This is my second yearly list of AI predictions. My first AI predictions from December 2023 are available here.

AI APIs for everything by 2026

APIs allow services to talk to each other. The late 2000s and early 2010s were the golden era of public APIs. Facebook, Google, and Twitter, etc offered API access to their platforms. Developers created loads of useful apps that improved the user’s experience.

Unified messaging apps like Pidgin meant that an individual could chat with friends from one single app no matter what platform they were on. Open API access enabled many other improved user experiences and companies like IFTTT made it easy for humans to connect things in novel ways. People could easily set up an automation that turned off a phone alarm and messaged the local parent group chat on WhatsApp when the local school district’s website contained the words “snow day” at 06:00am.

However, tech companies started locking down their APIs in an effort to control the user’s attention. Many tech companies introduced chat features for their products to lock the conversations to their platforms. Nowadays humans require many apps and platforms for the simple task of communicating in text with the occasional photo or video.

With AI, everything can be an API. Imagine the freedom of using a single app to send a group message to 20 humans and having an AI deal with the how. Some of those humans may only be accessible by SMS, WhatsApp, LinkedIn etc. An AI can navigate apps like WhatsApp and websites like LinkedIn, removing the need for those platforms to provide public APIs.

Given that AI can make phone calls, even offline services can be APIs.

Funny and obnoxious “are you human?” tests by 2025-06

“Jagged frontier” is a term that describes the (from a human perspective) uneven advances in AI. For instance, consider OpenAI’s gpt-4o model which was unable to correctly answer the “how many ‘r’ in ‘strawberry’” prompt even when challenged multiple times. Meanwhile, gpt-4o scored in the 90th percentile on the Uniform Bar Exam.

“Are you human?” tests need to be beyond the capabilities of AI but easy for humans. Fewer and fewer skills will meet this criterion over time. Tech companies are incentivised to find these skills to protect their platforms from AI users.

This will lead to some funny and obnoxious “Are you human?” tests in the short term. Be prepared to answer some strange questions before you can post your cat meme. Eventually, the platforms will have to pull back their “Are you human?” tests else they risk alienating the cash cows that are the actual humans using their platforms.

SaaS business models will struggle by 2027

Many Software as a Service (SaaS) companies have outsourced their server requirements to the cloud. They act as a solutions layer between customers who can move and cloud hardware which they don’t own. This is a sensible business model given that software is still relatively expensive to produce and maintain. However, AI will make software basically free very soon. It will become cheaper and easier for many current SaaS customers to migrate to their own software and access cloud providers directly.

Many SaaS customers may be hesitant to move to their own software. However, incumbent SaaS providers will also struggle to compete against the many new alternative AI-generated competition. Especially given how easy migration between SaaS platforms will become thanks to AI APIs.

It will be tough for employees of SaaS companies when software becomes free.

Novel creative approaches by 2026

Current AI models are great at writing generic boilerplate code. They are not currently good at coming up with novel solutions to coding problems. This is leading to a lot of inefficient and hard-to-maintain coding patterns being regurgitated by AI and introduced into codebases by eager yet inexperienced developers. However, AI frameworks are capable of generating non-derivative content.

In a recent hackdays project, we used techniques to extract comedic threads from a subject which were then used as context for non-derivative jokes. Similar patterns could be employed to generate efficient and novel coding solutions which could, in turn, be used directly or as training data for the next generation of models.

Open source o3 performance by 2026-07

Open source offerings will match o3 performance in common benchmarks by mid-2026. A model-agnostic framework could utilise the best open source models and experts for each specific task. Model specialisation allows for better utilisation of resources and this technique is likely already being employed by OpenAI for o3.