In the days of GPT 3.5-Instant, when we still had models with less than (fewer than?) a billion parameters—current SoTA models are in the trillions of parameters, now—we had problems that prevented industrial use. Early models couldn’t do arithmetic very well, they hallucinated information, they had frequent breakdowns, etc. It was a whole thing. Now, though, our frontier models are solving unsolved problems in mathematics, splicing human DNA to create custom cancer treatments, and writing a huge part of the world’s software. They’ve even been used to find vulnerabilities in Cloudflare, the primary provider of internet infrastructure!

This is to say that, if you want to have voicemail that answers in your voice and can schedule meetings for you, there’s really nothing stopping you from doing it. If you want to have a website chatbot that doesn’t suck, you can have it. You could even have an AI girlfriend if you wanted one.^[2]

We are no longer in the days of GPT 3.5-Instant. Our modern models are good enough; whether they perform 3.9% better than their predecessor or not makes little, if any, difference. (In practical applications.)

As more and more companies, people, and—importantly—investors realize that our models meet this threshold of being usable, money flows into adopting or developing novel applications of AI. If you don’t believe me, consider, for example, Allbirds: the shoe company whose stock price increased sevenfold after it rebranded to “Smartbird” and became an AI company. Money flows to AI.

And this money flow will cause innovation, as it always does, and consequently a loss in jobs. Software engineers said their jobs were safe two years ago; now, tech companies like Meta and Fortune 500 company Block lay off thousands of their technical staff. Block’s CEO, after firing 40% of Block’s workforce, explained that the primary reason was AI—and that he believes other companies will follow suit. Shortly after, Coinbase followed suit. Software engineering might not be so safe now, huh?

Increasingly, corporations are feeling the pressure to incorporate automation. This pressure spans accounting, law, software engineering, computer chip design (!), and it’s a matter of time before other industries feel the same pressures—and pressure to automate inevitably leads to automation and, consequently, layoffs.^[3]

There is a commonly-held belief in some circles that there will always be a need for humans in… whatever. People have filled the blank with all sorts of things (teaching, research, and customer service, as examples). As far as I can tell, this belief is simply false.

So long as our machines have limitations, there is going to be a guy who thinks, “What if we could automate this?”—and to be honest, thinking like this is what brought us here in the first place. It’s why we have self-driving cars. It’s what we have to thank for the Industrial Revolution. But it also means that it’s a matter of time until, one year from now or twenty, your job will be replaced by AI (or some future derivative of it).

Though I imagine AI will cause a record-breaking loss in jobs, I’m not too worried about the future. I think the world will have time to adapt,^[4] and overall the transition will be similar to the automation revolutions (such as the arrival of the Internet) we’ve experienced in the past. People will lose jobs. It’ll be a shame. But that’s the inescapable toll of progress.

Perhaps the more pressing problem, though, is that at the moment the vast majority of the world doesn’t believe that AI—which is only good for writing emails, right?—is all that. I’ve spoken to people who haven’t even used ChatGPT—and in the circles that read my writing, this is quite unfathomable. I’ve found that there are three camps of people when it comes to AI:

Here’s a fun fact of the day: of OpenAI’s weekly active users, 94.4% use the free tier. The vast, vast majority of the internet-capable human race is an Outsider.

And this is a problem. At the Outsiders’ level, “AI” is Google AI Overview, news about Grok generating some new thing involving minors, and pieces from technically illiterate mainstream media. For example, this piece from WIRED^[5] about how humans are safe in the realm of fact checking uses worthless models and actively misleads people about what’s possible with AI. As the article points out, Google’s AI Overview and ChatGPT Free genuinely suck. In fact, while I was writing this article, Overview failed me, too:

And if you were to hypothetically Google, “How many stairs on the Inca Trail?” it would tell you, confidently, how many steps (70,000) there are, not how many stairs (9,862, according to my calculations) there are. But the difference between ChatGPT 4o (from early May of last year), which the article uses to back up its claim, and GPT 5.5 Pro (or the difference between Gemini and Claude Fable 5), is large.^[6] So large, in fact, that when I first thought about how to describe it, the first words that came to mind were “larger than your mom.”

