by Jason Packer with help from Juliana Jackson

“.. the test of a first-rate intelligence is the ability to hold two opposed ideas in the mind at the same time, and still retain the ability to function.” – Scott Fitzgerald, 1936

We’re at a dangerous inflection point in AI, where soon there may be so many articles on this topic that new content may no longer be able to escape the event horizon.

I’m tired of reading about it, and you probably are too. But we’re writing about it again, and you’re reading about it again, so what gives? 

It’s because both of these statements are true:

• There is an ever-growing mountain of absolute garbage written on AI, and most of it should probably be ignored.
• AI is bar none the most important topic to learn about in data and tech, and we’re all still collectively trying to figure things out.

These are conflicting truths. We live in a complicated world full of dueling truths, and this is doubly true in the world of AI. 

To find success in this new world, we first have to learn how to separate fact from hype — and then internalize how facts that seem conflicting may both be true. All while retaining the ability to function.

In an effort to learn how to hold these conflicting truths, let’s walk through some of them.

AI will write our code and do analyses for us.
or
AI produces slop and won’t make our jobs easier.

AI boosters say wild-sounding things like, “AI will write 90% of code in 3-6 months and essentially all code in 12 months.” (Dario Amodei, CEO of Anthropic).

AI skeptics make strident declarations like, “This is, of course, all bullshit. AI isn’t sentient, it’s not going to make your job easier, and AI doctors aren’t going to cure what ails you” (The AI Con by Drs. Bender and Hanna, p. 17).

This is not a case where the “truth lies in the middle”. I’m not even sure what the middle would mean here.

It’s a case where both the skeptics and the boosters have important and truthful things to say, but their statements need a lot of unwinding and hype extraction.

Let’s look at that “90% of code in 3-6 months” quote. Right off the bat, that “90%” number seems like hyperbole from a CEO with something to sell.

But what exactly does “written by AI” even mean? What if I used ChatGPT to create some basic scaffolding code and then modified it by hand from there? What if all I did was use the tab completion in Cursor, was that written by AI? How is that different from tab-completion in my old IDE?

From @vasumanmoza on X

Whatever it means, we need to accept that the majority of developers do now use AI tools in their process. The pure skeptic perspective that AI can’t effectively code no longer tracks here, as amusingly detailed in Thomas Ptacek’s “My AI Skeptic Friends Are All Nuts” article.

Stack Overflow has been polling developers on AI usage in recent years.

The percentage that answered “yes” to this question has skyrocketed:
“Do you currently use AI tools in your development process?”
2025: to be released soon.. I bet 85%, I’ll update this when it comes out. Other polls have said 90%+.
2024: 62%
2023: 44%
2022: the question wasn’t even on the survey

0% to 80% in 3 years is an astonishing rate of increase. While developers do like shiny new tools, they also have a strong preference for sticking with what they know works. For example, in 2024 the 5th most popular IDE was Vim (with 22% of all users), a 33-year-old text editor based upon a 49-year-old text editor. Getting 80% of developers to agree on anything is quite a feat.

When used well, the productivity benefits of AI in software development are undeniable. That doesn’t mean we’re all going to be vibe coding our way to developer nirvana and/or losing our jobs.

AI tools have a lot of limits and need to be watched carefully if they are being used for production code. 90% of code in production coming straight from codegen tools is pure hype (also when someone says, “autonomous agentic coding”, just turn around and walk away). But the changes in the development process are real and shouldn’t be ignored or dismissed!

How does that square with the skeptics saying it will not make my job easier? After a year of integrating AI tools into my development process, I can definitively say that it has made me far more productive in terms of lines of code created and projects completed. 

It hasn’t made my job easier though, it’s just upped expectations about what can be produced. The 2025 HackerRank developer skills survey reported that 67% of developers say AI has increased pressure to deliver faster, and that 84% of engineering leaders say they’ve raised productivity expectations for their teams.

Like many people in the field, I have plenty of angst about being made redundant. It seems inevitable that there will be less of a need for people that write code.

Both statements have a lot of truth in them. AI is already a big part of the software development process. It also produces lots of garbage and will not lead to cushy jobs with lots of leisure time.

AGI is right around the corner
or
LLMs are “Stochastic Parrots” that can never reason.

First let’s be clear, AIs aren’t self-aware in any way. They are nowhere near that point and there is no current path to that. 

But it’s human nature to anthropomorphize, and we can’t help but see these tools as human-like in intelligence. After all, they are designed to mimic us.

Humans use language to express thoughts and ideas. LLMs aren’t communicating in this same way because for an LLM, the words it outputs are just a game of probabilities. It doesn’t ‘know’ what it’s saying, it simply picks the most statistically likely sequence. If you’ve heard this bit before, that’s because it’s true and bears repeating.

We rely on language to build trust. To communicate facts. To navigate reality. None of this applies when you interact with an LLM! You’re not talking to a sentient entity, you’re talking to a figment, an illusory ghost built on a mountain of data in volumes beyond human understanding.

It’s true that there have been many cases of so-called “emergent behavior”, where once the model gets big enough, there seem to be sudden functionality improvements.

