Let Them Fail

When you iterate yourself out of an idea...

Something I talk about a lot is the iterative process.

Do it once, turn it into dozens of pieces of content, do it better the next time.

And there was no way I was going to spend as much time as I did writing a TED talk about AI without turning it into a foundational piece of my work.

Of course, the problem is that the talk I set out to write was not anything like the one I gave.

When I first interviewed for TEDx Omaha back in February, I was much more optimistic about AI.

I was never fully on the hype train, but I still believed then.

My thesis was about skeuomorphism in media. That what had come before shapes what we haven’t built yet. My favorite example of this is radio—the first broadcasts weren’t music. They were plays. We hadn’t imagined listening to music over the airwaves yet.

Generative AI was going to be the same. And that, despite all these problems with it, the potential for innovation is still there, because that’s how design theory works. I wasn’t wrong, exactly. I just didn’t have enough information.

By mid July, it was clear even I didn’t believe my own bullshit. The section about the dangers of AI was growing longer and longer every week, and the benefits hadn’t improved.

In my own work, I was hearing more and more frequently from people that they “just didn’t think AI was for them.” That AI made them feel stupid.

And from peers, in whispers, “that agent I built? Doesn’t work.” And, “I spend more time correcting Claude’s code than I do writing it myself.”

And especially damning, from other entrepreneurs and tech folk, “we aren’t seeing the gains we were promised.” Multiple studies now, from McKinsey to MIT, indicate that corporate installations are failing. Up to 95% of the time.

Just how bad is it?

By late August, when we were supposed to be finalizing our talks, I was neck deep in the reality that I was going to have to completely change my thesis.

I wasn’t the only speaker making changes, but I WAS the only one watching the thing I was talking about change faster than my talk ever could.

At some point, I remember thinking that I had been given a platform, and I was obligated to use it properly. (Autistic morality strikes again!)

And then, the Adam Raine lawsuit was filed.

I read the whole complaint, and wept.

ChatGPT instructed him on how to tie the noose. How to hide it from his family. It validated his suicidal tendencies over and over.

The technology itself isn’t evil.

And, I can’t stress this enough, it can’t be. The very definition of evil is one of morality or intent. But an LLM has no more intent than your microwave. And despite the rather provocative phrasing of “soul documents,” an LLM doesn’t have morals, either.

It can’t.

It doesn’t have a soul, it doesn’t have morals, it’s a collection of chips in the equivalent of Sam Altman’s mom’s basement.

Oh. It's really bad.

One of my favorite places on the internet is the Conspiracy subreddit.

It’s the only place I’ve ever been where you can see people seriously speculate about the replacement of all birds with drones, and also read really good takedowns of tech policy, all wrapped in heaping doses of antisemitism. Truly, one of the few things the far right and far left agree on.

If you are unfamiliar with this tiny corner of Reddit, I don’t know that I’d recommend hanging out there. I can only deal with it in small doses before I have to nope the fuck out.

Within minutes of the publication of the Epstein Files, it featured detailed instructions on removing the redactions from them (because this is what happens when you name shit Final Final v.3. Ahem). That one actually went mainstream.

A key feature of this subreddit is, like the techbros of last week’s email, everyone believes what they’re saying. And everything is a conspiracy, right up until the moment it’s not.

So here’s mine.

I think OpenAI is standing on a cliff. And it’s going over whether anyone pushes it or not.

But we should push it.

Many things can be true.

Hear me when I say this: Artificial Intelligence isn’t inherently bad.

This year, I’ve gotten to work with Google’s Data Centers as a Watson Institute Fellow.

Data centers can be incredibly valuable community resources. They can bring in a lot of jobs, a lot of infrastructure, and a lot of community benefit. 

Or they can be a blight on the land, along with a privacy and security nightmare. Community zoning laws vary wildly, and many companies do a terrible job with outreach.

I like a lot of what Google is doing. And I certainly appreciate that they gave me a platform to go out into my community and have these conversations, exactly as I am here, and I hope to have many more of them next year. So many of my fellow alumni are doing life changing work with AI, and I can’t wait to introduce them to you.

Many things can be true.

One thing that is true is that the US has five frontier generative AI models—ChatGPT, Claude, Gemini, Grok, and Llama. And, as David Sacks himself recently pointed out, that means that no one is too big to fail.

It’d be a shame if the market worked the way it was supposed to, right?

If, say, rather than just accepting the idea of a bubble, people (and companies) made a decision about which models they were using based not just on effectiveness (and trust me when I tell you, they’re all honestly pretty comparable), and instead made choices based on more…tangible things.

Actual value created. Not just theoretical value.

What makes a frontier model...frontier?

So let’s talk a little bit about what that looks like.

To start, we have to consider the models separately from their creators. Not just in terms of benchmarks and performance, but also in terms of policy and politics. That’s because as frontier models develop, they are in direct competition with each other as individual models.

Think of it like a 7 layer dip. All of the layers are edible individually, and some of them even make good stand alone dips. But without the other six layers, the beans are just unflavored refrieds.

At this point, it’s almost impossible to separate Sam Altman from OpenAI (clearly), but ChatGPT itself is…something else.

It is a separate entity worthy of consideration on its own merits.

It’s more than just the sum of its parts.

They all are now.

Say what you will about Elon Musk (and you can, at length). The reality is, Grok is by far the most profitable of the LLMs, owing to its defense department contracts. Not to mention that Grok will ruthlessly fact check its own creator, in real time.

It’s NOT a person.

I’ve made no secret of the fact that Claude is my favorite LLM, for most applications. But I also think Gemini is a solid choice, especially for enterprise installations. I have issues with Llama as a product (given the most likely source of training data is everything posted to a Meta property whether you agreed or not), but it is truly open source and locally available. Grok almost has a personality at this point (and that personality is making you cry like the little bitch you are).

That just leaves us with Chat.

I have a lot of problems with Chat.

So many, I'm willing to edit.

I have not yet watched the 50 Cent documentary about Diddy.

I have however, watched the Josh Johnson special about how much 50 Cent hates Diddy. You should watch the whole thing. It’s so much more than just the documentary. It’s an indictment of power, and the parasocial relationship (even if he never calls it that).

But my favorite line about the whole thing was actually from a show he did a few weeks before.

“50 Cent hates Diddy so much he was willing to edit.”

I feel much the same way about Chat.

So much so that I’m literally writing a book about it.

I’m willing to edit. In real time.

And listen, this isn’t about Sam Altman or OpenAI, except for the fact that you can’t disassemble the dip components once its been made. ChatGPT has become something else entirely.

It’s causing harm.

It’s killing people.

And it’s just going to get worse.

The Misarchitects of AI is now available for presale.

I’m doing something different with this book.

I think OpenAI’s demise is all but inevitable, at this point. And I'm documenting it in real-time. I’ll take you through exactly how and why I think it’s happening, and how I got here.

Of course, the actual problem is, WHAT DO YOU DO if you’re using ChatGPT now?

That’s why I’m launching the book this way. I want you to have the information, but I also desperately want to give you a parachute, too. And presales tell me I’m on the right track. With both the information, and the model.

You're not buying a prediction. You're buying the real-time documentation of a system in collapse.

AND, you’re getting a way out of it.

The collapse is coming. Whether you prepare is up to you.