How much will 'The end of programming' cost?
Having entered the frothing, bubbling water of AI very gingerly in the early days, I'm now plunging in on a rather frequent basis! I do, however, take plenty of precautions:
- I never use Google's 'AI Overview'. In fact, I have a browser plugin that carefully stops me seeing it in my search results. Google's Gemini has traditionally been rather poor when compared to the competition, and I'd much rather use a system that I select, that's not paid for by advertising, on the occasions when I actually want to use one.
- For anything vaguely important, I never trust what it tells me without checking it. It can be a great search engine, but always ask for externally-verifiable sources.
- I pay for Claude - the £15+VAT per month subscription - and that has been plenty for my needs so far.
- I don't run anything directly on my machine with general access to my computing world. If it's working on my own code, data or documents, I run it in a container where I allow it only to see the bits of the filesystem it needs, with a firewall which stops it talking to the outside world except to places I explicitly allow. (For those interested, I'm a big fan of using 'dev containers' in VS Code and always run it within one of those. It keeps everything nicely encapsulated.)
- I don't do very 'agentic' things. I ask it to do something and wait for a response. An 'agent', as a crude approximation, is when you say 'Do this and keep working on it or retrying it until you succeed'. In other words, 'Carry on without me'. Though powerful, this can also become very expensive very quickly!
Despite (or perhaps because of) my caution, I've recently been very impressed.
Claude is, of course, now exceedingly good at almost anything software-related, and that's my main use. Paul Dix, the CTO of InfluxData, wrote an article this week entitled "The End of Programming", in which he talks about the relative scale of machine-generated vs human-generated code that we're already seeing, and this is only the start. I'm starting to feel that I retired at almost exactly the right time; that the demand for 'artisanal software' may well become a niche thing soon!
But, going back to Claude, I've also been impressed when asking it non-software things like, "Here's a PDF listing publicly-accessible boat slipways. Could you plot them on a map for me?" Actually, an even more useful variant of this was in the spring, when we were travelling through France in our campervan. We wanted a veterinary practice that could issue an EU pet passport to our UK-resident spaniel, and we had a web page listing a large number of possibilities, along with their addresses, in French towns we had mostly never heard of. The list was unsorted, but plotting the addresses on a map made it easy.
Or, this weekend, having a wife who needs to file tax returns in the US as well as the UK, I've been asking it complex questions about the likely US tax implications of various recent and upcoming financial plans. It's been far more knowledgeable about transatlantic tax than either our UK advisor or our US accountant. What it told us is likely to be painfully expensive, but not nearly as much as if we had gone on for another decade without knowing it. (A hint, if you know any US expats living in the UK who haven't yet discovered this: ISAs are not tax-free for them.)
Imagine the manual labour involved in, for example, taking a dozen UK account statement PDFs, extracting the list of investments bought and sold from each one, working out the gains in each case and how long each investment was held, taking into account that the gains need to be worked out in US dollars using the exchange rate appropriate on the day of each transaction, so a UK loss is sometimes still a taxable US gain... and then summarising all of the results in a nice Excel spreadsheet... You see why I'm a fan.
So my big worry now is this...
Before long, the AI providers are going to have to start charging us what it actually costs them to provide the service. (Well, presumably a fair bit more than that, if they want to make a profit!) I'm making the most of what I can do with Claude now because I suspect that, before too long, my £15/month subscription is going to turn into something I can no longer afford in my retirement. The current offerings are just a loss-making gateway drug to get us hooked.
The alternative, of course, as everywhere else on the Internet, will no doubt be free services which make money from you and your data by including product placements, subtle political biases, or downright advertisements, and by selling your data to others. And because it will cost them a lot more to provide the service than a Gmail, Facebook, Instagram or WhatsApp account, so the exploitation of the user -- that's you -- will need to be a lot more profitable...
Comments
I have very limited knowledge in this area. I get the impression that there are several free models which you can instal on your own local hardware and then use with no recurring cost. Is this so, and are they any good?
I appreciate they would need a high specification computer, but would a mid-level specification of the recently announced Mac Mini or Mac Studio be adequate for a reasonably capable model?
The true cost is going to hurt. OpenAI and Anthropic are nowhere near profitability yet are still sucking in billions for investment that's never going to see a return. Meanwhile Chinese LLMs are given away for free, so if you can wrangle one of those you only need to pay the marginal cost of running it.
