2025 year-end reflections: We're just entering the golden age of AI
December 28, 2026
I spend my days building and trying AI tools and workflows for myself and for our VC firm; and investing in companies making these tools. And that's where these reflections come from.
I spent a good chunk of 2025 believing the "AI bubble" was going to pop in late-2025 or early-2026. (more precisely I don't believe there's a bubble, yet even a non-bubble can pop).
I no longer think so.
For the reasons I outline below, I think we're just entering the golden age of how AI and AI-powered tools and workflows remake individual and org productivity.
1. Companies are experimenting wildly, and it's all nutrition for markets
I have a hard time believing markets and value will stall when so many teams big, small, and gargantuan are getting so bold in their experimentation, esp. in re-imagining our work surfaces.
Just to list out some personal examples:
- Gamma agent has changed how I build visual materials for meetings
- Perplexity Comet has changed how I use browsers (more on this later)
- Granola has changed how our team records meetings, write, and share results
- Rork is changing how I interact with my phone (more on this later)
- Happenstance is changing how I research people
But we MUST mention Google.
Just take a look at Google Labs with their experiments:
- Pomelli to auto-build entire marketing campaigns
- Disco to reimagine browsers as agents-in-tabs
- Stitch to change the design interface
- and a lot more.
It's bonkers how quickly Google shifted from stagnant to "faster than startups". When the biggest of them all is taking such wild swings, as a user you cannot but rejoice. It's going to create an insane amount of value for all of us.
This is terrifying for startups, but invigorating for the best startups.
A lot of founders will be disheartened and believe that they don't have the manpower, money, and distribution to compete. But the best founders will be energized and see nothing but gold in all the experimentation!
The simple truth is this: not even Google has the resources (and way more importantly, focus) to cover all the new surfaces their experiments are opening up. And these are all surfaces that great founders and product teams can attack - especially when Google has done the hard work of telling billions of people that these surfaces exist!
2. Software will evolve with you
We’re entering an era where software doesn’t just serve you — it evolves with you.
And this is of course because the cost of code has declined to nearly nothing. Why wouldn't you build your own app, and update it as your needs update!?
My prime example is Rork (disclosure: I became an investor in Rork in late 2025, after using it for 1.5 months and being blown away).
I wanted to track things that were impacting my energy, mood, and focus. There are tons of apps out there, but why even spend days to search, try, and compare a bunch of apps, when I could just build what I needed?
So I did this in Rork:
And over the following weeks, I continued to tinker with what to track, how to log, how to visualize data, etc. often by conversing with Rork and other AI.
As I used it, it led to new insights, and I had new kinds of things I wanted to track (e.g. more granularity on food intake). In the past, I would've messaged the app makers, and if I was lucky, there'd be a feature change I wanted. In a few months.
Now? I simply told Rork I wanted to start tracking different things, and Rork promptly tweaked the app to track those things. In five minutes.
Craziness.
I believe this will have immense implications for how we interact with the most important device in our lives.
3. We need opinionated AI, not just opinionated software
A decade ago, Basecamp’s founders talked about “opinionated software” — products that made clear, deliberate choices for the user instead of drowning them in options.
Thanks to the fact that AI can literally create anything now, we're all drowning in AI options.
In the past couple of years, I've been tempted to invest in a number of do-everything software products (other than the foundational LLMs), e.g. tools that allow you to build literally any automated workflow just by describing them.
(I didn't end up investing because I couldn't get into the rounds 🤣).
Looking at those tools now? Literally every single one has struggled to sell — other than I think n8n.
I believe it's the "blank canvas problem" at fault.
And actually it's not even just a blank initial canvas. It's the fact that even if you had an idea and got started, the next step, and the next step… were all blank canvases in themselves.
Because AI allows you to do anything, your follow-on steps also had an infinite number of possible pathways - creating even more opportunities for decision paralysis.
That's why we need opinionated AI.
By opinionated, I mean after absorbing your context and learning of your intentions, AI should be highly opinionated regarding what you should do next, how you should do it - and then it should help you get it done.
This is already starting to happen.
Take Lovable, for instance — after taking your initial prompt, which are often pretty vague, it continues with carefully designed multiple-choice questions, dynamically generated:
Two points worthy of note:
- Clicking on questions & submit is just a lot less effort than having to type in prompts and answers. Even if you are speaking to the computer with something like Wispr Flow.
- Questions are exactly what the best experts would do in a real-life interaction with me.
Btw, when I'm building apps in Rork, the first thing I do is also to make the AI opinionated. I first ask one of the LLMs — could be ChatGPT or Gemini — to simulate an expert relevant to the app I'm building, then I converse with that expert until I have a strong opinion regarding what the app should and should NOT do (the latter is far more important). Then I generate a PRD for Rork to build.