This is surprising, but it’s magical thinking to believe that something like AGI would emerge similarly. Emergence so far has been things like “around 13B parameters the LLM seems to get a lot better at two-digit multiplication”, not “around 100T parameters the LLM becomes sentient and exterminates humanity”. It’s also human nature to worry about self-aware killer robots taking over, when the real imminent risk is misuse of AI by humans.

If you’re looking for us to give you a clear answer to the epic argument happening right now about whether or not AIs can reason… I’m sorry to inform you that your expectations of us are way too high. 

The Torch, CC BY 2.0, via Wikimedia Commons. Cropped/pixelated/gloomed.
A non-AI generated Macao riding a bicycle.

A few things we can say:

It’s hard to define what “AGI” or “reasoning” actually is.

Let’s face it, the really hard part is to define what “intelligence” is, artificial or otherwise.

The field of AI has been all over the place in what they consider that “AGI” benchmark to be and when we might reach it.

Coming soon, real soon, since 2019.

This is solid reporting from Sherwood News, but we should consider that almost all of these dudes have a major financial interest in AI. Also how is it they provide a clear date for AGI when they can’t even provide a clear definition of what it is?

The OpenAI charter says AGI is: “highly autonomous systems that outperform humans at most economically valuable work”. This avoids trying to describe “intelligence” and focuses on economic output. This is a strange and fundamentally depressing definition. A Roomba is highly autonomous and more economically valuable than Jason’s music career. It also seems to ignore that a lot of economically valuable work humans do which doesn’t involve sitting at a computer.

“Reasoning” models like o3 and Gemini Pro 2.5 are incredibly powerful tools that can solve problems that require logical steps. 

Their internal workings remain a black box, and what they declare as their reasoning steps might not be what they actually do. It can solve many problems that require reasoning. As Douglas Adams said, “If it looks like a duck, and quacks like a duck, we have at least to consider the possibility that we have a small aquatic bird of the family Anatidae on our hands.”

A previous article by Jason tested different models with increasingly difficult quizzes about Google Analytics. OpenAI’s o3 was beat out all but the most knowledgeable industry experts. An example question on the most challenging quiz was:

A question from Quantable’s expert-level GA4 quiz

Reasoning models o3 and Gemini Pro 2.5 could answer this question correctly, but most of the other LLMs got it wrong, as did 40% of the human test-takers. 

It’s a purposefully tricky question, because many of the answers are accurate, but not relevant to the question. Sending IP addresses to GA4 is forbidden by its terms, but that’s not why the filter breaks.

This question is designed to be challenging for an LLM because it requires multiple logical steps and can’t rely directly on training data (since minimally that IP wouldn’t be associated with this issue explicitly).

To solve this problem, you have to:

1. Know that GA4 can filter out internal traffic by IPs.
2. Recognize this particular IP is non-routable.
3. Know that a non-routable IP wouldn’t typically show up in GA.

It’s impressive, but it’s still more about information retrieval than analytical problem solving. In other words, o3 could help you debug your GA4 implementation, but you can’t just give it access to your data and expect to get “insights”.

Besides being very slow and computationally expensive, the reasoning models tend confabulate more than their simpler counterparts. In other words, advanced large reasoning models actually score worse for hallucination. The Vectara Hallucination Leaderboard rates o3-mini-high second on the list with a .8% hallucination rate, but regular o3 has a terrible 6.8% hallucination rate and is 96th on the list!

So again, both statements have truth in them. With a weak definition of AGI, we may be getting close (though we’d argue not that close). However, the stochastic nature of LLMs may make closing that final distance infeasible with current technology.

AI is an unprecedented tool for learning.
or
AI is undermining critical thinking skills and threatening the education system.

In Juliana’s newsletter, we’ve previously talked about the dangers of AI to critical thinking. When you outsource your intellectual work to the AI, you’re not going to learn. 

If you made a robot to go to the gym and lift weights for you, would you expect your muscles to get stronger? So why would you expect anything different with your brain?

A recent widely covered study from MIT (a good overview from Scientific American here) shows this effect through an EEG watching brain function while writing. Maybe.

So is this your brain on drugs AI? Chat, are we cooked?

An always-relevant 1980s PSA

One less-covered part of the research seems to show that using a chatbot after writing the essay without digital help showed the most brain connectivity. Maybe this shows a way forward where AI doesn’t do the heavy lifting for us, but shows us afterwards how to do it better? Going back to our metaphor about the gym, it’d be like using modern technology to show us how to exercise and recover more efficiently, but making sure we continue to do the sweaty work ourselves.

Sounds great, but what if I don’t want to go to the (intellectual) gym? Hard work is, you know, hard.

Chatbots can be a great tool for motivated learners. At the same time, the temptation to have it just do the work for us is big. This was a bomb dropped on already overloaded teachers with no warning. 

The Richard Feynman quote “names don’t constitute knowledge” is helpful here. AI could get us the species names of the ducks in that video, but what use is the name alone? We’d have to follow-up and ask more questions to actually learn (and retain) something.

