2025 Predictions

Making predictions is a fool's errand, but here we are.

January 06, 2025

I’m (un)surprisingly bad at predicting tech trends.

Reviewing my predictions last year, the smaller models prediction is obviously wrong near term; models have a minimally useful size and innovation is happening at the periphery. The rest I’m meh about. They feel at best true-ish and mostly ambiguous. I suspect my time horizons are not well calibrated.

There’s no getting good without practice, so let’s give it another go.

  1. Strong differentiation emerges between commercial AI models. It feels like we’re hitting the limits of pre-training. My guess is that while there are experiments on new model architectures, they have yet to find a generational leap in performance1 and will not see any breakthroughs there. Test-time compute however, creates space for new approaches and techniques. The major AI labs will develop their own methodologies (i.e., differentiated from O3) to coax models into distinct performance characteristics. As a result, we will see stronger differentiation between the models shipped by the major vendors.

  2. We have reached peak RAG and copilot. The most common shape of AI product today is a thin wrapper around AI APIs coupled with access to a proprietary data source. Over the last 2 years, we have seen the limitations in these types of deployments. I predict that momentum and growth of these apps will slow in the coming year and this generation of features will start getting deprecated in favor of new emerging paradigms. RAG remains useful but diminishes in importance and becomes a commodity technology layer.

  3. ChatGPT becomes the juggernaut in D2C AI products. I’m stating the obvious at this point, but let me make a stronger case. ChatGPT becomes the one AI Chatbot to rule them all. Competing efforts from Anthropic and Perplexity2 lose steam and market share. The compelling mix of OpenAI innovating on the model layer plus its relentless focus on the ChatGPT product makes it difficult for others to compete without significant outspend.

  4. Apple Intelligence does not live up to the hype. Following the hyped WWDC presentation, most of what’s being shipped up to 18.2 feels like a dud. I don’t think it will get better this year. My guess is this is due to a combination of Apple being constrained by having to run a smaller, weaker model on local hardware and the baggage of Siri. AI is also in the hyper growth phase right now and we users are quick to acclimate to new improvements, and Apple, lacking the internal expertise will have to aggressively play catch up to avoid falling behind. It’s a pity, because I still think the AirPods is the most potent AI UI with a build in captive audience but it feels squandered right now.

  5. There will be signs of Clay Christensen disruption in Enterprise software. This is how it will go: a Small medium enterprise company decides to forgo a SAP installation in favor of a custom AI build on top of conventional databases. AI software today has a quality that makes it orthogonal to enterprise software. AI is great at being flexible, while enterprise software like ERP is hard and rigid like a rock. The potent mix of getting into the new new while at the same time saving money by not using luxury vendors will be attractive to shrewd and discerning bosses.

  6. Augmented Reality will take a step forward but fail to reach mass-market. Even though Apple Vision Pro has faltered, new AR devices or AR-adjacent devices like Smart glasses will continue to debut this year. One major trend will be the integration of voice AI as a primary user interface. It will be impressive and reviewers will swoon, but the lack of immediate use cases3 mean that these devices remain in the niche of technology enthusiasts.

  7. There will be a new UI design paradigm for AI. The Chat UI is getting old, and it feels as if the air is pregnant with the potential for disruption. My prediction is that someone will come up with a compelling demo this year and it will get broadly copied and mimicked. It may not be as compelling as the introduction of ChatGPT, but it will represent a page turn on how we think about the way we interact with AI.

There you have it. We will revisit this next year and see how I fare.

  1. We might see architecture swaps that provide better inference time efficiency but they will perform at best equal if not strictly worse than transformers. ↩

  2. I wonder if Perplexity is monetizing too quickly without figuring out its user moat. ↩

  3. AR Glasses as a category really needs an iPhone “phone, Safari, iPod” moment. ↩