Journal tags: models




I’m sure you’ve heard the law of the instrument: when all you have is a hammer, everything looks like a nail.

There’s another side to it. If you’re selling hammers, you’ll depict a world full of nails.

Recent hammers include cryptobollocks and virtual reality. It wasn’t enough for blockchains and the metaverse to be potentially useful for some situations; they staked their reputations on being utterly transformative, disrupting absolutely every facet of life.

This kind of hype is a terrible strategy in the long-term. But if you can convince enough people in the short term, you can make a killing on the stock market. In truth, the technology itself is superfluous. It’s the hype that matters. And if the hype is over-inflated enough, you can even get your critics to do your work for you, broadcasting their fears about these supposedly world-changing technologies.

You’d think we’d learn. If an industry cries wolf enough times, surely we’d become less trusting of extraordinary claims. But the tech industry continues to cry wolf—or rather, “hammer!”—at regular intervals.

The latest hammer is machine learning, usually—incorrectly—referred to as Artificial Intelligence. What makes this hype cycle particularly infuriating is that there are genuine use cases. There are some nails for this hammer. They’re just not as plentiful as the breathless hype—both positive and negative—would have you believe.

When I was hosting the DiBi conference last week, there was a little section on generative “AI” tools. Matt Garbutt covered the visual side, demoing tools like Midjourney. Scott Salisbury covered the text side, showing how you can generate code. Afterwards we had a panel discussion.

During the panel I asked some fairly straightforward questions that nobody could answer. Who owns the input (the data used by these generative tools)? Who owns the output?

On the whole, it stayed quite grounded and mercifully free of hyperbole. Both speakers were treating the current crop of technologies as tools. Everyone agreed we were on the hype cycle, probably the peak of inflated expectations, looking forward to reaching the plateau of productivity.

Scott explicitly warned people off using generative tools for production code. His advice was to stick to side projects for now.

Matt took a closer look at where these tools could fit into your day-to-day design work. Mostly it was pretty sensible, except when he suggested that there could be any merit to using these tools as a replacement for user testing. That’s a terrible idea. A classic hammer/nail mismatch.

I think I moderated the panel reasonably well, but I have one regret. I wish I had first read Baldur Bjarnason’s new book, The Intelligence Illusion. I started reading it on the train journey back from Edinburgh but it would have been perfect for the panel.

The Intelligence Illusion is very level-headed. It is neither pro- nor anti-AI. Instead it takes a pragmatic look at both the benefits and the risks of using these tools in your business.

It has excellent advice for spotting genuine nails. For example:

Generative AI has impressive capabilities for converting and modifying seemingly unstructured data, such as prose, images, and audio. Using these tools for this purpose has less copyright risk, fewer legal risks, and is less error prone than using it to generate original output.

Think about transcripts of videos or podcasts—an excellent use of this technology. As Baldur puts it:

The safest and, probably, the most productive way to use generative AI is to not use it as generative AI. Instead, use it to explain, convert, or modify.

He also says:

Prefer internal tools over externally-facing chatbots.

That chimes with what I’ve been seeing. The most interesting uses of this technology that I’ve seen involve a constrained dataset. Like the way Luke trained a language model on his own content to create a useful chat interface.

Anyway, The Intelligence Illusion is full of practical down-to-earth advice based on plenty of research backed up with copious citations. I’m only halfway through it and it’s already helped me separate the hype from the reality.

Browser history

I woke up today to a very annoying new bug in Firefox. The browser shits the bed in an unpredictable fashion when rounding up single pixel line widths in SVG. That’s quite a problem on The Session where all the sheet music is rendered in SVG. Those thin lines in sheet music are kind of important.

Browser bugs like these are very frustrating. There’s nothing you can do from your side other than filing a bug. The locus of control is very much with the developers of the browser.

Still, the occasional regression in a browser is a price I’m willing to pay for a plurality of rendering engines. Call me old-fashioned but I still value the ecological impact of browser diversity.

That said, I understand the argument for converging on a single rendering engine. I don’t agree with it but I understand it. It’s like this…

Back in the bad old days of the original browser wars, the browser companies just made shit up. That made life a misery for web developers. The Web Standards Project knocked some heads together. Netscape and Microsoft would agree to support standards.

So that’s where the bar was set: browsers agreed to work to the same standards, but competed by having different rendering engines.

There’s an argument to be made for raising that bar: browsers agree to work to the same standards, and have the same shared rendering engine, but compete by innovating in all other areas—the browser chrome, personalisation, privacy, and so on.

Like I said, I understand the argument. I just don’t agree with it.

One reason for zeroing in a single rendering engine is that it’s just too damned hard to create or maintain an entirely different rendering engine now that web standards are incredibly powerful and complex. Only a very large company with very deep pockets can hope to be a rendering engine player. Google. Apple. Heck, even Microsoft threw in the towel and abandoned their rendering engine in favour of Blink and V8.

And yet. Andreas Kling recently wrote about the Ladybird browser. How we’re building a browser when it’s supposed to be impossible:

The ECMAScript, HTML, and CSS specifications today are (for the most part) stellar technical documents whose algorithms can be implemented with considerably less effort and guesswork than in the past.

I’ll be watching that project with interest. Not because I plan to use the brower. I’d just like to see some evidence against the complexity argument.

