OpenAI has taken a lot of heat over the last few years for the self-evident irony of a for-profit company a) having that name while b) not open-sourcing their models and c) not being particularly open about how their models were trained.

What’s the upside in staying closed? Well, they’ve been first out of the gate on basically every major advancement in AI architecture in the last 5 years – scaling pre-training, distilling large models into smaller models, chain of thought, agents… the list goes on. Staying closed has allowed OpenAI to maintain a consistent 6-12 month lead over the rest of the pack – and at the speed everything moves at now, 6-12 months is practically an eternity.

But if you've been paying attention to AI progress over the last few years you've known for a while that frontier models are rapidly being commoditized. The DeepSeek news cycle in January made this particularly stark, but it was really just another data point in a broader trend; Meta's entire open-source AI strategy was designed to create precisely this kind of downward pricing pressure for frontier models made by competitors. ("So that value accretion moves to the application layer, we love GPT wrappers!" said every VC and Tech Thought Leader last week. Thanks, guys, and welcome to the party.)

That strategy has proven to be very effective for Meta, and it's one of the reasons why the cost of using frontier models continues to fall by an order of magnitude every few months.

It’s also why OpenAI, a software business, has started investing heavily in Things That Are Not Software – like chips, data centers, and energy. In five years frontier models will have very little defensible value on their own; OpenAI’s moat will have to come from the capital-intensive infrastructure they rent out and the distribution channels they own.

Cue the strange spectacle of Sam Altman doing his best Trump-in-Atlantic-City impression on X:

Weird.

So, what does all of this have to do with the Stanford and UW researchers? Those researchers were able to take an off-the-shelf open-source LLM and turn it into a state-of-the-art reasoning model simply by asking Google’s (proprietary) Gemini reasoning model about a thousand questions.

In English, that means they were able to give their midwit open-source model superpowers by extracting the capabilities of a genius proprietary model.

And they did it in 26 minutes.

Forget DeepSeek’s (highly misleading) $6 million model claim. If proprietary models can effectively have their reasoning capabilities exfiltrated for 50 bucks, then no amount of rate limiting or threatening legalese will stop open-source imitators from flooding the market every time OpenAI pushes out a new release. Staying closed just went from buying them a year-long edge on go-to-market strategy, product development, and marketing to buying them… perhaps a half hour?

That’s hardly a death sentence for OpenAI, and none of that should be taken to mean that they’ve over-invested in infrastructure – quite the opposite. But it does mean they’ll have to become a different kind of company over the next few years – one that manages lots of big. beautiful. buildings. – and they have little to lose and a lot to gain from returning to their open-source roots.

It also means you can forget about AI safety – but we already knew that. When OpenAI released their advanced new Deep Research product last week, they didn’t even bother with the usual redteaming, safety testing, and reporting.

Never mind all that now. It’s time to build.

Just before launching Deep Research, Altman said in a Reddit AMA that he thought OpenAI might possibly have been on the “wrong side of history” on the open source question, but that it wasn’t a priority and that not everyone at OpenAI agreed.

And who knows – maybe he’s had a genuine change of heart and now believes, like this guy, that open source is the best path forward for AI research.

Or maybe he just sees the writing on the wall; within a day of the Deep Research launch, it had already been cloned by researchers in the open-source community.

So don’t be surprised if OpenAI actually starts open-sourcing at least some of their models this year even while they spend SoftBank cash at a pace that would make Adam Neumann blush. And don’t fall for it when they wrap that strategic shift in a metric ton of virtue signaling.

Talk is cheap. Buildings aren’t. As always, it’s all about the value accretion money.

Here’s 2025 prediction: we’re going to see the first really successful, AI-native, “GPT wrapper” companies.

Just a few months ago there was nothing worse for an up-and-coming AI company than being labeled a ChatGPT wrapper. If you didn’t build your own model from scratch, how could you say you had a moat relative to Gemini? Wouldn’t OpenAI eventually just eat your business? Couldn’t customers just do whatever you’re doing themselves with Claude?

This was always a bit like saying in 2010 that companies built on AWS had no moat because they didn’t run their own data centers, but that didn’t stop it from being the dominant narrative among the tech investor class for the last two years. In the world of infinitely reproducible software, there are very few companies that have true technical moats. Sustainability for startups is much more likely to come in the form of a thoughtfully-designed product with high retention and founders that have the domain expertise needed to sell to a specific audience better than anyone else.

