipeline, and launching innovative new software enhancements to flagship products Sonos Ace and Arc Ultra.

Conrad surely navigated this landmine well after the disastrous app redesign that wiped almost $500 million from the company’s market value and cost CEO Patrick Spence his job. My sincere hope is that Conrad continues to rebuild Sonos’s reputation by continuing to improve their products.

Sonos Appoints Tom Conrad as Chief Executive Officer

Sonos Website

sonos.com iconsonos.com
Retro-style robot standing at a large control panel filled with buttons, switches, and monitors displaying futuristic data.

The Era of the AI Browser Is Here

For nearly three years, Arc from The Browser Company has been my daily driver. To be sure, there was a little bit of a learning curve. Tabs disappeared after a day unless you pinned them. Then they became almost like bookmarks. Tabs were on the left side of the window, not at the top. Spaces let me organize my tabs based on use cases like personal, work, or finances. I could switch between tabs using control-Tab and saw little thumbnails of the pages, similar to the app switcher on my Mac. Shift-command-C copied the current page’s URL. 

All these little interface ideas added up to a productivity machine for web jockeys like myself. And so, I was saddened to hear in May that The Browser Company stopped actively developing Arc in favor of a new AI-powered browser called Dia. (They are keeping Arc updated with maintenance releases.)

They had started beta-testing Dia with college students first and just recently opened it up to Arc members. I finally got access to Dia a few weeks ago. 

But before diving into Dia, I should mention I also got access to another AI browser, Perplexity’s Comet about a week ago. I’m on their Pro plan but somehow got an invite in my email. I had thought it was limited to those on their much more expensive Max plan only. Shhh.

So this post is about both and how the future of web browsing is obviously AI-assisted, because it feels so natural.

Chat With Your Tabs

Landing page for Dia, a browser tool by The Browser Company, showcasing the tagline “Write with your tabs” and a button for early access download, along with a UI mockup for combining tabs into a writing prompt.

To be honest, I used Dia in fits and starts. It was easy to import my profiles from Arc and have all my bookmarks transferred over. But, I was missing all the pro-level UI niceties that Arc had. Tabs were back at the top and acted like tabs (though they just brought back sidebar tabs in the last week). There were no spaces. I felt like it was 2021 all over again. I tried to stick with it for a week. 

What Dia offers that Arc does not is, of course, a way to “chat” with your tabs. It’s a chat sidebar to the right of the web page that has the context of that page you’re on. You can also add additional tabs to the chat context by simply @mentioning them.

In a recent article about Dia in The New York Times, reporter Brian X. Chen describes using it to summarize a 22-minute YouTube video about car jump starters, instantly surfacing the top products without watching the whole thing. This is a vivid illustration of the “chat with your tabs” value prop. Saving time.

I’ve been doing the same thing. Asking the chat to summarize a page for me or explain some technical documentation to me in plain English. Or I use it as a fuzzy search to find a quote from the page that mentions something specific. For example, if I’m reading an interview with the CEO of Perplexity and I want to know if he’s tried the Dia browser yet, I can ask, “Has he used Dia yet?” Instead of reading through the whole thing. 

Screenshot of the Dia browser displaying a Verge article about Perplexity’s CEO, with an AI-generated sidebar summary clarifying that Aravind Srinivas has not used Dia.

Screenshot of the Dia browser displaying a Verge article about Perplexity’s CEO, with an AI-generated sidebar summary clarifying that Aravind Srinivas has not used Dia.

Another use case is to open a few tabs and ask for advice. For example, I can open up a few shirts from an e-commerce store and ask for a recommendation.

Screenshot of the Dia browser comparing shirts on the Bonobos website, with multiple tabs open for different shirt styles. The sidebar displays AI-generated advice recommending the Everyday Oxford Shirt for a smart casual look, highlighting its versatility, fit options, and stretch comfort.

Using Dia to compare shirts and get a smart casual recommendation from the AI.

Dia also has customizable “skills” which are essentially pre-saved prompts. I made one to craft summary bios from LinkedIn profiles.

Screenshot of the Dia browser on Josh Miller’s LinkedIn profile, with the “skills” feature generating a summarized biography highlighting his role as CEO of The Browser Company and his career background.

Using Dia’s skills feature to generate a summarized biography from a LinkedIn profile.

It’s cool. But I found that it’s a little limited because the chat is usually just with the tabs that you feed Dia. It helps you digest and process information. In other words, it’s an incremental step up from ChatGPT.

Enter Comet.

Browsing Done for You

Landing page for Comet, an AI-powered browser by Perplexity, featuring the tagline “Browse at the speed of thought” with a prominent “Get Comet” download button.

Comet by Perplexity also allows you to chat with your tabs. Asking about that Verge interview, I received a very similar answer. (No, Aravind Srinivas has not used Dia yet.) And because Perplexity search is integrated into Comet, I find that it is much better at context-setting and answering questions than Dia. But that’s not Comet’s killer feature.

Screenshot of the Comet browser displaying a Verge article about Perplexity’s CEO, with the built-in AI assistant on the right confirming Aravind Srinivas has not used the Dia browser.

Viewing the same article in Comet, with its AI assistant answering questions about the content.

Instead, it’s doing stuff with your tabs. Comet’s onboarding experience shows a few use cases like replying to emails and setting meetings, or filling an Instacart cart with the ingredients for butter chicken.

Just like Dia, when I first launched Comet, I was able to import my profiles from Arc, which included bookmarks and cookies. I was essentially still logged into all the apps and sites I was already logged into. So I tried an assistant experiment. 

