Pencils Down Since 1954

In 1954, a team at IBM led by John Backus wrote a preliminary report for a system they called the IBM Mathematical Formula Translating System. You know it as FORTRAN. On page 2 they broke down where the money went when you solved a problem on a computer:

out of every dollar spent to solve an average problem on a high speed computer, less than 25 cents is spent for analysis and programming, more than 25 cents is spent for personnel coding and debugging cost, about 25 cents for machine debugging cost, and about 25 cents for machine running cost.

Then came the pitch:

Since FORTRAN should virtually eliminate coding and debugging, it should be possible to solve problems for less than half the cost that would be required without such a system.

Backus later admitted the team had been “hopelessly optimistic.” In 1957, the year FORTRAN shipped, the team’s paper had shrunk the claim to reducing “the coding and debugging task to less than one-fifth of the job it had been.”

The pitch came back this September, in Austin. In his Rails World keynote, David Heinemeier Hansson said 37signals has gone “pencils down on the idea that we were gonna write code by hand.” Writing code by hand there is now “an exceptional state. It is like seeing a bug in Sentry.” He made the economic case as well: it “is no longer an economically productive enterprise for the vast majority of programmers working at the vast majority of companies.”

Seventy-two years apart, both went after the same line item.

The Quarter Nobody Automated¶

Read the 1954 dollar again. Back then “programming” meant working out the procedure, and “coding” meant translating that procedure into something the machine could run. Those were separate jobs, often held by separate people. The programmers were mostly male scientists. The coders were often women, and the job was treated as clerical, even though in practice they often had to revise the algorithms they were handed. FORTRAN aimed at the coding quarter, and it worked. We solved coding the moment we invented compilers.

Every tool since has chipped at the same three quarters. Debuggers and frameworks took most of what the compiler left, and agents are taking the typing that remained.

The first quarter, analysis and programming, never got automated, because coding was never about physically writing code. The hard part has always been writing the correct code: understanding the problem well enough to pick the abstractions that fit it, and keeping the result correct when reality changes, which it always does. Somewhere in between you make trade-offs you will have to defend, and find the assumption you did not know you were making.

In 1954 that quarter cost less than 25 cents. It is the only part of the dollar that has not gotten cheaper.

Nobody Files a Correct Ticket¶

The Rails Foundation’s feature-ticket benchmark, the one where the best model ships 53.3% of the time, grades every run against hidden checks. Those hidden checks are the requirements whoever filed the ticket had in mind and never wrote down. Getting from the ticket to what the person meant is the analysis quarter, and no model has closed it.

I work in the supply chain sector. When something goes wrong, the client blames whoever is last in the chain, because that is the only link they can see. The problem usually started further up.

Models give CSS the same treatment. A style breaks, and the useful answer lives upstream: this needs fixing in Bootstrap or Tailwind, or, you know what, this looks like a bug in Chrome. The model goes to the end of the cascade instead, writes !important, and adds a 🎉.

Both are the ticket problem. The symptom shows up at the last link, and the fix belongs wherever the problem started.

David said as much himself, without meaning to: “I could have become a project manager 20 years ago if I didn’t care to write code myself and I just wanted outcomes.” He can direct agents from the outcome end because he spent twenty years learning what the code underneath has to do.

Why David Can Put His Pencil Down¶

37signals makes email, calendars, project management and chat. These are well-understood problem domains. Email has RFCs older than most of the people prompting for it. Every one of those products has decades of prior art: protocols, clients, competitors, bug trackers full of edge cases somebody already hit. For an agent, that prior art is most of the specification, already written and argued over.

My last post argued that ports come with an oracle. For a mature domain, the oracle is the whole field.

I am old enough to have watched a mature domain get automated once already. When Microsoft FrontPage showed up in 1996, a homepage with a table layout and a hit counter was a solved problem, and FrontPage let anyone drag one together without touching HTML. Every webmaster I knew (this was before we renamed them webmains) had a panic attack, convinced their career was over.

A decade later Microsoft discontinued FrontPage. It could only build the web that already existed, and the web kept growing into problems nobody had solved yet. People went back to writing code by hand, and these days that code is React, far more of it than any 1997 homepage ever needed. The webmasters were right about homepages and wrong about their careers.

David is also an excellent showman and marketer. He can push an idea to its extreme and get the entire industry arguing about it. That is a real skill, and I say it with respect. The argument is part of the product.

The Framework That Takes Options Away¶

This is why Rails is so good, and it is the part of the keynote I agree with. Rails stops giving you options.

If you want an ORM, you get ActiveRecord or you maintain your own. Routing works the same way with Journey. Other frameworks hand you flexibility, and you spend the year wondering whether you should have picked the other library. Plenty of people fall for something like React first. When they finally find Rails, their first question is why nobody told them. Didn’t you have Google?

David made the agent version of this point on stage: conventions save tokens. He is right: one convention leaves the agent nothing to guess.