So I believe that in 2026, we'll see even more opinionated AI, especially for verticals, but also for horizontal workflows.
4. The AI Browser: closing the expertise gap
AI browsers have seriously changed my life.
It allows me to go places, learn things, and take actions that I never would've been capable of.
Here's just an example: For one of our firm's investments, I had to look into software for electronics components. I have zero background and technical capabilities for this. Here's what I did:
- I asked Comet to simulate a PCB designer, pick a common product, and to take over my browser, perform components search using different software products in the industry, and report back results.
- Comet reported back the results, complete with play-by-play, screenshots, and full evaluations from the POV of a PCB expert.
- I then tried the same thing with ChatGPT Atlas Browser, and then had Comet and Atlas check each other's work to get to a more realistic result.
- I then took this to real-life PCB design experts to final-check learnings.
None of this would have been possible before the age of AI browsers.
I used the same method to learn how to use database products and tools meant for UI/UX designers. The AI browser would click through all the tools and then teach me step-by-step what each part of the tool was for, why they were there, why they were laid out according to that logic, what each terminology meant, etc.
Btw, with capabilities like this in the hands of anyone who wants it, how could you ever conclude that AI will make the next generation of students dumber and not infinitely smarter?
5. Enterprise adoption will hit an inflection point in 2026
Enterprises have been severe laggards, but I don't think this will be true for much longer.
Btw, this point was inspired by a chat with the awesome LF Chien (簡立峰) who is a famed AI/ML researcher and built Google's largest R&D base in Asia — he told me: "Pay very close attention to what's happening inside enterprises."
So I started to pay attention, and quickly got the feeling that enterprise adoption of AI will soon hit an inflection point.
The first reason is simply the capability of AI agents
Two years ago, when I began to build no-code workflows for our VC firm, I had to do everything manually — connecting APIs, tweaking integrations, mapping data fields, debugging endless logic paths, etc.
Not because I wanted to, but because AI simply couldn't do those things.
Now? AI can handle around 80% of that automatically.
Zapier, Gumloop, Lutra AI, Google Opal (lagging a bit honestly), etc. they can all do ~80% of the wiring for you.
You describe your process in natural language, and it sets up multi-step automations across tools. It still breaks in the final 10–20%, but that’s not the point — it’s close enough that you can feel the future coming.
The second is the sheer number of CEOs posting about real AI transformation
I would say the first two years of this Gen AI wave, most press releases about corporate AI transformation were mostly BS. And we knew it was BS because they didn't contain any level of detail. It was a lot of general descriptions followed by bold claims of performance gains.
I knew they were BS because I spent years writing these press releases for both public (Taiwan's presidential office) and private (public companies) orgs.
Now? I'm seeing posts like this all the time:
Sorry it's in mandarin, but it's about how Dcard - the reddit of Taiwan - reconfigured their whole company including data infrastructure, finance operations, and team org to suit a future of AI agents.
And it goes into extreme detail - a sure sign that it's happening, even if claimed ROI might be overblown.
It's this convergence of capability <> intent that has me optimistic that we'll see enterprise adoption of AI finally catch up a bit to consumer in 2026.
And public markets will be very happy about this.
Looking Ahead to 2026
So I guess that leads me to a few thoughts for 2026.
These are not predictions, I'm not nearly smart enough for those. They're simply where I feel things are headed.
- Google calms down, and focuses on integration: those experiments are awesome, but at some point it's too much, especially when even basic integrations between apps are missing. They've gotta tighten things up for these apps to be truly useful. So I think Google puts the limelight back on integrations in a big way, by making either Gemini the place you get everything done, or (slightly bolder but not really) making NotebookLM the heart of it all.
- We see novel GTM for AI in enterprises: this isn't me, btw. LF Chien alerted me to this during our coffee chat: the top-down method of software adoption isn't working for AI. Hiten Shah's post response to Notion's Ivan Zhao also talks about this - the fact that how AI is spreading throughout orgs and empowering individuals isn't aligned with the way things are. So my bet is that we see new, creative ways for AI to take over enterprises — likely closer to east-west infection rather than north-south decrees, or centered around community activation.
Hiten Shah on Twitter / X
I read Ivan’s post several times.The metaphors are strong. Steel. Steam. Infinite minds.It’s easy to see why it’s spreading. The piece gives language to something many people already feel. Work is accelerating. Familiar rhythms are breaking. Bolting AI onto existing tools is… https://t.co/f2F88NENrp— Hiten Shah (@hnshah) December 23, 2025
- I like to write things in trios, but I really don't have a third one… so I'll just make a wish: Microsoft, you have a product problem. Seriously. You've had one for years, and it's really getting bad. Please just go out and buy the best product teams money can get you, and get back into the race.
Regardless, I truly believe that we're just entering the golden age of AI.