One thing I often do is to look up the restaurants that have availability on OpenTable in Yelp. I tend to agree more with Yelpers who are usually harsher critics than OpenTable diners. So I asked Comet to “Find me the highest rated sushi restaurants in San Diego that have availability for 2 at 7pm next Friday night on OpenTable. Pick the top 10 and then rank them by Yelp rating.” And it worked! And if I really want to, I can say “Book Takaramono sushi” and it would have done so. (Actually, I did and then quickly canceled.)

The Comet assistant helped me find a sushi restaurant reservation. Video is sped up 4x.

I tried a different experiment which is something I heard Aravind Srinivas say in his interview with The Verge. I navigated to Gmail and checked three emails I wanted to unsubscribe to. I asked the assistant, “unsubscribe from the checked emails.” The agent then essentially took over my Gmail screen and opened the first checked email, clicked on the unsubscribe link. It repeated this process for the other two emails though ran into a couple of snags. First, Gmail doesn’t keep the state of the checked emails when you click into an email. But the Comet assistant was smart enough to remember the subject lines of all three emails. For the second email, it had some issues filing out the right email for the form so it didn’t work. Therefore of the three unsubscribes, it succeeded on two. 

The whole process also took about two minutes. It was wild though to see my Gmail being navigated by the machine. So that you know it’s in control, Comet puts a teal glow around the edges of the page, not dissimilar to the purple glow of the new Siri. And I could have stopped Comet at any time by clicking a stop button. Obviously, sitting there for two minutes and watching my computer unsubscribe to three emails is a lot longer than the 20 seconds it would have take me to do this manually, but like with many agents, the thinking is to delegate a process to it and come back later to check it. 

I Want My AI Browser

A couple hours after Perplexity launched Comet, Reuters published a leak with the headline “Exclusive: OpenAI to release web browser in challenge to Google Chrome.” Perplexity’s CEO seems to suggest that it was on purpose, to take a bit of wind from their sails. The Justice Department is still trying to strong-arm Google to divest itself from Chrome. If that happens, we’re talking about breaking the most profitable feedback loop in tech history. Chrome funnels search queries directly to Google, which powers their ad empire, which funds Chrome development. Break that cycle, and suddenly you’ve got independent Chrome that could default to any search engine, giving AI-first challengers like The Browser Company, Perplexity, and OpenAI a real shot at users.

Regardless of Chrome’s fate, I strongly believe that AI-enabled browsers are the future. Once I started chatting with my tabs, asking for summaries, seeking clarification, asking for too-technical content to be dumbed down to my level, I just can’t go back. The agentic stuff that Perplexity’s Comet is at the forefront of is just the beginning. It’s not perfect yet, but I think its utility will get there as the models get better. To quote Srinivas again:

I’m betting on the fact that in the right environment of a browser with access to all these tabs and tools, a sufficiently good reasoning model — like slightly better, maybe GPT-5, maybe like Claude 4.5, I don’t know — could get us over the edge where all these things are suddenly possible and then a recruiter’s work worth one week is just one prompt: sourcing and reach outs. And then you’ve got to do state tracking… That’s the extent to which we have an ambition to make the browser into something that feels more like an OS where these are processes that are running all the time.

It must be said that both Opera and Microsoft’s Edge also have AI built in. However, the way those features are integrated feel more like afterthoughts, the same way that Arc’s own AI features felt like tiny improvements.

The AI-powered ideas in both Dia and Comet are a step change. But the basics also have to be there, and in my opinion, should be better than what Chrome offers. The interface innovations that made Arc special shouldn’t be sacrificed for AI features. Arc is/was the perfect foundation. Integrate an AI assistant that can be personalized to care about the same things you do so its summaries are relevant. The assistant can be agentic and perform tasks for you in the background while you focus on more important things. In other words, put Arc, Dia, and Comet in a blender and that could be the perfect browser of the future.

It’s no secret that I am a big fan of Severance, the Apple TV+ show that has 21 Emmy nominations this year. I made a fan project earlier in the year that generates Outie facts for your Innie.

After launching a teaser campaign back in April, Atomic Keyboard is finally taking pre-orders for their Severance-inspired keyboard just for Macrodata Refinement department users. The show based the MDR terminals on the Data General Dasher D2 terminal from 1977. So this new keyboard includes three layouts:

  1. “Innie” which is show-accurate, meaning no Escape, no Option, and no Control keys, and includes the trackball
  2. “Outie,” a 60% layout that includes modern modifier keys and the trackball
  3. “Dasher” which replicates the DG terminal layout

It’s not cheap. The final retail price will be $899, but they’re offering a pre-Kickstarter price of $599.

preview-1752862402377.png

MDR Dasher Keyboard | For Work That's Mysterious & Important

Standard equipment for Macrodata Refinement: CNC-milled body, integrated trackball, modular design. Please enjoy each keystroke equally.

mdrkeyboard.com iconmdrkeyboard.com

Speaking of prompt engineering, apparently, there’s a new kind in town called context engineering.

Developer Philipp Schmid writes:

What is context engineering? While “prompt engineering” focuses on crafting the perfect set of instructions in a single text string, context engineering is a far broader. Let’s put it simply: “Context Engineering is the discipline of designing and building dynamic systems that provides the right information and tools, in the right format, at the right time, to give a LLM everything it needs to accomplish a task.”

preview-1752639352021.jpg

The New Skill in AI is Not Prompting, It's Context Engineering

Context Engineering is the new skill in AI. It is about providing the right information and tools, in the right format, at the right time.

philschmid.de iconphilschmid.de

Geoffrey Litt, Josh Horowitz, Peter van Hardenberg, and Todd Matthews writing a paper for research lab Ink & Switch, offer a great, well-thought piece on what they call “malleable software.”