Meanwhile most other browser projects are building on the raised bar of a shared browser engine. Blisk, Brave, and Arc all use Chromium under the hood.

Arc is the most interesting one. Built by the wonderfully named Browser Company of New York, it’s attempting to inject some fresh thinking into everything outside of the rendering engine.

Experiments like Arc feel like they could have more in common with tools-for-thought software like Obsidian and Roam Research. Those tools build knowledge graphs of connected nodes. A kind of hypertext of ideas. But we’ve already got hypertext tools we use every day: web browsers. It’s just that they don’t do much with the accumulated knowledge of our web browsing. Our browsing history is a boring reverse chronological list instead of a cool-looking knowledge graph or timeline.

For inspiration we can go all the way back to Vannevar Bush’s genuinely seminal 1945 article, As We May Think. Bush imagined device, the Memex, was a direct inspiration on Douglas Engelbart, Ted Nelson, and Tim Berners-Lee.

The article describes a kind of hypertext machine that worked with microfilm. Thanks to Tim Berners-Lee’s World Wide Web, we now have a global digital hypertext system that we access every day through our browsers.

But the article also described the idea of “associative trails”:

Wholly new forms of encyclopedias will appear, ready made with a mesh of associative trails running through them, ready to be dropped into the memex and there amplified.

Our browsing histories are a kind of associative trail. They’re as unique as fingerprints. Even if everyone in the world started on the same URL, our browsing histories would quickly diverge.

Bush imagined that these associative trails could be shared:

The lawyer has at his touch the associated opinions and decisions of his whole experience, and of the experience of friends and authorities.

Heck, making a useful browsing history could be a real skill:

There is a new profession of trail blazers, those who find delight in the task of establishing useful trails through the enormous mass of the common record.

Taking something personal and making it public isn’t a new idea. It was what drove the wave of web 2.0 startups. Before Flickr, your photos were private. Before Delicous, your bookmarks were private. Before, what music you were listening to was private.

I’m not saying that we should all make our browsing histories public. That would be a security nightmare. But I am saying there’s a lot of untapped potential in our browsing histories.

Let’s say we keep our browsing histories private, but make better use of them.

From what I’ve seen of large language model tools, the people getting most use of out of them are training them on a specific corpus. Like, “take this book and then answer my questions about the characters and plot” or “take this codebase and then answer my questions about the code.” If you treat these chatbots as calculators for words they can be useful for some tasks.

Large language model tools are getting smaller and more portable. It’s not hard to imagine one getting bundled into a web browser. It feeds on your browsing history. The bigger your browsing history, the more useful it can be.

Except, y’know, for the times when it just make shit up.

Vannevar Bush didn’t predict a Memex that would hallucinate bits of microfilm that didn’t exist.


Picture someone tediously going through a spreadsheet that someone else has filled in by hand and finding yet another error.

“I wish to God these calculations had been executed by steam!” they cry.

The year was 1821 and technically the spreadsheet was a book of logarithmic tables. The frustrated cry came from Charles Babbage, who channeled his frustration into a scheme to create the world’s first computer.

His difference engine didn’t work out. Neither did his analytical engine. He’d spend his later years taking his frustrations out on street musicians, which—as a former busker myself—earns him a hairy eyeball from me.

But we’ve all been there, right? Some tedious task that feels soul-destroying in its monotony. Surely this is exactly what machines should be doing?

I have a hunch that this is where machine learning and large language models might turn out to be most useful. Not in creating breathtaking works of creativity, but in menial tasks that nobody enjoys.

Someone was telling me earlier today about how they took a bunch of haphazard notes in a client meeting. When the meeting was done, they needed to organise those notes into a coherent summary. Boring! But ChatGPT handled it just fine.

I don’t think that use-case is going to appear on the cover of Wired magazine anytime soon but it might be a truer glimpse of the future than any of the breathless claims being eagerly bandied about in Silicon Valley.

You know the way we no longer remember phone numbers, because, well, why would we now that we have machines to remember them for us? I’d be quite happy if machines did that for the annoying little repetitive tasks that nobody enjoys.

I’ll give you an example based on my own experience.

Regular expressions are my kryptonite. I’m rubbish at them. Any time I have to figure one out, the knowledge seeps out of my brain before long. I think that’s because I kind of resent having to internalise that knowledge. It doesn’t feel like something a human should have to know. “I wish to God these regular expressions had been calculated by steam!”

Now I can get a chatbot with a large language model to write the regular expression for me. I still need to describe what I want, so I need to write the instructions clearly. But all the gobbledygook that I’m writing for a machine now gets written by a machine. That seems fair.

Mind you, I wouldn’t blindly trust the output. I’d take that regular expression and run it through a chatbot, maybe a different chatbot running on a different large language model. “Explain what this regular expression does,” would be my prompt. If my input into the first chatbot matches the output of the second, I’d have some confidence in using the regular expression.

A friend of mine told me about using a large language model to help write SQL statements. He described his database structure to the chatbot, and then described what he wanted to select.

Again, I wouldn’t use that output without checking it first. But again, I might use another chatbot to do that checking. “Explain what this SQL statement does.”

Playing chatbots off against each other like this is kinda how machine learning works under the hood: generative adverserial networks.

Of course, the task of having to validate the output of a chatbot by checking it with another chatbot could get quite tedious. “I wish to God these large language model outputs had been validated by steam!”

Sounds like a job for machines.