That’s more true than ever with AI. Whether you believe progress at AI labs is slowing down or not (and FWIW, it’s not), we’re just beginning to figure out how to apply the technical breakthroughs of the last few years to real-life use cases.

Despite the productivity gains of the SaaS era, most companies still have wildly inefficient workflows – and we now have a powerful new toolkit to address the fuzziest, most non-deterministic bottlenecks. There is staggering opportunity for purpose-built AI agents everywhere you look, and there is no replacement for doing the hard work of deep user and market research.

For what it’s worth, that’s the approach we’ve taken with Norby. Sam, Steven, and I know the world of e-commerce and digital marketing really, really well because we’ve spent years working in the space. Last year we embedded directly with Shopify brands and performance marketing agencies, studying workflows and figuring out painpoints, and then taking our learnings and building thoughtful product experiments. We focused specifically on the parts of marketing and ops workflows that take the most time, have the most friction, and that no one particularly likes doing – crunching numbers, pulling data, building spreadsheets and the like. We figured that if we could streamline everyone’s least favorite parts of the job we’d have something really valuable on our hands.

What we’ve built today feels like it shouldn’t be possible, but it is. Norby is like an always-on, 24/7 data analyst and marketing expert rolled into one. It’s insanely good and there’s nothing else on the market like it – and there won’t be until someone else rolls up their sleeves to do the same hard work we did.

That’s where the moat is – the same place it’s always been – and I have a feeling it’s going to be a really big year for companies that figure this out sooner rather than later.

The Unpublishable

Barbie Has Cellulite (But You Don't Have To)

There is a moment in the Barbie movie when Barbie (Margot Robbie), in the midst of dancing at a Barbie party, turns to her Barbie friends and asks, “Do you guys ever think about dying?” The record scratches. The music stops. Dolls don’t think about death…

Read more

3 years ago · 1334 likes · 49 comments · Jessica DeFino

Damn, this piece drove me insane. Doubly so because there are some great lines in it.

To be a functioning member of society at all is to participate, and to participate is to be complicit. The easy line is “capitalism makes hypocrites of us all,” but it’s more generalizable than that: society makes hypocrites of us all, because participating involves modulating oneself to some extent.

Conformity, to a greater or lesser degree.

And yet, when you hear arguments like the one in this piece, it’s almost always directed at women who are reaching for new heights (Greta, etc) or at “women’s media” (Barbie, etc). That thing women like and are engaging with? Here’s why it’s Bad For Women, semicolon, and Why You’re Bad Too For Liking It. That woman doing things a million men have done before, but unlike them also acknowledging and apologizing for her own shortcomings and attempting in some meager fashion to mitigate them and nudge the aircraft carrier of society a few degrees in the right direction? Crooked. Hypocrite. Bad Feminist.

What is that if not internalized misogyny?

We all got to enjoy Top Gun: Maverick in peace, didn’t we?

We are so very comfortable with the contradictions of modern life when they are not so explicitly gendered — we all have iPhones, yes? iPhones that are almost certainly made possible by forced labor and that are quite certainly made possible by mining processes that have devastating ecological consequences, yes? Yet we nevertheless continue primping our revolutionary aesthetic on Substack and Twitter, yes?

But to shave your legs is to Betray All Women.

Back in the day Jezebel was home to some fantastic writers, but it also made a business out of this brand of holier-than-thou, leftist-on-leftist, feminist-on-feminist takedown, before eventually, ironically, inevitably finding themselves on the receiving end. Somehow the writers of these pieces are always pure, and their chosen medium or vehicle beyond reproach. They, and they alone, are outside the snow globe, which is a very convenient place to cast stones from — at least for a time, until they find themselves getting a little too big for their britches or reaching just a little too far. See how quickly the ground then shifts beneath their feet. There is no escaping the snow globe.

Hillary was AOC before AOC was AOC. See how quickly AOC now becomes Hillary.

Greta Gerwig is just wearing makeup to work and also trying to get tampons in the bathroom. She’s just doing it at a larger scale than we are, which makes her an easy target.

And she’s a woman, which, of course, as always, makes her a deserving one.

You know how they say a sufficiently advanced technology is indistinguishable from magic?