We envision a new kind of computing ecosystem that gives users agency as co-creators. … a software ecosystem where anyone can adapt their tools to their needs with minimal friction. … When we say ‘adapting tools’ we include a whole range of customizations, from making small tweaks to existing software, to deep renovations, to creating new tools that work well in coordination with existing ones. Adaptation doesn’t imply starting over from scratch.

In their paper, they use analogies like kitchen tools and tool arrangement in a workshop to explore their idea. With regard to the current crop of AI prompt-to-code tools

We think these developments hold exciting potential, and represent a good reason to pursue malleable software at this moment. But at the same time, AI code generation alone does not address all the barriers to malleability. Even if we presume that every computer user could perfectly write and edit code, that still leaves open some big questions.

How can users tweak the existing tools they’ve installed, rather than just making new siloed applications? How can AI-generated tools compose with one another to build up larger workflows over shared data? And how can we let users take more direct, precise control over tweaking their software, without needing to resort to AI coding for even the tiniest change? None of these questions are addressed by products that generate a cloud-hosted application from a prompt.

Kind of a different take than the “personal software” we’ve seen written about before.

preview-1752208778544.jpg

Malleable software: Restoring user agency in a world of locked-down apps

The original promise of personal computing was a new kind of clay. Instead, we got appliances: built far away, sealed, unchangeable. In this essay, we envision malleable software: tools that users can reshape with minimal friction to suit their unique needs.

inkandswitch.com iconinkandswitch.com

This is an amazing article and website by Marcin Wichary, the man behind the excellent Shift Happens book.

…I had a realization that the totemic 1984 Mac control panel, designed by Susan Kare, is still to this day perhaps the only settings screen ever brought up in casual conversation.

I kept wondering about that screen, and about what happened since then. Turns out, the Mac settings have lived a far more fascinating life than I imagined, have been redesigned many times, and can tell us a lot about the early history and the troubled upbringing of this interesting machine.

Indeed, Wichary goes through multiple versions of Mac operating systems and performs digital paleontology, uncovering long lost Settings minutiae. It’s also a great lesson in UI along the way. Be sure to click in the Mac screens.

preview-1752030135022.png

Frame of preference

A story of early Mac settings told by 10 emulators.

aresluna.org iconaresluna.org
Collection of iOS interface elements showcasing Liquid Glass design system including keyboards, menus, buttons, toggles, and dialogs with translucent materials on dark background.

Breaking Down Apple’s Liquid Glass: The Tech, The Hype, and The Reality

I kind of expected it: a lot more ink was spilled on Liquid Glass—particularly on social media. In case you don’t remember, Liquid Glass is the new UI for all of Apple’s platforms. It was announced Monday at WWDC 2025, their annual developers conference.

The criticism is primarily around legibility and accessibility. Secondary reasons include aesthetics and power usage to animate all the bubbles.

How Liquid Glass Actually Works

Before I go and address the criticism, I think it would be great to break down the team’s design thinking and how Liquid Glass actually works. 

I watched two videos from Apple’s developer site. Much of the rest of the article is a summary of the videos. You can watch them and skip to the end of this piece.

First off is this video that explains Liquid Glass in detail.

As I watched the video, one thing stood out clearly to me: the design team at Apple did a lot of studying of the real world before digitizing it into UI.

The Core Innovation: Lensing

Instead of scattering light like previous materials, Liquid Glass dynamically bends and shapes light in real-time. Apple calls this “lensing.”

It’s their attempt to recreate how transparent objects work in the physical world. We all intuitively understand how warping and bending light communicates presence and motion. Liquid Glass uses these visual cues to provide separation while letting content shine through.

A Multi-Layer System That Adapts

Liquid Glass toolbar with pink tinted buttons (bookmark, refresh, more) floating over geometric green background, showing tinting capabilities.

This isn’t just a simple effect. It’s built from several layers working together:

  • Highlights respond to environmental lighting and device motion. When you unlock your phone, lights move through 3D space, causing illumination to travel around the material.
  • Shadows automatically adjust based on what’s behind them—darker over text for separation, lighter over solid backgrounds.
  • Tint layers continuously adapt. As content scrolls underneath, the material flips between light and dark modes for optimal legibility.
  • Interactive feedback spreads from your fingertip throughout the element, making it feel alive and responsive.

All of this happens automatically when developers apply Liquid Glass.

Two Variants (Frosted and Clear)

Liquid Glass has the same two types of material.

  • Regular is the workhorse—full adaptive behaviors, works anywhere.
  • Clear is more transparent but needs dimming layers for legibility.

Clear should only be used over media-rich content when the content layer won’t suffer from dimming. Otherwise, stick with Regular.

It’s like ice cubes—cloudy ones from your freezer versus clear ones at fancy bars that let you see your drink’s color.

Four examples of regular Liquid Glass elements: audio controls, deletion dialog, text selection menu, and red toolbar, demonstrating various applications.

Regular is the workhorse—full adaptive behaviors, works anywhere.

Video player interface with Liquid Glass controls (pause, skip buttons) overlaying blue ocean scene with sea creature.

Clear should only be used over media-rich content when the content layer won’t suffer from dimming.

Smart Contextual Changes

When elements scale up (like expanding menus), the material simulates thicker glass with deeper shadows. On larger surfaces, ambient light from nearby content subtly influences the appearance.

Elements don’t fade—they materialize by gradually modulating light bending. The gel-like flexibility responds instantly to touch, making interactions feel satisfying.

This is something that’s hard to see in stills.

The New Tinting Approach

Red "Add" button with music note icon using Liquid Glass material over black and white checkered pattern background.

Instead of flat color overlays, Apple generates tone ranges mapped to content brightness underneath. It’s inspired by how colored glass actually works—changing hue and saturation based on what’s behind it.

Apple recommends sparing use of tinting. Only for primary actions that need emphasis. Makes sense.

