SAN FRANCISCO — These last weeks found me off the grid, cruising the Pacific Coast Highway toward San Francisco. The vistas were awesome: Oregon’s cliffs battered by ocean winds, Northern California’s beaches stretching into myth, and the redwoods towering in regal silence. I couldn't help but take in the beautiful landscapes and, of course, think about AI (and the current state of the world).
Once by the Bay, I took the train over to Berkeley to hang out with my high school buddy Jakob at Arcadia house. The same hacker house I was at last year with my good friend Jack.
If you remember, the frontier in 2024 was agents. So that summer Jack and I worked on building our own browser agent. The air was electric! Using JSON tool calls, our AI could interact with the internet and complete rudimentary tasks. It felt like only a matter of time before even operating systems would be replaced with an AI that controls your entire computer. Still, it was clear a lot more work needed to be done. Vision models were not accurate enough to generate precise UI x,y coordinates for mouse movements. Tool calling was nascent and prompting posed a constant challenge — getting the model to output the right tool call in the correct format.
Fast forward to today, and these problems persist, but at a lower rate. Frameworks like browser-use, launched back in February, abstract away prompting and tool parsing. VLMs, big and small, have gotten much better at pointing. And MCP “solved” tooling.
Although agents are being used now by businesses more than ever, there are still very few consumer use cases. I suppose that's fine as long as they’re working better and making money. But I can’t help and feel underwhelmed by the still significant progress of agents.
The vibes at Arcadia this year feel a little more “whelmed.” Don’t get me wrong, there is still much excitement in the air, but the runaway, exponential, intelligence explosion that was promised is canceled. And in its place, a slower, more gradual, sustainable, rate of improvement has been installed. The scene now feels more research focused, as if everyone realized the low hanging fruit has been picked. Now the real work begins.
This is not the vibe you’d get if you strictly based your perception on media reports. PBS, CNN, and Fortune all wrote about AI being in a bubble. To that extent, I suppose, I’m no different.
But I have a different take.
Nearly all articles on this topic reference MIT’s state of AI in business study (July 2025) and Mckinsey’s state of AI study (June 2025), so let's dive in because these headlines are a bit misleading.
First and foremost, it is true that the MIT study states: “Despite $30–40 billion in enterprise investment into GenAI, this report uncovers a surprising result in that 95% of organizations are getting zero return.”
However, the authors later acquiesce: “While official enterprise initiatives remain stuck on the wrong side of the GenAI Divide, employees are already crossing it through personal AI tools.”
Adding: “While only 40% of companies say they purchased an official LLM subscription, workers from over 90% of the companies we surveyed reported regular use of personal AI tools for work tasks. In fact, almost every single person used an LLM in some form for their work.”
So, employees are using services like ChatGPT or Gemini to augment their work rather than the official enterprise AI initiatives. That seems like enterprise traction, albeit from a more ground-up approach.
Additionally, to collect their data, researchers asked the following question to business leaders: How many GenAI pilots have been launched since Jan 2024?
January 2024 is months before o1-preview (the first reasoning model) hit the scene in September 2024! Imagine coding with AI before reasoning. Obviously some of those pilots failed — the models weren't good enough yet!
Mckinsey’s survey was also conducted before reasoning models (July 16 to July 31 2024) and claims only 20% of enterprises are seeing measurable increased profits as a result of company wide AI projects. (All while 40% of Mckinsey’s revenue comes from “AI consulting.”)
So, no, the sky is not falling. But sentiment around AI is changing. You don't even need to read the studies to feel it; you just have to scroll.
The vibe has gone from “AI will change everything tomorrow” to “AI is everywhere, and I’m not sure I like it.” In the digital trenches of X, people are tired. The novelty of AI-generated everything has worn off, replaced by a sense of being steamrolled by “AI slop.” The backlash isn’t just about quality, though. Users feel like AI is being foisted on them by every online platform, and not in a useful way.
My theory is this fatigue is what's feeding the media’s “AI bubble” narrative. It’s partly a reflection of this, but also a reaction to the classic gap between promise and reality. If there is a bubble, it’s one of expectations. Ultimately, AI’s impact will be measured by the quiet revolutions in how we each work and live. The work continues, the pace is steady, and the cliffs along the Pacific remind me that life and progress is anything but linear.