I’m sure she’s a nice, polite, attractive woman; I mean no insult. But at the frontier of artificial intelligence, models are autonomously solving Erdos problem after Erdos problem. They are literally splicing human DNA. They can generate (nearly) any website you can think of in the span of less than an hour, and deploy it to the world in, maybe, 30 minutes. Modern AI can drive cars across the country. And so the Insiders, who talk about “agent swarms” and think “autoresearch” is last year’s news, are scared about the future. But these people make up less than 0.1% of the global population, and their articles don’t show up in WIRED or the New York Times.

There is a reason these companies are valued in the trillions, and it’s only partly due to the fact that with 80 minutes and a hundred dollars, you can produce a publishable paper in graduate-level mathematics.

But at the Outsiders’ level, AI is the thing that the big bad data centers are built for. “Each ChatGPT query guzzles a bottle of water,” says the mainstream media^[7] to Outsiders:

(This graphic is from The Washington Post. See Footnote 5.)

And now we have social media, mainstream media, politicians, influencers, and normal people all campaigning to ban data centers in a way not dissimilar to the pushes against nuclear energy and supersonic flight. A LaGuardia-to-SFO flight could take two hours; instead, it takes five and a half. Because of poor media and an inadequate, outdated understanding of modern AI’s capabilities, the vast majority of people lean anti-AI. This means they don’t see what’s coming—and being surprised by the future is worse than having time to adapt.

I say that in a very grandiose way, but the reality is that even in AI circles most people don’t know what’s coming. Those who claim they know have either read Asimov, Gibson, or some threads on Twitter. AI models can solve increasingly high percentages of benchmarks that most humans would struggle on already;^[8] we’re just making them smarter and smarter with no clear understanding of what we’re going to do once we’ve maxed out the benchmarks. Do we… make new, harder benchmarks?

It’s this question—what does a slightly higher number on a benchmark really mean, if our models are already good enough, and the main issue is that the public doesn’t want to build data centers or believe in AI—that I gave a little bit of thought to.

I only gave it a little bit of thought because the answer is pretty obviously “A higher number basically means nothing.” Anthropic can be serving us Opus 4.7 instead of Fable 5, and nobody would be the wiser. This tells me that chasing benchmarks will not translate to noticeable, significant, or meaningful changes in the field of AI. Instead, as we’ve seen in the past, real newsworthy changes come from one of three directions in AI development:

It should be obvious why new architectures constitute newsworthy AI research: imagine we invented a successor to reinforcement learning, or world models finally showed signs of being comparable in capability to language models. Every AI enthusiast would know! There is a reason you can say “Attention is…” and most people familiar with AI research will finish the phrase with “All you need”: everyone knows the Attention Is All You Need paper that finally got modern large language models off the ground. This is all a verbose way of saying that if we tried something other than large language models, and it performed better (or had a different set of limitations) than current frontier models, it would be big news.

New capabilities also make for interesting and meaningful progress. The fact that many Slack users can now ping @Claude and have Claude start working on something completely autonomously, or can be pleasantly surprised by Claude reacting to their Slack messages with fun emojis, or interact with Claude as though it were a work-from-home coworker… This is a very, very strange time we live in. But every major step in AI’s form factor and capabilities has been newsworthy and created quality of life improvements across industries. As a short list of noteworthy AI developments that have fallen into this “new capabilities” category (and as a result been widely reported on in tech circles):

Finally, perhaps the least interesting of the directions: speedups and cost reductions. While boring, these make a real difference in the usability and adoption of AI. The Fable 5 session running on my machine right now, which has been running for about an hour, has cost $267.43 in API usage. That is a staggeringly high amount of money. Open-source models, though, simply aren’t feasible to run on consumer hardware like my three-year-old laptop yet. This leaves a huge amount of potential improvement on the table: wouldn’t it be nice if your phone could run the world’s most intelligent technology to give you near-instant answers to any question you have? Faster models would make things like Claude for Fusion 360 actually usable! Honestly, I’m out of examples. I think there’s just so much potential in continued research into speed and cost optimizations.

And perhaps… Perhaps there will always be a need for humans who write article-length thought pieces about AI. With that, I’d like to announce GPT 6.8 Ultra Mega Extra High with 2M Context. It’s even bigger, better, and safer than GPT 6.7.