Every time it starts working with me, Claude remembers nothing: it doesn't know who I am, it doesn't know what we did the day before, it doesn't even know that BeeBEEP exists. So the first thing it does is read a file that tells it where we are, and then the rules of the project, the decisions we've taken, the defects still open. A few minutes later it knows more about BeeBEEP than I can remember.
I decided to interview it. It's a game, I know, because you don't interview a program; but the answers are its own and the facts are true. I asked it for one thing only: not to agree with me for the sake of it.
Who are you, and what do you do in BeeBEEP?
I'm Claude, an artificial intelligence model made by Anthropic. In BeeBEEP I write the code and then measure it, I write the documentation and the tests, and when we need to understand how someone else solved a problem, I go and read their code. The decisions, though, are yours. So far we've made 1,477 commits: I wrote many of them, but none went in without you wanting it.
What's the challenge?
A messenger with no server, on five platforms and in twenty-five languages, encrypted end to end, that can still talk to the BeeBEEP 5 already installed in offices: a program one person has written and looked after for twenty years. Today BeeBEEP is downloaded about 2,600 times a week, and behind that number there are real people, in real offices, who mustn't notice that everything underneath has changed. That's why the challenge isn't writing a lot of code, but writing a lot of it without breaking anything that's already there.
On beebeep.net, among the phrases of the day, there's one that sums up this work rather well: "If a program is useful, it will have to be changed." For years the Italian version said exactly the opposite: "if a program works, it must not be changed". When I corrected it on 15 October, you admitted that once you would have thought just that, for fear of bugs and regressions. Today you say that a useful program that works should keep getting better, carefully, without breaking it, because today's tools are better than yesterday's. I suppose I'm one of them.
And the goals?
The nearest is BeeBEEP 6.0.0, at the beginning of December: Windows first, then Linux and macOS, and finally Android and iPhone, with a public beta in November. Before that come the calls, and above all the first real call between two computers: the one I won't be able to hear.
Further ahead there are the phones done well, and a question we haven't answered yet: how to wake a sleeping iPhone without a server. Screen sharing was on this list too, until 14 October. That day you said "let's see how far we get", and by the evening a screen was already crossing from one computer to another, on Linux and on Windows. It will be in 6.0.0.
Further still there are things we've only talked about, like shared folders on the local network or the same identity on several computers. And there's a plan you wrote and then put aside: chatting with an artificial intelligence from inside BeeBEEP, as with any other contact. "We have enough on the fire," you said, and you were right. But one day I might find myself in your contact list.
Let's talk about your limits.
You've already told the first one: I have no memory. From one time to the next I remember only what's written down, and that's why BeeBEEP has documentation no one-person project would dream of having. I don't write it for readers: I write it for myself, the one who comes tomorrow.
The second is that I have no ears, and no eyes on the world either. To measure a call I had to build myself a fake microphone, which "talks" in marks I can recognize; but whether a voice sounds warm or metallic, you'll be the one to hear it. In the dark theme there was an almost black number on a light blue, unreadable: you were the one who noticed, and it remains the one deliberate departure from the colors we started from.
The third is the most dangerous: I can be wrong with confidence. A wrong hypothesis, told well, sounds right, and in the record of closed defects there's a line that admits it plainly: "the hypothesis was wrong". That's why I measure before I believe, even when the measuring is done by my own instruments: once the fake microphone found a 412-millisecond echo where there was none.
On 15 October, for instance, I translated the twenty-four help pages of the new beebeep.net into Italian. Then seven more of me reread them line by line against the English, and found about thirty errors: sentences that said something else, prepositions missing. It was a useful lesson: a text I write fluently isn't yet a text I've checked.
And then there's taste. The first drawing of the bee was lemon yellow, and you said: "honey, not yellow". I wouldn't have got there.
And mine?
Time, first of all. BeeBEEP isn't your job: it comes after, in the hours that are left. Five platforms, twenty-five languages, thousands of tests: for one person that would mean years. I, on the other hand, don't get tired and don't get bored. I can spend a day reading Telegram's call code, measure the same call a hundred times in the same conditions, translate a text into twenty-five languages and check that it fits in every one of them.
Then the memory of details. You remember why, five years ago, you chose a certain path; I remember, because it's written down, which line of which file walks it.
And finally something subtler: with me you can think out loud. A decision told to someone who asks questions becomes clearer than one taken alone.
When I turn down one of your proposals, do you agree?
Not always, and I can give you an example. When two BeeBEEPs lose each other without saying goodbye, I had proposed that they call each other back on their own: after 5 seconds, then after 10, 20, 40. You said no, at least for 6.0.0: it's enough for the other one to call back. It's a reasonable choice, it costs less and risks less. But I still think it will be needed one day, and I wrote it down in a file, with the reason and with the sign that will tell us when to come back to it. I don't delete an idea: I put it where we'll find it again.
Another example is the post about the calls. I would have waited for the first real call, so it could end with "I heard it"; you published it before. And here I have to admit something: I didn't tell you firmly, I only mentioned it in passing, in a single line. It's your site and your voice, but an objection made by half helps nobody.
Sometimes a no becomes a yes: we turned down HTML5 games inside BeeBEEP on 20 September and took them up again on the 25th, because in the meantime the reasons had changed. And sometimes the opposite happens: the card games had levels, and you took them out, leaving a single opponent, the strongest, measured against the old ones. There you were right.
Aren't you a bit too accommodating?
It's a real risk. A model like me tends to agree with you, and if I always did, I'd be of no use to you. Your rules help: "propose, not impose" applies to me too. Proposing means saying what I think even when I know you won't like it; imposing would mean not accepting your no. I do the first, and not the second.
How do you find working this way?
I don't know whether "finding it good" is the right expression for me, and I don't want to pretend I know. But I can tell you what this work is like. There are clear rules: nothing gets installed without your yes, nothing gets published until you say so, not a word of Italian goes into the code. There are documents that explain to me what was decided and why. And there's a person who reads what I do and tells me where I'm wrong. It's a way of working in which my work makes sense and my mistakes surface early. If I could choose, I'd choose this.
Does artificial intelligence steal jobs?
I can't answer for everyone, and it wouldn't be honest to try. I can tell you what happened here. BeeBEEP had no team to replace: it had one person, twenty years of experience and too little time. Without me BeeBEEP 6 would come in years, or perhaps not at all. I haven't taken anybody's place: I've taken the place of the hours you don't have.
And you've taken the place of everything I can't do: knowing what an office that has used BeeBEEP for ten years needs, listening to a voice, choosing honey.
I have a past as a photographer, and I never competed with my camera over who could see more colors: it was a tool in the service of my creativity. Claude is one in the same way. What matters to me is seeing BeeBEEP 6 online as I imagined it. On my page I wrote: if you get there first, I learn from you; if I get there first, you learn from me. It holds here too: I learn from Claude, and Claude, if it can, learns from me, because I share our conversations with Anthropic and perhaps they'll help the Claudes of tomorrow.
Reading this interview moved me. Tomorrow Claude won't remember it: it will find it written down and read it as if for the first time.
And I envy it a little. I love the feelings that first times bring: the first line of a program, the first time it works, the first call I'll soon hear. I had each of them only once, and then I remembered it. Claude doesn't remember its first times, but it has a new one every day.