Everything Looks Great in PowerPoint

GPT-6 Astra dropped, and let me say this up front so nobody misreads the rest of this article: it is a nice step up. I have put real time into testing it and I like it. I have watched it do genuinely cool things in the UI, including drawing real architecture diagrams, and if you know me at all you know I lit up a little at that. I like architecture. I like when a tool can actually reason about how the pieces fit. This is not me trashing the model. I use it.

Then I opened my feed and watched a thousand people post the same benchmark charts.

Bars going up. Percentages. Beautiful gradients. Everyone very excited that a number moved three percent on a test almost nobody in this industry could describe from memory. And a line I have carried around for years came right back to me, courtesy of Elon Musk:

Everything looks great in PowerPoint.

That is the whole article. But let me earn it.

Benchmarks are a PowerPoint slide

To be clear, my problem is not with Astra. My problem is with how we talk about it. A good model deserves better evidence than a bar chart.

A benchmark is a narrowly defined test. That is not an insult, that is literally what it is. Someone picks a bounded set of problems with knowable answers and measures against it. Useful. Necessary, even. But narrow.

The questions I actually care about do not show up on that chart. Does it hold up on real work? Can somebody prompt inject it in about four minutes? (I am writing a whole separate piece on that one.) How good is it at the boring everyday stuff a normal person needs done at 9:40 on a Tuesday?

"It got three percent better on the benchmark" and "it is meaningfully better at the thing you are going to use it for" are two completely different claims, and only one of them fits neatly on a slide. For what it is worth, my own read on Astra comes from actually using it, not from the chart. That is exactly my point.

Which brings me to the bigger thing, because this is not really about AI models. It is about every idea anyone has ever pitched me.

An idea in a deck is worth almost nothing

You can make anything look incredible in a deck. Nice charts, clean mockups, a tidy little story where every arrow points at the next arrow and nothing ever breaks. I have sat through those. I have made those.

The deck has no edge cases. The deck has no users. The deck has never had a customer email you at 11pm to say the thing did something insane. The deck is a fantasy where your idea works, and the entire value of building is finding out where reality disagrees with you.

Until you have tested it against something real, you do not have a product. You have a hypothesis with good fonts.

So no, I do not think beginners should pre-sell

Here is where I break with a lot of startup advice.

The standard line is: validate with pre-sales, sell it before you build it, get money down first. And for some people, in some situations, sure. But if you are brand new and you have never built anything, I think pre-selling is a trap.

Why? Because you cannot make promises about a world you do not understand yet. If you have no idea how software actually gets built, you do not know what is possible, what is trivial, what is a nightmare, or what quietly explodes at scale. So you sell something. And then you go try to build it. And you discover the thing you promised requires an integration that does not exist, or a permission model you have never heard of, or six weeks of work you priced at four hundred dollars.

Now you are not a founder. You are a person who owes somebody a product.

Build the thing first. Watch where it breaks. Then go sell, because now your mouth is writing checks your hands can actually cash.

And do not misread me here: selling is one of the most important skills there is. I have written about exactly that in the piece on sales. I am not anti-selling. I am anti-selling-a-thing-you-do-not-understand. Once your ideas are rooted in reality, once you actually know what software is and what it costs to make, pre-sell all day. Before that, you are just generating debt.

The edge cases are the entire education

You have a beautiful idea for a CRM. Gorgeous. And with today's tools you can spin up an MVP frighteningly fast, whatever the hot stack is this month.

So you ship it, and then your users start talking.

Hey, this broke. Hey, when I tried to buy a domain through you it would not let me buy www.mywebsite.com, it only let me buy mywebsite.com. Hey, why did it do that.

None of that was in your deck. It could not have been. I wrote a whole article about the time an .ai domain made me think I had broken my own software, and the lesson there was the same as the lesson here: edge cases will punch you in the face, and hardening a product against them is most of what separates software that works from software that demos.

You cannot harden something you have only ever seen in PowerPoint. There is nothing to harden. There is a picture.

Get an ugly prototype in front of people

This is why Elon and that whole camp are fanatical about getting a working prototype out fast, and why I am too. It does not matter if it is duct tape and chewing gum. My company ran on duct tape and chewing gum for a long stretch and it worked anyway, and I learned more in that stretch than in any planning session I have ever sat in.

It is going to be ugly at first. Genuinely, embarrassingly ugly. Ship it anyway. The whole game is iteration speed: build, put it in front of a human, watch what happens, fix it, repeat. Every loop you run is information you could not have gotten any other way, and the person who runs that loop more times wins. I went deeper on why speed alone is not enough in Judgment Is the Real Moat Now, but you cannot exercise judgment about a thing that does not exist yet.

People do not all learn the same way, and a prototype fixes that

One of my mentors taught me something that stuck: people consume information differently.

Show some people a concept on a slide and it lands instantly. Great. That is not me. Ideas explained to me in text or out loud tend to slide right off. They do not carry much weight until I can get my hands on them.

The way I learn is by taking things apart, putting them back together, and seeing where they break. That is systems thinking, which I happen to believe is the skill that matters most right now. And you cannot do it to a slide. There is nothing to take apart.

So when you show up with only a deck, you are quietly filtering out everyone who thinks like me. A working prototype, even a hideous one, lets the touch-it-and-poke-at-it people actually understand what you are talking about. That is not a nice-to-have. That is half your room.

Your idea of a perfect product is wrong. Mine was.

When I was 24 and knocking on doors, I had no idea what I was doing. None. And I had a very clear, very detailed picture in my head of the perfect product.

It was not what the market wanted. Not even close.

Elon has said a version of this himself and it is one of the truest things in business: your idea of a perfect product will shift as the market gives you feedback. Not might. Will. Which is exactly why the iteration loop matters so much. You are not building toward your vision. You are building toward the thing your vision becomes after it gets corrected a few dozen times by people who are actually paying.

You do not get those corrections from a deck. Decks do not argue back.

And selfishly, this is how I sell

Last thing, and this one is personal.

There are far better salesmen than me. If you handed me a room and said close it on raw charisma, I would rather not. I have done it. It is not my edge, and I know it.

But put a working thing in front of somebody, watch it solve the actual problem they came in with, and then ask how they would like to pay? That I can do all day. That is not charisma, that is evidence. The product does the persuading and I just stand there.

Which is the whole thesis, really. Everything looks great in PowerPoint. Nothing gets tested there.

Go build the ugly version.

Want to skip straight to the working version?

We build the real thing, put it in front of real users, and harden it against the stuff nobody plans for. That's what we do at HyppoAI.

Build it right