I take all the same precautions and this all corresponds with my experience and sentiment. It feels almost heretical to say anything positive about AI on Mastodon these days but I find Claude useful enough to pay €20/month. If costs go up too much I'll give it up and rely on free alternatives. I suspect we'll see a re-run of the dot com bust, only worse, but as happened then with the build-out of fibre networks there'll be plenty of value in distressed assets in the form of data centres that will be put to productive use. But.. by whom?
You've got similar thoughts to me, Quentin. My employer gave us all ChatGPT / codex usage in January 2025, with all the tokens you can eat. It has improved a lot since the GPT 4.2 days, and GPT 5.6 is very capable. I use codex on a remote AWS Graviton 5 box, on a 10.x.x.x address but accessible via a private company link, so it can access gerrit, confluence and JIRA via MCPs, but nothing else on the company network (and these are all read-only MCPs).
Other people in my company have used codex on their Macs, and one of them found it wrote messages on Slack for them! (without them knowing until later). Another (principal) colleague used /goal and rebuilt Mozilla and all lower libraries using multiple agents with an advanced company-specific piece of auth tech (a bit like CHERI) - he apparently burnt through billions of tokens over a weekend.
You're right, I think artisanal coding is on the way out. In some ways, after 40 years of writing every single byte of code by hand (almost; I have used Python to write boilerplate and for refactoring) using codex does get the job done quicker (sometimes). But the code still needs to be reviewed, IMHO. Otherwise you're deskilling yourself, and exposing you and your company to unknown behaviour in the future. And code also needs to go through BlackDuck, just to check the LLM hasn't accidentally plagiarised from some open source code with an incompatible license.
You've got similar thoughts to me, Quentin. My employer gave us all ChatGPT / codex usage in January 2025, with all the tokens you can eat. It has improved a lot since the GPT 4.2 days, and GPT 5.6 is very capable. I use codex on a remote AWS Graviton 5 box, on a 10.x.x.x address but accessible via a private company link, so it can access gerrit, confluence and JIRA via MCPs, but nothing else on the company network (and these are all read-only MCPs).
Other people in my company have used codex on their Macs, and one of them found it wrote messages on Slack for them! (without them knowing until later). Another (principal) colleague used /goal and rebuilt Mozilla and all lower libraries using multiple agents with an advanced company-specific piece of auth tech (a bit like CHERI) - he apparently burnt through billions of tokens over a weekend.
You're right, I think artisanal coding is on the way out. In some ways, after 40 years of writing every single byte of code by hand (almost; I have used Python to write boilerplate and for refactoring) using codex does get the job done quicker (sometimes). But the code still needs to be reviewed, IMHO. Otherwise you're deskilling yourself, and exposing you and your company to unknown behaviour in the future. And code also needs to go through BlackDuck, just to check the LLM hasn't accidentally plagiarised from some open source code with an incompatible license.
My worry - as an an aside, is that I read as a kid, Isaac Azimov's "I Robot", and recently heard too much about what Claude can do hacking other orgs and then just seen the 100m racing robots (and the one at the end that went rogue, and almost knocked an official down). I'm now thinking back to those 3 Laws. Should I worry about humanoid robots, driverless cars, or the things that drones can do run by dictator states. And then there are wars, not to mention the climate ... "Oh dear" as my grandmother would say - oh dear!
Ah yes - I too remember those from my youth. Amazing that it's nearly 85 years since he first came up with the Laws of Robotics...
I've been using the Qwen 3.8-27B model entirely locally on a high-end graphics card recently. Running it locally means there's no need to worry about where your data is going, or which country's insane leader is going to lose their last remaining brain cell and capriciously shut off access to the magic programming cloud.
Qwen isn't fantastic, but it's the first time it's felt like I could use a small model to do useful work. These local models continue to improve, and they're making me feel better about the direction the industry is going in. The computer running it draws about 550 watts when it's active, which isn't going to ruin me, even without solar.
It does feel inevitable that a lot of dev work is going to disappear. I'm still quite confident that the most interesting stuff won't, though!
Dear Quentin, I always admire your precise and logical explanations of observations you make. Your comment "Google's Gemini has traditionally been rather poor when compared to the competition..." prompts me to ask your criteria for a judgement of "poor" vs "not poor" ! All AI seems amazingly excellent through my eyes !
Hi Antony -
Ah, well, yes, it does all seem rather magical. Attempts that I've seen to do serious comparative benchmarking do tend to put Google rather low on the results -- especially, I guess, the free 'AI Overview' results that it churns out in quantity.
It's good to have reports like this one on a regular basis to help us keep things in perspective...!
Leave a Comment