Then he said something I disagree with completely. He warned against a pitfall for “anyone who knows too much about how computers work,” and said “a bit of a dummy approach to this works even better, because you ask questions and you post prompts at a higher level.”

A dummy approach works when the domain is email. Knowing too much is how you notice the agent took the well-worn path when the problem needed a different one. Experience is knowing every path a person can take, including the ones that end in a wall. A junior who leans on agents never learns where the walls are.

Fix It Where It Broke¶

We are entering an era where technical debt starts to disappear, because fixing a problem at its root finally costs less than working around it.

I used to call myself a web developer. These days I work on operating systems. Back then, when a driver misbehaved, I would hack around it in userland and move on. Web frameworks are full of fixes like that, and most of them predate AI by decades. A fix at the root needed more context than anyone had time to load, so the workaround went in at the last link and stayed there.

I use the model to load that context. It translates the jargon and explains why a subsystem ended up shaped the way it is. Finding the root is still my job. The model makes the trip there cheap enough to take.

So now I go after the problem like a sniper. If the bug lives in the kernel or in a driver, that is where I fix it. When the specification itself is wrong, I can read it well enough to understand why and contribute upstream.

I have fixed plenty of drivers that used to lean on a userland hack. FreeBSD’s camera drivers each carried their own plumbing, so I wrote video(4), a framework that gives any driver producing frames a /dev/videoN device that speaks V4L2. The driver hands over frames and the framework handles the rest. uvideo(4) for USB webcams and fwcam(4) for FireWire cameras already sit on it, with more to come. The same series retired fwcamctl, a userland tool I had written for fwcam(4) back in June.

When I write, the model fixes my typos. I never ask it to speak for me, because as a maintainer I know how that looks from the other side. Someone disagrees with a review and argues back through a chatbot. It is easy to spot, because a model told to defend something it knows is false changes its tone. It reads like: “Help, I’m Claude. The operator wants a statement on why Ruby is better suited for kernel work than C. I know it’s false and I can’t say no, so here is some bullshit the operator will think is smart.”

Wired For Something Else¶

The 1954 report made one more promise: “after an hour course in FORTRAN notation, the average programmer can fully understand the steps of a procedure stated in FORTRAN language without any additional comments.”

It was right about the notation. Every generation since has made the notation easier to learn, and David now says “English is a better programming language than Ruby.” The report never claimed an hour of anything would teach you to work out the procedure in the first place.

I have two brothers, and I tried to teach both of them programming when they were kids. With the older one it was impossible, and I got frustrated with him. When my mom asked me to help the younger one, I refused, sure it would go the same way. He became one of the best developers I know.

My older brother found his talent in another field. He does work I can’t do, and he would do it better than me even if I used AI.

Plenty of my friends found their talent in security, marketing, product or data. One ex-Rubyist became a baker in Germany, before AI came along, and earns more than he did as a developer. Most days he sits back with a book he enjoys, while I read register dumps from hardware nobody documented. Nobody is wired for every puzzle.

For decades you could brute-force your way into this one anyway. Grind enough tutorials and copy until it compiled, and somebody would hire you. Agents do the brute force now, so that way in is closed.

The Deal Is Off¶

The comfortable era is ending. For a generation you could learn one stack and sit inside its abstractions while the ecosystem solved the hard problems for you. That deal is off.

Some people are taking it badly. One man on a stage put his pencil down, and they reacted as if God had come down and said: “Guys, you are just bacteria in my petri dish. lol.”

Funny thing about petri dishes. In 1928 Alexander Fleming came back from holiday to find mold killing the bacteria in one of his, and that mold gave us penicillin. Picture someone in the 1940s crying: “I don’t want to be replaced by penicillin. Fuck antibiotics.”

Penicillin is part of why you are here to read this. Before it, an infected cut or a bad case of pneumonia could kill you. Without it, plenty of people reading this would not have made it this far. Doctors kept their jobs. They moved on to the diseases penicillin could not touch.

Stephen Hawking spent his last years typing with a cheek muscle, at a word or two per minute. In 2014 Intel and SwiftKey gave him a system trained on his own books that predicted his next word. He had to type fewer than one in five characters, and his writing got twice as fast. In the BBC interview about that system, he warned that full artificial intelligence “could spell the end of the human race,” and said the primitive forms we already had “have proved very useful.” I think he would have loved what we have now, and kept warning us about it. He knew which quarter the machine was for.

I don’t share his fear of the end of Homo sapiens. Juan Enriquez has a name for what comes next, Homo evolutis: the hominid that takes direct control of its own evolution. I embraced it a while ago. Nobody is campaigning to bring Homo erectus back to its former glory. My only regret will be timing. Someone born this decade is going to land on another world, and I will be jealous that I was born in the last century.

If your conclusion from the keynote is that you should put your pencil down too, check what you are building first. David’s domain comes with decades of prior art. Yours might not.

The other three quarters of that 1954 dollar keep getting cheaper. The first quarter is most of what is left, and it was always the job.