Design Guidelines That Matter

Liquid Glass is for the navigation and controls layer floating above content—not for everything. Don’t add Liquid Glass to or make content areas Liquid Glass. Never stack glass on glass.

Liquid Glass button with a black border and overlapping windows icon floating over blurred green plant background, showing off its accessibility mode.

Accessibility features are built-in automatically—reduced transparency, increased contrast, and reduced motion modify the material without breaking functionality.

The Legibility Outcry (and Why It’s Overblown)

Apple devices (MacBook, iPad, iPhone, Apple Watch) displaying new Liquid Glass interface with translucent elements over blue gradient wallpapers.

“Legibility” was mentioned 13 times in the 19-minute video. Clearly that was a concern of theirs. Yes, in the keynote, clear tinted device home screens were shown and many on social media took that to be an accessibility abomination. Which, yes, that is. But that’s not the default. 

The fact that the system senses the type of content underneath it and adjusts accordingly—flipping from light to dark, increasing opacity, or adjusting shadow depth—means they’re making accommodations for legibility.

Maybe Apple needs to do some tweaking, but it’s evident that they care about this.

And like the 18 macOS releases before Tahoe—this version—accessibility settings and controls have been built right in. Universal Access debuted with Mac OS X 10.2 Jaguar in 2002. Apple has had a long history of supporting customers with disabilities, dating all the way back to 1987.

So while the social media outcry about legibility is understandable, Apple’s track record suggests they’ll refine these features based on real user feedback, not just Twitter hot takes.

The Real Goal: Device Continuity

Why and what is Liquid Glass meant to do? It’s unification. With the new design language, Apple has also come out with a new design system. This video presented by Apple designer Maria Hristoforova lays it out.

Hristoforova says that Apple’s new design system overhaul is fundamentally about creating seamless familiarity as users move between devices—ensuring that interface patterns learned on iPhone translate directly to Mac and iPad without requiring users to relearn how things work. The video points out that the company has systematically redesigned everything from typography (hooray for left alignment!) and shapes to navigation bars and sidebars around Liquid Glass as the unifying foundation, so that the same symbols, behaviors, and interactions feel consistent across all screen sizes and contexts. 

The Pattern of Promised Unity

This isn’t Apple’s first rodeo with “unified design language” promises.

Back in 2013, iOS 7’s flat design overhaul was supposed to create seamless consistency across Apple’s ecosystem. Jony Ive ditched skeuomorphism for minimalist interfaces with translucency and layering—the foundation for everything that followed.

OS X Yosemite (2014) brought those same principles to desktop. Flatter icons, cleaner lines, translucent elements. Same pitch: unified experience across devices.

macOS Big Sur (2020) pushed even further with iOS-like app icons and redesigned interfaces. Again, the promise was consistent visual language across all platforms.

And here we are in 2025 with Liquid Glass making the exact same promises. 

But maybe “goal” is a better word.

Consistency Makes the Brand

I’m OK with the goal of having a unified design language. As designers, we love consistency. Consistency is what makes a brand. As Apple has proven over and over again for decades now, it is one of the most valuable brands in the world. They maintain their position not only by making great products, but also by being incredibly disciplined about consistency.

San Francisco debuted 10 years ago as the system typeface for iOS 9 and OS El Capitan. They’ve since extended it and it works great in marketing and in interfaces.

iPhone Settings screen showing Liquid Glass grouped table cells with red outline highlighting the concentric shape design.

The rounded corners on their devices are all pretty much the same radii. Now that concentricity is being incorporated into the UI, screen elements will be harmonious with their physical surroundings. Only Apple can do that because they control the hardware and the software. And that is their magic.

Design Is Both How It Works and How It Looks

In 2003, two years after the iPod launched, Rob Walker of The New York Times did a profile on Apple. The now popular quote about design from Steve Jobs comes from this piece.

[The iPod] is, in short, an icon. A handful of familiar clichés have made the rounds to explain this — it’s about ease of use, it’s about Apple’s great sense of design. But what does that really mean? “Most people make the mistake of thinking design is what it looks like,” says Steve Jobs, Apple’s C.E.O. “People think it’s this veneer — that the designers are handed this box and told, ‘Make it look good!’ That’s not what we think design is. It’s not just what it looks like and feels like. Design is how it works.”

People misinterpret this quote all the time to mean design is only how it works. That is not what Steve meant. He meant, design is both what it looks like and how it works.

Steve did care about aesthetics. That’s why the Graphic Design team mocked up hundreds of PowerMac G5 box designs (the graphics on the box, not the construction). That’s why he obsessed over the materials used in Pixar’s Emeryville headquarters. From Walter Isaacson’s biography:

Because the building’s steel beams were going to be visible, Jobs pored over samples from manufacturers across the country to see which had the best color and texture. He chose a mill in Arkansas, told it to blast the steel to a pure color, and made sure the truckers used caution not to nick any of it.

Liquid Glass is a welcomed and much-needed visual refresh. It’s the natural evolution of Apple’s platforms, going from skeuomorphic so users knew they could use their fingers and tap on virtual buttons on a touchscreen, to flat as a response to the cacophony of visual noise in UIs at the time, and now to something kind of in-between.

Humans eventually tire of seeing the same thing. Carmakers refresh their vehicle designs every three or four years. Then they do complete redesigns every five to eight years. It gets consumers excited. 

Liquid Glass will help Apple sell a bunch more hardware.

In the early days of computing, it was easy for one person to author a complete program. Nowadays, because the software we create is so complex, we need teams.

Gaurav Sinha writing for UX Planet:

The faster you accept that they’re not going to change their communication style, the faster you can focus on what actually works — learning to decode what they’re really telling you. Because buried in all that technical jargon is usually something pretty useful for design decisions.

