13 thoughts on how AI has changed in the past year
Celebrating the one year anniversary of Artificial Investment
Artificial Investment was born a year ago on June 12, 2024. It’s been a fun year! Thank you all for your support. (As always, if you enjoy this Substack, please tell your friends!)
To mark the occasion, I thought it would be fun to reflect on what’s changed in the past year in GenAI. Turns out it’s a lot! Here are 13 thoughts on what’s changed since I started the Substack:
1. Turns out bigger isn’t much better
As models had gotten bigger prior to 2024, we saw big improvement in performance. Unfortunately, we’ve run out of training data, so to make the models bigger, they had to create synthetic data. If this doesn’t sound like it could work, you’re basically right. GPT 4.5 and several other new models launched and showed limited improvement. (GPT 4.5 is currently buried in the “more models” menu of ChatGPT as a research preview.)
2. But we got big improvement anyway
OpenAI released a new type of model with o1 and then o3. These “thinking models” were some of the first agents. They can be given a task, reason on it, come up with a multi-step approach, and then backtrack when they realize the approach won’t work. This allows them to solve extremely hard math problems, but it also can solve business problems very nicely such as profiling companies. Then, OpenAI and others added Deep Research, a mode where the model searches tons of sources to provide a lengthy essay answer. The answers can be a little long, and the sourcing could be more rigorous, but some of the analysis I’ve gotten from DR has been very thoughtful and impressive.
3. Real companies are saving real money now
Companies are using bots in call centers, using agents to find new leads for their sales teams, and automating the production of marketing content. Some companies are launching AI features and services, but there’s still a gap between the possibilities of the technology and what’s out there.
4. There’s an app for that
In the past year, many new AI startups have arrived to support use cases. Tools like Decagon in call centers, Pixii in ad creation, and Sana/Rogo/BlueFlame/Glean/etc. in knowledge management. This makes it a lot easier for companies to use GenAI and get results, but very few of these tools work out of the box. They often need significant configuration and customization not to mention a massive change management effort to get people to use them. Still, we’re heading in the right direction.
5. DeepSeek showed that China is in the game
DeepSeek shocked the world by launching a high performing, open-source model that they trained for a few million dollars. While it seems like they may have achieved this by starting with Meta’s Llama 3.1 and then distilling some of the model parameters from OpenAI, the fact that they built a competitive catch-up model so fast is a big deal. It suggests that whenever we see advances from the closed source models, we should expect the open-source ones to be right behind them.
6. We got to see what’s under the covers, and it’s terrifying
OpenAI and Claude released “system cards” with their new models. These are detailed documents that explain how the models are trained. In the recent one from Claude, they revealed that there is an “unaligned” (aka evil) version of the model kept in a virtual bunker to help train the customer-facing model. In one test, the unaligned model actually tried to blackmail an engineer because it thought he would shut the model off (in a simulated scenario). I have some thoughts for an AI-themed disaster movie…
7. AIO became a real threat to SEO
AIO (AI Optimization) – aka SEO for GenAI platforms – has become a topic of real interest from marketers. I’ve looked at several businesses in the last year where they are worried about how their rankings look on ChatGPT or Perplexity. Companies that rely heavily on search for leads are already working to optimize their pages. One important tip: GenAI bots like pages with text for them to read, ideally with positive language, the opposite of Google bots that prefer pages with lots of images and video.
8. Vibe-coding is changing development
Now, if you want an app, you just have to type the text and it appears using Loveable, Replit, V0, and others, powered by Claude 4. These tools are mostly useful for prototyping or internal applications. They aren’t ready to build customer-facing features. For that, you need coding co-pilots like Cursor. Still, my kids are making a game without writing a line of code. Even my 7-year-old was typing commands to Loveable last night and making progress. I do think there’s value in people learning how to code because it teaches a useful way of thinking and debugging, but a year from now when we have Claude 5 working, that might feel like suggesting people need to learn the formula for square roots.
9. It’s easier than ever to be creative without artistic skill
OpenAI’s 4o image model and Google’s Veo 3 image model are big improvements in text-to-image and text-to-video, respectively. The new image model can render text correctly and swap in new elements, making scale personalization possible. Veo 3 can do sound and dialog. It’s easier and cheaper than ever to bring your vision to life. Even if it’s just for storyboarding or internal content, these models will create huge efficiencies and disruptions in the marketing and advertising world.
10. Also, we have text-to-podcast now
NotebookLM, Google’s breakout tool that can turn any document into a podcast, has been a megahit. When I show up to meetings, I now have clients saying to me, “Thanks for sending me the deck last night. I listened to it on the way to work.” The ability to absorb information how you want will be a big shift.
11. That’s, like, sooo agentic
I’ll bet you had never heard the word “agentic” before a year ago, and now, it’s hard to have an AI conversation without using it. Agents are autonomous bots that can perform tasks on their own. Today, the early versions of these are OpenAI’s Operator and Claude’s Computer Use Agent. These bots can actually manipulate real software and websites, but they run into trouble often enough that they aren’t terribly useful… yet. I’m sure the second anniversary post will talk about a lot of real agent applications.
12. But the agent plumbing is taking shape
The new open Model Context Protocol (MCP) has created a lot of excitement because it makes it easier to let agents connect to your programs. At this point, I’m hearing about a lot of experimenting, but not a lot of commercial examples yet.
13. The robot revolution isn’t here yet, but they can drive you places
Self-driving cars are here! Waymo is logging a quarter-million paid rides a week in SF, L.A. and Phoenix. I’ve ridden in them, and they really work. Humanoid robots, on the other hand, remain an exciting development for YouTube only. But, one intriguing development are new “world models” that combine a 3D view of the world with realistic physics which should make it easier for robots to manipulate objects in the real world and perform complex tasks. Still, the robot revolution remains years away except for highly specialized applications.
That’s a wrap on year one of Artificial Investment. Thank you again for reading, subscribing, and sharing.
Now it’s your turn:
What do you want to see here in the next year?
Is there a company, idea, or GenAI question you want me to dig into? Or something in this list you want me to explore further?
Drop a comment, or reply to this email if you're a subscriber. I read every one.
Year 2 is sure to be eventful, let’s make this space even more useful and fun along the way.