ZD 25.48: Happy Birthday ChatGPT
It has been an exciting three years. What's next?
The State of AI in 2025
Three years ago, ChatGPT burst onto the scene. Since then, the rise of Large Language Models has dominated nearly every conversation about technology. Here are five thoughts on the current state of AI, knowing what we know today.
1. Forget the benchmarks
There is a lot of digital ink spilled about which LLM model is the latest in what benchmark. It is largely immaterial. Switching tools to whoever has the best benchmark score this week is a false strategy. You will just confuse yourself. There is always going to be a marginally faster model coming, but you don’t need to pay attention. Focus on tools that connect to your data for the big wins.
2. Connectivity Becomes the Differentiator
As AI becomes commoditized, connectivity starts to become the big differentiator. Technologies like MCP are starting to unlock this. Today, most of your major tools let you connect with your cloud files, your email and other tools. On the other hand, many system of record sorts of tools are developing their own AI integrations. You can start to see the world where your data is easily accessible to AI take shape.
Tools like Notion and Airtable have pretty impressive AI baked into them, letting you use AI alongside your data in real time — as explored further in this piece. The future is now.
3. Google 🚀🚀
Google stumbled hard out of the AI gates. For readers who may not have tracked this closely, Google’s early releases—like Bard—were widely seen as rushed responses to ChatGPT that lacked polish and competitiveness. Their first versions were not good, and they took a long time to get the product together. The giant is now awake and ready to rumble. Gemini 3 is very, very good. Nano Banana’s image creation prowess is unsurpassed; 🍌 has supplanted ChatGPT for this publication’s image needs. NotebookLM has grown from side project to an amazing knowledge tool I could not live without.
Behind the scenes, Google has long built their own hardware, and their homegrown TPU computing units now power much of their AI infrastructure. Recently, they have started looking at selling the devices commercially. Beyond hardware there is a huge data moat. YouTube and Gmail are huge assets. The internet graph is another.
Last and certainly not least, Google can afford to do this out of the massive cashflow generated from their existing business. They do not need anyone’s capital to build it out. They are very well positioned to lead with AI. The $160 share prices we saw not too long ago might have been a generational buy.
4. OpenAI 🤷♂️
Google and the other hyperscale tech companies in AI are largely spending current cashflow to fund future dreams. OpenAI is burning capital at an impressive clip, underscoring their urgent need to monetize more effectively without repeating the same point about scale. They have an impressive number of active users, but not on Google’s or Meta’s scale. Sam Altman is forced to keep pushing the envelope and increasing the leverage because he has no other options. The ChatGPT dream is getting bigger because it has to get bigger or die trying. Currently it seems they are at some sort of American super-app a la WeChat. You can see this starting — group chat and shopping are here, ads are coming. They need to figure out some way to monetize a lot of those users fast.
When I hear people talking about the AI bubble, I wonder if they are really talking about how far over their skis OpenAI might be and using that as a proxy for all of the technology being that far over its skis. Don’t mistake one early adopter’s stumbles for failing technology. My Space did not last either.
5. Limits of LLMs?
It is not entirely impossible that we will soon hit a hard wall in the performance of Large Language Models. Compared with the dramatic capability leaps of 2022 and 2023, today’s advancements feel far more incremental, reinforcing the sense that we may be approaching the limits of this architecture. Recent updates are certainly more incremental. Gary Marcus, who has long been skeptical of the longevity of this technology, is bordering on doing victory laps. If the LLMs do run into a wall, what happens with OpenAI and the other AI-only companies that are largely bets on continually increasing LLM performance?
Year four of the AI boom will likely hinge on how quickly connectivity matures, whether Google’s resurgence continues, and whether LLMs can push past their potential ceiling. Year four of the AI boom will be exciting.
The Distilled Spirit
AI Depreciation Bubbles
AI, depreciation and bubbles were all over the news again.
🌍 AI is Eating the World (Benedict Evans)
Benedict Evans’ annual presentation on technology reminds you that AI is advancing like every wave of tech before it in many ways. The waves are also getting bigger every cycle.
👀 Google > OpenAI + NVidia (Stratechery)
Google strikes back — do they have the wherewithal to penetrate NVIDIA’s moats and OpenAI’s lead?
🌐 Financial and Other Bubbles (
)Why technology bubbles like what might be happening with AI are different than financial bubbles like we had in 2008.
📝 The Depreciation Battleground (
)Michael Burry was a main character in the 2008 crash. He is betting on this bubble, and his thesis is centered on the deprecation of GPUs and datacenters. Does he have a point?
Things to Think About
The most interesting reads of the week.
📰 Read the Damn Article ()
A very eloquent reminder on the value of reading the article, not just the re-written headline.
😀 What Captures Our Attention Now? ()
Media and entertainment are deeply fragmented. Figuring out what captures attention now required a exploration and aggregation across a number of data sets to find a thread that might not be there.
😠 The World’s Polarization Side Hustle (404 Media)
A new feature on X showed that many of the divisive and hateful accounts fanning the flames of social discord hailed from places like Bangladesh, Vietnam, Cambodia and Russia. Most of it is tied to monetization programs where American eyeballs pay the most.
AI Applied
How you can apply AI to your life and work.
🧑🎨 Can AI Replace Powerpoint? ()
Gemini 3.0 and NotebookLM can actually make decent presentations you can really present from. Very close — you can only re-prompt, not edit. But the writing is on the wall.
🤖 Reverse Engineer Yourself to Train Your AI ()
A really cool prompt to examine your own writing, spot your patterns and train a tool to help you compensate for them.
🥇 How to AI Like the Best ()
Using AI effectively requires teaching AI exactly how to do it so you don’t get randomly generated slop.
Gift Guides
I heard people here like gear. We had a great gift guide last week, but we were not alone. Here are a few I liked.
- has great curation and a great and wide-ranging holiday list.
Steven Raichlen writes great BBQ and a wonderful gift guide for the outdoor cook in your life.
- writes about product and AI has a solid list of modern gifts.
The
writes about fashion and has a cool sub-$150 list.- is running a series of gift guides for the soccer fan in your life. They started with stocking stuffers and intend to continue all week.
PI.FYI is an amazing community. They have a crowdsourced gift guide worth exploring.
The Look
does it again in the latest Six Chart Sunday where he shares how to navigate this age of disruption we share.I’ll add subscribing to this publication speaks to the first row explicitly.
Did you enjoy reading this post? Hit the ♥ button above or below because it helps more people discover great Substacks like this one and it helps train your algorithm to get you more posts you like. Please share here or in your networks to help us grow!
Thanks a lot for the mention!
The explosion in capability is crazy. Just take your title image. That couldn't have been done just 1 year ago with the letters and words spelled properly like that. It just couldn't. 3 Years ago it was insanely bad!