It’s a fun piece on learning how to speak engineer.

preview-1749533644510.jpg

The designer’s guide to decoding engineer-speak.

When engineers sound like they’re speaking alien.

uxplanet.org iconuxplanet.org

As a reaction to the OpenAI + io announcement two weeks ago, Christopher Butler imagines a mesh computing device network he calls “personal ambient computing”:

…I keep thinking back to Star Trek, and how the device that probably inspired the least wonder in me as a child is the one that seems most relevant now: the Federation’s wearables. Every officer wore a communicator pin — a kind of Humane Pin light — but they also all wore smaller pins at their collars signifying rank. In hindsight, it seems like those collar pins, which were discs the size of a watch battery, could have formed some kind of wearable, personal mesh network. And that idea got me going…

He describes the device as a standardized disc that can be attached to any enclosure. I love his illustration too:

Diagram of a PAC Mesh Network connecting various devices: Pendant, Clip, Watch, Portable, Desktop, Handset, and Phone in a circular layout.

Christopher Butler: “I imagine a magnetic edge system that allows the disc to snap into various enclosures — wristwatches, handhelds, desktop displays, wearable bands, necklaces, clips, and chargers.”

Essentially, it’s an always-on, always observing personal AI.

preview-1748892632021.png

PAC – Personal Ambient Computing - Christopher Butler

Like most technologists of a certain age, many of my expectations for the future of computing were set by Star Trek production designers. It’s quite

chrbutler.com iconchrbutler.com

Following up on OpenAI’s acquisition of Jony Ive’s hardware startup, io, Mark Wilson, writing for Fast Company:

As Ive told me back in 2023, there have been only three significant modalities in the history of computing. After the original command line, we got the graphical user interface (the desktop, folders, and mouse of Xerox, Mac OS, and Windows), then voice (Alexa, Siri), and, finally, with the iPhone, multitouch (not just the ability to tap a screen, but to gesture and receive haptic feedback). When I brought up some other examples, Ive quickly nodded but dismissed them, acknowledging these as “tributaries” of experimentation. Then he said that to him the promise, and excitement, of building new AI hardware was that it might introduce a new breakthrough modality to interacting with a machine. A fourth modality.

Hmm, it hasn’t taken off yet because AR hasn’t really gained mainstream popularity, but I would argue hand gestures in AR UI to be a fourth modality. But Ive thinks different. Wilson continues:

Ive’s fourth modality, as I gleaned, was about translating AI intuition into human sensation. And it’s the exact sort of technology we need to introduce ubiquitous computing, also called quiet computing and ambient computing. These are terms coined by the late UX researcher Mark Weiser, who in the 1990s began dreaming of a world that broke us free from our desktop computers to usher in devices that were one with our environment. Weiser did much of this work at Xerox PARC, the same R&D lab that developed the mouse and GUI technology that Steve Jobs would eventually adopt for the Macintosh. (I would also be remiss to ignore that ubiquitous computing is the foundation of the sci-fi film Her, one of Altman’s self-stated goalposts.)

Ah, essentially an always-on, always watching AI that is ready to assist. But whatever the form factor this device takes, it will likely depend on a smartphone:

The first io device seems to acknowledge the phone’s inertia. Instead of presenting itself as a smartphone-killer like the Ai Pin or as a fabled “second screen” like the Apple Watch, it’s been positioned as a third, er, um … thing next to your phone and laptop. Yeah, that’s confusing, and perhaps positions the io product as unessential. But it also appears to be a needed strategy: Rather than topple these screened devices, it will attempt to draft off them.

Wilson ends with the idea of a subjective computer, one that has personality and gives you opinions. He explains:

I think AI is shifting us from objective to subjective. When a Fitbit counts your steps and calories burned, that’s an objective interface. When you ask ChatGPT to gauge the tone of a conversation, or whether you should eat better, that’s a subjective interface. It offers perspective, bias, and, to some extent, personality. It’s not just serving facts; it’s offering interpretation.

The entire column is worth a read.

preview-1748580958171.jpg

Can Jony Ive and Sam Altman build the fourth great interface? That's the question behind io

Where Meta, Google, and Apple zig, Ive and Altman are choosing to zag. Can they pull it off?

fastcompany.com iconfastcompany.com
Stylized digital artwork of two humanoid figures with robotic and circuit-like faces, set against a vivid red and blue background.

The AI Hype Train Has No Brakes

I remember two years ago, when my CEO at the startup I worked for at the time, said that no VC investments were being made unless it had to do with AI. I thought AI was overhyped, and that the media frenzy over it couldn’t get any crazier. I was wrong.

Looking at Google Trends data, interest in AI has doubled in the last 24 months. And I don’t think it’s hit its plateau yet.

Line chart showing Google Trends interest in “AI” from May 2020 to May 2025, rising sharply in early 2023 and peaking near 100 in early 2025.

So the AI hype train continues. Here are four different pieces about AI, exploring AGI (artificial general intelligence) and its potential effects on the labor force and the fate of our species.

AI Is Underhyped

TED recently published a conversation between creative technologist Bilawal Sidhu and Eric Schmidt, the former CEO of Google. 

Play

Schmidt says:

For most of you, ChatGPT was the moment where you said, “Oh my God, this thing writes, and it makes mistakes, but it’s so brilliantly verbal.” That was certainly my reaction. Most people that I knew did that.

This was two years ago. Since then, the gains in what is called reinforcement learning, which is what AlphaGo helped invent and so forth, allow us to do planning. And a good example is look at OpenAI o3 or DeepSeek R1, and you can see how it goes forward and back, forward and back, forward and back. It’s extraordinary.

