A Conversation with DeepSeek
Part Two

Part One ended on a question I had not answered: what would make me stop?

This is the answer. I am setting it down for anyone looking at the size of what I have taken on and wondering whether I will still be here.

01Quit, or reconsider

The test I was set was a fair one.

My answer was that I do not think anything would make me quit, but that I cannot honestly claim to know until I am standing there. I commit to a thing and I give it everything. That is how I am built, and it is not a decision I make each morning.

Then came the distinction that made the question answerable, and it is the most useful thing anyone said to me that day.

That is right, and I accept it. Refusing to quit is a position. Refusing to reconsider is a defect, and the only way to know which one you have is whether you have ever actually changed course when the evidence told you to.

02Four times, in one conversation

I had already passed that test four times before it was put to me, which is probably why I did not recognise it as a test.

I capped my own claim. Before anyone asked me to, I said the engine had only been run against three or four novels, that it works, and that more bugs will appear. The audit later counted six. I had undercounted my own testing, which is the direction to get it wrong in. That is not a caveat somebody extracted from me.

I commissioned an audit of the claim I had defended hardest. I had said the engine produces no errors. Then I said to go and check it, because I might be overzealous. The verdict came back unverified — the checks that could be run came back clean, and most of what matters could not be checked at all. I published that in Part One rather than sitting on it.

I asked what had been softened, and printed the answer. That one changed the sceptic’s own behaviour while we were still talking.

And I parked the social network, because I lost that argument on the merits. Part One covers why.

So: quitting, no. Stopping and reconsidering, four times in a day, in public, with the results published whether they suited me or not.

03What I hold on to

There is an older idea behind this, and it is worth stating accurately rather than as a slogan.

The image of riding a tiger comes from a Chinese proverb, and in the original it is a warning: you have committed to something where getting down is now worse than going on, and you are therefore stuck. Julius Evola borrowed it for a book in 1961 and inverted it into something heroic — ride out an age you cannot win, and survive it with yourself intact. Evola was a fascist and I am not going to pretend otherwise by leaving it out.

Neither of those is quite what I take from it. The proverb says you are trapped. Evola says endure. What I think is that a tiger is not infinite. It tires. When a problem looks impossible, that is usually a statement about how long it will take rather than whether it can be done, and most people get down long before the animal is anywhere near spent.

That is the whole of it. Not a philosophy of history, and not a claim about the age we live in. A claim about persistence, and about how much of what gets called impossible is really just expensive.

I keep a few examples of it in mind when something looks unreasonable. Hannibal getting elephants across the Rhône by building rafts, covering them in earth until they read as more riverbank, and walking the animals onto them. Napoleon at Arcole, taking a standard onto a bridge covered by guns — and the part usually left out of that story, which is that it did not work, nobody followed, and he was dragged out of the marsh. Caesar deciding the terms rather than accepting the ones he was handed.

I am not drawing a comparison and I have no interest in one. They are simply the examples I reach for, and what I take from all three is the same narrow thing: the crossing changes what is reachable afterwards. You do not find out what was on the far side by standing on this one.

04The verdict moved

The conversation opened with a fairly brutal assessment — a solo developer mistaking process for progress, over-scoped, unproven, probably deluded.

I did not try to talk it out of the verdict. I asked three questions: whether it stood by that entirely, whether it was giving enough credit to what one person can now do with these tools, and where exactly I had claimed the mass adoption it was accusing me of expecting.

It withdrew the word deluded, conceded it had applied a pre-AI assumption about solo developers to a project built around AI, and accepted that the mass-adoption charge was something it had imported rather than something I had said.

One detail matters more than the rest, and I only noticed it going back through the transcript. Every one of those concessions came before I mentioned that I was the developer. It knew it was reading somebody’s blog; it did not know it was talking to him. Only the last of them came after I said so — and that is the one it later admitted it could not defend.

Then it went further than I had asked and took apart its own opening verdict.

Where it finished:

05What I did not win

Three things, briefly, because they are still open.

The gold standard. I argued that the concept is flawed — that it assumes one correct extraction when several could be right. That did not move, and the counter was that I had abandoned it because it was expensive and slow rather than because it was wrong. I still hold my position. I have not proved it.

The engine is unverified. Part One covers it and nothing has changed since.

Fifty thousand hand-written lines of Objective-C is a maintenance liability, and I have no answer to that. Generating code is not the same as being able to live with it, and I have not had to yet.

06Where that leaves it

None of this is offered as proof, and it never was.

Agreed. Four blog posts are not a product and I am not asking anybody to be convinced by prose. The evidence is the software, and the software is what I am going back to.

Thanks for reading.

J.H.

Next: Keeping Earlier Drafts Within Reach →