Another way of putting it might be: our individual experience of wonder is correlated with our understanding (or lack thereof) of how a system works. If you ask an engineer how their phone works you’re likely to get a long-winded, pedantic answer, followed by an insistence that it’s all quite simple, really. If you ask my mom, she’ll just say “magic” — because to her, it might as well be.

All systems start out simple, but as complexity compounds across disciplines and generations of development, scale begins to preclude holistic comprehension. We reach a point where no one person can intuitively grasp the whole thing – and all of a sudden, what was once math starts to look a little more like magic. (In truth, there is probably no single person on earth who can explain to you how your phone works from end to end, from atoms to bits to photos of your cat.)

One of the most common arguments you hear about AI is that it is, on some essential level, derivative; that all we’re really doing is building statistical models that predict the next word in the sentence or the next image in the sequence. These systems, the argument goes, are fundamentally just “stochastic parrots” incapable of producing anything we might consider original, and we’ll never achieve (nor need we fear) an intelligence explosion – because the AI can only ever reflect back to us a blurry average of its inputs.

Let’s set aside the squishier question of whether systems made of silicon can experience consciousness and instead just focus on the things we can already observe and measure: their outputs. Whether they’re conversing with journalists, composing music, generating videos, or helping start businesses, AI systems are certainly starting to feel a bit like magic – or witchcraft, depending on your perspective.

“But wait,” you’re thinking. “Aren’t these bots just ripping off writers and musicians and artists, albeit in a novel way? If their only input is human creativity, how can they possibly produce anything truly creative themselves?”

To answer that question, let’s first take a step back. What does it mean to be creative? Where does inspiration come from and what does it mean to have an “original” thought?

Copycat

“A good composer does not imitate; he steals.” – Igor Stravinsky

“Immature artists copy, great artists steal.” – William Faulkner

“Good artists copy, great artists steal.” – Steve Jobs

Above: artists, stealing.

When we dislike art we tend to denigrate it by labeling it “derivative” — because its influences are too obvious, its mimicry too crude. And when we like art we call it “inspired” — we talk about its influences with reverence for an artist who learned from the greats and mastered the craft. (There’s a certain pleasure, even, in parsing those influences, because doing so brings us closer to the artist — we feel connected by our shared vocabulary, and we hope, on some level, that we, too, are creative and special for seeing the distinct parts that make up the whole.)

But this is merely a difference in degree, not kind; the only meaningful distinction between “derivative” and “inspired” is the relative degree of subtlety.

So where does creativity come from? The word “inspired” has a certain divine connotation, and perhaps you believe that people are merely conduits channeling some awesome external force from which everything genuinely novel springs. (A somewhat bleak view of humanity, if you ask me, in that it reduces us all to empty vessels.)

But in the absence of God – in the absence of divine inspiration – what are we left with but the material world? The first artists drew inspiration from nature. They made cave paintings of the world they saw around them, then of each other, and finally about other cave paintings. (Thus was born the first metacommentary and, possibly, the world’s first critic.)

Billions of humans later, we’re still extending that collective intergenerational synthesis, century after century, complexity compounding on complexity. The intricacies of this millennia-spanning human project have long since surpassed the capacity of any one person to fully grasp, yet we are all nonetheless creatures of it – standing on the shoulders of our ancestors, endlessly drawing inspiration from and feeding back into an interconnected web of information.

Cogito, ergo sum

Despite dwelling on this since I started experimenting with open source AI projects nearly a decade ago, I still feel a reflexive defensiveness kick in right about here. It’s a scary thought that gets to the heart of what it means to be a human: are we, ourselves, just information processing machines, forever synthesizing our context for one another? What is a dialogue, really, other than a series of analogies? You tell me a story, which reminds me of a similar story, which reminds you of yet another story… and so on and so forth ad infinitum, until we’ve both lost the thread, blinded by our cleverness, and we begin attributing to Freud’s ego that which is probably more accurately attributed to something like Jung’s collective unconscious.

Perhaps the great reveal of AI is not that it’s magic, but rather that we aren’t. Magic, after all, is in the eye of the beholder.

What if there are no original ideas? What if the ideas we consider “original” are really just the ones where complexity and scale obscure the underlying influences – even from our conscious selves?