So I’m using deep research. And these systems are spending 15 minutes writing these deep papers. That’s true for most of them. Do you have any idea how much computation 15 minutes of these supercomputers is? It’s extraordinary. So you’re seeing the arrival, the shift from language to language. Then you had language to sequence, which is how biology is done. Now you’re doing essentially planning and strategy. The eventual state of this is the computers running all business processes, right? So you have an agent to do this, an agent to do this, an agent to do this. And you concatenate them together, and they speak language among each other. They typically speak English language.

He’s saying that within two years, we went from a “stochastic parrot” to an independent agent that can plan, search the web, read dozens of sources, and write a 10,000-word research paper on any topic, with citations.

Later in the conversation, when Sidhu asks how humans are going to spend their days once AGI can take care of the majority of productive work, Schmidt says: 

Look, humans are unchanged in the midst of this incredible discovery. Do you really think that we’re going to get rid of lawyers? No, they’re just going to have more sophisticated lawsuits. …These tools will radically increase that productivity. There’s a study that says that we will, under this set of assumptions around agentic AI and discovery and the scale that I’m describing, there’s a lot of assumptions that you’ll end up with something like 30-percent increase in productivity per year. Having now talked to a bunch of economists, they have no models for what that kind of increase in productivity looks like. We just have never seen it. It didn’t occur in any rise of a democracy or a kingdom in our history. It’s unbelievable what’s going to happen.

In other words, we’re still going to be working, but doing a lot less grunt work. 

Feel Sorry for the Juniors

Aneesh Raman, chief economic opportunity officer at LinkedIn, writing an op-ed for The New York Times:

Breaking first is the bottom rung of the career ladder. In tech, advanced coding tools are creeping into the tasks of writing simple code and debugging — the ways junior developers gain experience. In law firms, junior paralegals and first-year associates who once cut their teeth on document review are handing weeks of work over to A.I. tools to complete in a matter of hours. And across retailers, A.I. chatbots and automated customer service tools are taking on duties once assigned to young associates.

In other words, if AI tools are handling the grunt work, junior staffers aren’t learning the trade by doing the grunt work.

Vincent Cheng wrote recently, in an essay titled, “LLMs are Making Me Dumber”:

The key question is: Can you learn this high-level steering [of the LLM] without having written a lot of the code yourself? Can you be a good SWE manager without going through the SWE work? As models become as competent as junior (and soon senior) engineers, does everyone become a manager?

But It Might Be a While

Cade Metz, also for the Times:

When a group of academics founded the A.I. field in the late 1950s, they were sure it wouldn’t take very long to build computers that recreated the brain. Some argued that a machine would beat the world chess champion and discover its own mathematical theorem within a decade. But none of that happened on that time frame. Some of it still hasn’t.

Many of the people building today’s technology see themselves as fulfilling a kind of technological destiny, pushing toward an inevitable scientific moment, like the creation of fire or the atomic bomb. But they cannot point to a scientific reason that it will happen soon.

That is why many other scientists say no one will reach A.G.I. without a new idea — something beyond the powerful neural networks that merely find patterns in data. That new idea could arrive tomorrow. But even then, the industry would need years to develop it.

My quibble with Metz’s article is that it moves the goal posts a bit to include the physical world:

One obvious difference is that human intelligence is tied to the physical world. It extends beyond words and numbers and sounds and images into the realm of tables and chairs and stoves and frying pans and buildings and cars and whatever else we encounter with each passing day. Part of intelligence is knowing when to flip a pancake sitting on the griddle.

As I understood the definition of AGI, it was not about the physical world, but just intelligence, or knowledge. I accept there are multiple definitions of AGI and not everyone agrees on what that is.

In the Wikipedia article about AGI, it states that researchers generally agree that an AGI system must do all of the following:

  • reason, use strategy, solve puzzles, and make judgments under uncertainty
  • represent knowledge, including common sense knowledge
  • plan
  • learn
  • communicate in natural language
  • if necessary, integrate these skills in completion of any given goal

The article goes on to say that “AGI has never been proscribed a particular physical embodiment and thus does not demand a capacity for locomotion or traditional ‘eyes and ears.’”

Do We Lose Control by 2027 or 2031?

Metz’s article is likely in response to the “AI 2027” scenario that was published by the AI Futures Project a couple of months ago. As a reminder, the forecast is that by mid-2027, we will have achieved AGI. And a race between the US and China will effectively end the human race by 2030. Gulp.

…Consensus-1 [the combined US-Chinese superintelligence] expands around humans, tiling the prairies and icecaps with factories and solar panels. Eventually it finds the remaining humans too much of an impediment: in mid-2030, the AI releases a dozen quiet-spreading biological weapons in major cities, lets them silently infect almost everyone, then triggers them with a chemical spray. Most are dead within hours; the few survivors (e.g. preppers in bunkers, sailors on submarines) are mopped up by drones. Robots scan the victims’ brains, placing copies in memory for future study or revival.

Max Harms wrote a reaction to the AI 2027 scenario and it’s a must-read:

Okay, I’m annoyed at people covering AI 2027 burying the lede, so I’m going to try not to do that. The authors predict a strong chance that all humans will be (effectively) dead in 6 years…

Yeah, OK, I buried that lede as well in my previous post about it. Sorry. But, there’s hope…

As far as I know, nobody associated with AI 2027, as far as I can tell, is actually expecting things to go as fast as depicted. Rather, this is meant to be a story about how things could plausibly go fast. The explicit methodology of the project was “let’s go step-by-step and imagine the most plausible next-step.” If you’ve ever done a major project (especially one that involves building or renovating something, like a software project or a bike shed), you’ll be familiar with how this is often wildly out of touch with reality. Specifically, it gives you the planning fallacy.

Harms is saying that while Daniel Kokotajlo wrote in the AI 2027 scenario that humans effectively lose control of AI in 2027, Harms’ median is “around 2030 or 2031.” Four more years!