AI is the greatest thing since sliced bread.
or
AI is a massively over-hyped bubble.

AI use can feel like magic, and that’s a feeling people are excited to share. Sensing there might be magic in the air, people are hunting up and down for that special combination of prompt, model, and je ne sais quoi that will let them summon the magic for themselves.

I know it’s not magic. You know it’s not magic. But it’s important to remind ourselves of that and remain skeptical. In particular our skepticism should be focused upon those trying to sell us this thing and convince us it is magic, rather than simply a very powerful tool.

OpenAI now has over 500M weekly active users, and has broken all kinds of growth records. The only thing bigger than its meteoric growth might be its own hype. Still, you don’t get 500M weekly active users from hype alone, and you don’t get & keep millions of paying business users unless there is actual utility provided.

Even noted analytics curmudgeon Tim Wilson counts himself as someone who has “had a lot more positive results than negative ones“.

This still feels more like dot-com déjà vu than fire from the gods via Generative Prometheus Transformer.

Valuations have reached stratospheric levels echoing the late ’90s bubble. OpenAI’s valuation soared from $157 billion in October 2024 to around $300 billion after raising $40 billion in March 2025. This was the largest private tech funding round ever recorded. Meanwhile, xAI secured $6 billion at a valuation exceeding $40 billion. OpenAI now trades around 13.5x forward revenue, premium multiples detached from traditional economics.

Startup grifters are everywhere: wrapper companies offering little more than GPT wrapped in slick UIs, and bootcamps peddling prompt-engineering “secrets” you can find on a hundred different blogs for free.

The harm goes well beyond vaporware and overvalued stocks though. The scramble to build and power more data centers is a serious environment problem heaped upon a world already in a climate crisis.

Data centers consumed less than 300 terawatt-hours in 2020 but may approach 1,000 terawatt-hours by 2030, more electricity than Japan consumes annually. In the U.S., data center electricity usage is expected to nearly triple, rising from 4.4% of total consumption in 2024 to around 12% by 2028. Claims like “AI will help solve the climate crisis” ring pretty hollow in the face of hundreds of billions spent on new data centers.

AI tools can absolutely help the climate: better weather modeling of extreme events, smarter energy grids, better sorting of recyclables, automated drones planting trees, etc. But A race of tech giants to outdo one another on chatbot Elo ratings does not get us there. Neither does some kind of hope that an artificial super-intelligence is going to present us with new answers to a problem we already know most of the answers to.

How to deflate the hype.

You’ve heard the quotes. Here’s a few choice ones:

“AI will be bigger than the internet” – Sam Altman (2023)

“In the last 40 years, nothing has been this big. It’s bigger than PC, it’s bigger than mobile, and it’s gonna be bigger than the internet, by far.” – Jensen Huang (2023)

“AI is one of the most important things that humanity is working on. It’s more profound than, I don’t know, electricity or fire.” – Sundar Pichai (2018)

“More popular than Jesus” – John Lennon (1966)

“AI is a much bigger deal than the industrial revolution, electricity, and everything that’s come before.” – Larry Ellison (2025)

Sounds like time to sell your stock in tinder boxes and steam-powered looms to invest in AI!!

Flavor Flav, via ChatGPT and repeatedly instructions for more clock.

First, let’s acknowledge that nobody knows how this all plays out in the future. It’s possible that LLMs, having gobbled up all the available digital data, are near their performance ceiling. Perhaps there are revolutionary leaps to be made via neuro-symbolic AI being worked on in a lab somewhere right now that are going to change everything again? Or maybe both those things are true.

This is a long-term, ongoing debate.

In 1960, shortly after the coining of the term “artificial intelligence”, Science magazine featured the computer scientists Norbert Wiener and Arthur Samuel having this same debate.

Wiener: “…when a machine constructed by us is capable of operating on its incoming data at a pace which we cannot keep, we may not know, until too late, when to turn it off”

Samuel: “A machine is not a genie, it does not work by magic, it does not possess a will, and […] nothing comes out which has not been put in”

Samuel did also add this prophetic addendum:

“An apparent exception to these conclusions might be claimed for projected machines of the so-called ‘neural net’ type… Since the internal connections would be unknown, the precise behavior of the nets would be unpredictable and, therefore, potentially dangerous.”

How to stay functional

As F. Scott Fitzgerald told us at the top of this article, we’ve got to retain the ability to function. There’s no need to read every single new article on AI (though obviously you should read all the ones we write). Switching between model providers every time OpenAI outdoes Anthropic who has outdone Google who has outdone OpenAI… is usually a waste of your time.

Similarly, sticking your head in the sand and believing that this all is going to go the way of Web3 or NFTs is a mistake.

If you’re feeling too much hype, there are some great books out there like “The AI Con” and “Empire of AI” that will help deflate that bubble. If you’re feeling like you want to better understand how AI works and how it can be useful to you, blogs like Simon Willison‘s and Timothy Lee’s Understanding AI both bring a lot of accessible information with minimal hype.

Share this post: on Twitter on Facebook on LinkedIn