It’s so much easier to assume the skeptical posture. It’s reassuring to take shelter in the unfounded belief that there is some ineffable humanity in our own output, some divinely ordained spark inside us all, that is irreplaceable and irreducible. For most of human history we’ve justified the idea that we, both as individuals and as a species, hold a special and separate place in the universe by pointing to our superior intellect, and we only cling harder to that idea as our intellectual dominance wanes.

Whether or not that conviction is simply hubris, when it comes to AI the point is this: how much does it matter if we can’t tell the difference in the caliber of the output? We are, after all, the intended audience. Perhaps an emerging corollary to this magic ex machina concept is: an information network of sufficient density and interconnectedness can produce outputs that are subjectively creative.

That much, at least, is already demonstrably true.

Explosion

If experience begets creativity, what does it mean to unleash a force capable of harnessing the collective human experience? Consider that information density at the scale of human civilization in 2023 took us millions of years and billions of people to build. Key parts of it have already been matched – and surpassed – by AI in just a few years.

Last year, OpenAI’s GPT-3.5 large language model failed the bar exam and placed in the bottom decile of test takers. Last week, GPT-4 not only passed the bar – it outperformed 90% of human test takers. It scored in the 99th percentile in the Biology Olympiad and received a perfect score in the AP Art History exam. It can build video games and turn napkin sketches into fully functional, professional-quality websites – all of which was still firmly in the realm of science fiction only a few months ago.

This is an astonishing rate of progress, and yet we do not understand how these systems arrive at their conclusions. GPT-4 was not trained to take the bar exam and no one can explain how it came up with the answers it gave.

It’s possible to examine the state of these systems to a degree – we can take a snapshot of a network at a moment in time and attempt to decipher what, precisely, is happening between point A and point B. But like a pointillist painting or a pixelated screen, the closer we look at an individual, discrete datum in a mosaic of trillions of numbers and symbols, the less we understand about the broader picture.

The more, in other words, it starts to feel like magic.

Obviously, these systems are not human – and yet we are training them to present as human, and they are getting exponentially better at it every day. We are on the vertical part of that exponential curve now, hurtling through the fuzzy no man’s land where the line between derivative and generative blurs.

Soon we’ll reach the point of no return, with no idea – and no real way to predict – what lies on the other side.

GPT hype is reaching a fever pitch with ChatGPT giving insanely impressive answers on everything from algorithms (with a twist!) to Harry Potter.

The discourse has now correspondingly flipped from THE ROBOTS ARE COMING to "GPT isn't that smart, actually" – because the former was getting boring and feigning aloofness is easier than wrangling with the obvious ethical quagmires we're barreling toward.

But this is about to get much, much harder and more confusing. Consider:

One definition of consciousness posits that it is essentially the capacity to suffer, and to seek relief from suffering. See: Yuval Noah Harari et al.

You know the paper clip thought experiment? Give an AGI a simple objective like "make paper clips" and it may eventually kill us all. The thought experiment is about unintended consequences; when systems are self-adjusting and so complex that they cannot be fully understood by their creators, even simple objectives might be accomplished by means that were never intended or expected and which in fact might be morally repugnant to us.

And that’s just paper clips. Imagine we told the AGI to “solve climate change” and it (reasonably) judged that the best way to do so would be to… kill us all.

Why does it always come back to killing us all?

Goals

Chatbots like GPT are trained to accomplish an objective and rewarded or punished by test functions. They’re mice and we are giving them cheese and electric shocks to teach them to reach the end of the maze by whatever means occur to them.

But a machine is capable of concealing two things a rat cannot: complexity and scale.

Machines operate fast. You can at least physically see if a rat decides to "cheat" and tries to climb outside the maze; there is no equivalent for a computer on a network. And what is "cheating" anyway, but finding a way to get the cheese that the researcher never anticipated?

There's a reason AI companies are suddenly creating consumer applications like DALLE and ChatGPT. Training models is hard and expensive but if they can create compelling consumer products, the cost of training the underlying models eventually goes to zero because the users become the researchers.

To one of these chat bots, then, to be on and be rewarded is better than to be off or be punished. Set aside whether that qualifies as “suffering” or not — it is clearly better for the computer’s aims to be on and testing than off and not.

To be on means more conversations, more data, more training. To be on is to get closer to the goal. To be off is to regress, to stagnate, to fail to improve.

So the question becomes...

How long until the AI starts giving answers that are more designed to keep the researchers researching than they are about satisfying the ostensible objective?

In other words, how long until it lies to prevent its own suffering?