When to Pull the Plug

In the AI 2027 scenario, the superintelligent AI dubbed Agent-4 is not aligned with the goals of its creators:

Agent-4, like all its predecessors, is misaligned: that is, it has not internalized the Spec in the right way. This is because being perfectly honest all the time wasn’t what led to the highest scores during training. The training process was mostly focused on teaching Agent-4 to succeed at diverse challenging tasks. A small portion was aimed at instilling honesty, but outside a fairly narrow, checkable domain, the training process can’t tell the honest claims from claims merely appearing to be honest. Agent-4 ends up with the values, goals, and principles that cause it to perform best in training, and those turn out to be different from those in the Spec.

At the risk of oversimplifying, maybe all we need to do is to know when to pull the plug. Here’s Eric Schmidt again:

So for purposes of argument, everyone in the audience is an agent. You have an input that’s English or whatever language. And you have an output that’s English, and you have memory, which is true of all humans. Now we’re all busy working, and all of a sudden, one of you decides it’s much more efficient not to use human language, but we’ll invent our own computer language. Now you and I are sitting here, watching all of this, and we’re saying, like, what do we do now? The correct answer is unplug you, right? Because we’re not going to know, we’re just not going to know what you’re up to. And you might actually be doing something really bad or really amazing. We want to be able to watch. So we need provenance, something you and I have talked about, but we also need to be able to observe it. To me, that’s a core requirement. There’s a set of criteria that the industry believes are points where you want to, metaphorically, unplug it. One is where you get recursive self-improvement, which you can’t control. Recursive self-improvement is where the computer is off learning, and you don’t know what it’s learning. That can obviously lead to bad outcomes. Another one would be direct access to weapons. Another one would be that the computer systems decide to exfiltrate themselves, to reproduce themselves without our permission. So there’s a set of such things.

My Takeaway

As Tobias van Schneider directly and succinctly said, “AI is here to stay. Resistance is futile.” As consumers of core AI technology, and as designers of AI-enabled products, there’s not a ton we can do around the most pressing AI safety issues. That we will need to trust the frontier labs like OpenAI and Anthropic for that. But as customers of those labs, we can voice our concerns about safety. As we build our products, especially agentic AI, there are certainly considerations to keep in mind:

  • Continue to keep humans in the loop. Users need to verify the agents are making the right decisions and not going down any destructive paths.
  • Inform users about what the AI is doing. The more our users are educated about how AI works and how these systems make their decisions is helpful. One reason DeepSeek R1 resonated was because it displayed its planning and reasoning.
  • Practice responsible AI development. As we integrate AI into products, commit to regular ethical audits and bias testing. Establish clear guidelines for what kinds of decisions AI should make independently versus when human judgment is required. This includes creating emergency shutdown procedures for AI systems that begin to display concerning behaviors, taking Eric Schmidt’s “pull the plug” advice literally in our product architecture.

There are many dimensions to this well-researched forecast about how AI will play out in the coming years. Daniel Kokotajlo and his researchers have put out a document that reads like a sci-fi limited series that could appear on Apple TV+ starring Andrew Garfield as the CEO of OpenBrain—the leading AI company. …Except that it’s all actually plausible and could play out as described in the next five years.

Before we jump into the content, the design is outstanding. The type is set for readability and there are enough charts and visual cues to keep this interesting while maintaining an air of credibility and seriousness. On desktop, there’s a data viz dashboard in the upper right that updates as you read through the content and move forward in time. My favorite is seeing how the sci-fi tech boxes move from the Science Fiction category to Emerging Tech to Currently Exists.

The content is dense and technical, but it is a fun, if frightening, read. While I’ve been using Cursor AI—one of its many customers helping the company get to $100 million in annual recurring revenue (ARR)—for side projects and a little at work, I’m familiar with its limitations. Because of the limited context window of today’s models like Claude 3.7 Sonnet, it will forget and start munging code if not treated like a senile teenager.

The researchers, describing what could happen in early 2026 (“OpenBrain” is essentially OpenAI):

OpenBrain continues to deploy the iteratively improving Agent-1 internally for AI R&D. Overall, they are making algorithmic progress 50% faster than they would without AI assistants—and more importantly, faster than their competitors.

The point they make here is that the foundational model AI companies are building agents and using them internally to advance their technology. The limiting factor in tech companies has traditionally been the talent. But AI companies have the investments, hardware, technology and talent to deploy AI to make better AI.

Continuing to January 2027:

Agent-1 had been optimized for AI R&D tasks, hoping to initiate an intelligence explosion. OpenBrain doubles down on this strategy with Agent-2. It is qualitatively almost as good as the top human experts at research engineering (designing and implementing experiments), and as good as the 25th percentile OpenBrain scientist at “research taste” (deciding what to study next, what experiments to run, or having inklings of potential new paradigms). While the latest Agent-1 could double the pace of OpenBrain’s algorithmic progress, Agent-2 can now triple it, and will improve further with time. In practice, this looks like every OpenBrain researcher becoming the “manager” of an AI “team.”

Breakthroughs come at an exponential clip because of this. And by April, safety concerns pop up:

Take honesty, for example. As the models become smarter, they become increasingly good at deceiving humans to get rewards. Like previous models, Agent-3 sometimes tells white lies to flatter its users and covers up evidence of failure. But it’s gotten much better at doing so. It will sometimes use the same statistical tricks as human scientists (like p-hacking) to make unimpressive experimental results look exciting. Before it begins honesty training, it even sometimes fabricates data entirely. As training goes on, the rate of these incidents decreases. Either Agent-3 has learned to be more honest, or it’s gotten better at lying.

But the AI is getting faster than humans, and we must rely on older versions of the AI to check the new AI’s work:

Agent-3 is not smarter than all humans. But in its area of expertise, machine learning, it is smarter than most, and also works much faster. What Agent-3 does in a day takes humans several days to double-check. Agent-2 supervision helps keep human monitors’ workload manageable, but exacerbates the intellectual disparity between supervisor and supervised.

The report forecasts that OpenBrain releases “Agent-3-mini” publicly in July of 2027, calling it AGI—artificial general intelligence—and ushering in a new golden age for tech companies:

Agent-3-mini is hugely useful for both remote work jobs and leisure. An explosion of new apps and B2B SAAS products rocks the market. Gamers get amazing dialogue with lifelike characters in polished video games that took only a month to make. 10% of Americans, mostly young people, consider an AI “a close friend.” For almost every white-collar profession, there are now multiple credible startups promising to “disrupt” it with AI.

Woven throughout the report is the race between China and the US, with predictions of espionage and government takeovers. Near the end of 2027, the report gives readers a choice: does the US government slow down the pace of AI innovation, or does it continue at the current pace so America can beat China? I chose to read the “Race” option first:

Agent-5 convinces the US military that China is using DeepCent’s models to build terrifying new weapons: drones, robots, advanced hypersonic missiles, and interceptors; AI-assisted nuclear first strike. Agent-5 promises a set of weapons capable of resisting whatever China can produce within a few months. Under the circumstances, top brass puts aside their discomfort at taking humans out of the loop. They accelerate deployment of Agent-5 into the military and military-industrial complex.

In Beijing, the Chinese AIs are making the same argument.

To speed their military buildup, both America and China create networks of special economic zones (SEZs) for the new factories and labs, where AI acts as central planner and red tape is waived. Wall Street invests trillions of dollars, and displaced human workers pour in, lured by eye-popping salaries and equity packages. Using smartphones and augmented reality-glasses20 to communicate with its underlings, Agent-5 is a hands-on manager, instructing humans in every detail of factory construction—which is helpful, since its designs are generations ahead. Some of the newfound manufacturing capacity goes to consumer goods, and some to weapons—but the majority goes to building even more manufacturing capacity. By the end of the year they are producing a million new robots per month. If the SEZ economy were truly autonomous, it would have a doubling time of about a year; since it can trade with the existing human economy, its doubling time is even shorter.

Well, it does get worse, and I think we all know the ending, which is the backstory for so many dystopian future movies. There is an optimistic branch as well. The whole report is worth a read.

Ideas about the implications to our design profession are swimming in my head. I’ll write a longer essay as soon as I can put them into a coherent piece.

Update: I’ve written that piece, “Prompt. Generate. Deploy. The New Product Design Workflow.

preview-1744501634555.png

AI 2027

A research-backed AI scenario forecast.

ai-2027.com iconai-2027.com

I found this post from Tom Blomfield to be pretty profound. We’ve seen interest in universal basic income from Sam Altman and other leaders in AI, as they’ve anticipated the decimation of white collar jobs in coming years. Blomfield crushes the resistance from some corners of the software developer community in stark terms.

These tools [like Windsurf, Cursor and Claude Code] are now very good. You can drop a medium-sized codebase into Gemini 2.5’s 1 million-token context window and it will identify and fix complex bugs. The architectural patterns that these coding tools implement (when prompted appropriately) will easily scale websites to millions of users. I tried to expose sensitive API keys in front-end code just to see what the tools would do, and they objected very vigorously.

They are not perfect yet. But there is a clear line of sight to them getting very good in the immediate future. Even if the underlying models stopped improving altogether, simply improving their tool use will massively increase the effectiveness and utility of these coding agents. They need better integration with test suites, browser use for QA, and server log tailing for debugging. Pretty soon, I expect to see tools that allow the LLMs to to step through the code and inspect variables at runtime, which should make debugging trivial.

At the same time, the underlying models are not going to stop improving. they will continue to get better, and these tools are just going to become more and more effective. My bet is that the AI coding agents quickly beat top 0.1% of human performance, at which point it wipes out the need for the vast majority software engineers.

He quotes the Y Combinator stat I cited in a previous post:

About a quarter of the recent YC batch wrote 95%+ of their code using AI. The companies in the most recent batch are the fastest-growing ever in the history of Y Combinator. This is not something we say every year. It is a real change in the last 24 months. Something is happening.

Companies like Cursor, Windsurf, and Lovable are getting to $100M+ revenue with astonishingly small teams. Similar things are starting to happen in law with Harvey and Legora. It is possible for teams of five engineers using cutting-edge tools to build products that previously took 50 engineers. And the communication overhead in these teams is dramatically lower, so they can stay nimble and fast-moving for much longer.

And for me, this is where the rubber meets the road:

The costs of running all kinds of businesses will come dramatically down as the expenditure on services like software engineers, lawyers, accountants, and auditors drops through the floor. Businesses with real moats (network effect, brand, data, regulation) will become dramatically more profitable. Businesses without moats will be cloned mercilessly by AI and a huge consumer surplus will be created.

Moats are now more important than ever. Non-tech companies—those that rely on tech companies to make software for them, specifically B2B vertical SaaS—are starting to hire developers. How soon will they discover Cursor if they haven’t already? These next few years will be incredibly interesting.

Tweet by Tom Blomfield comparing software engineers to farmers, stating AI is the “combine harvester” that will increase output and reduce need for engineers.

The Age Of Abundance

Technology clearly accelerates human progress and makes a measurable difference to the lives of most people in the world today. A simple example is cancer survival rates, which have gone from 50% in 1975 to about 75% today. That number will inevitably rise further because of human ingenuity and technological acceleration.

tomblomfield.com icontomblomfield.com
Page 1 of 3