On June 10, 2024, Apple introduced its vision for a new Siri.
The promise was easy to understand. An assistant would know more about the information on your device, understand what was on your screen, and help you get things done across apps. You could ask for something in ordinary language and spend less time finding, copying, switching, and searching.
For customers waiting for AI to become useful in everyday life, that was a compelling idea. It brought a much larger technology story into something familiar: the phone in your hand.
Then came the wait.
Apple delivered some Apple Intelligence features in October 2024, but the more personal Siri capabilities were still to come. In March 2025, Apple acknowledged that those capabilities were taking longer than expected. For customers who expected those features, the gap between the announcement and the product was disappointing. The frustration was about more than a missing feature. It raised a reasonable question: how much confidence should we place in a future capability when deciding what to buy today?
The dispute eventually reached the courts. Apple agreed to a proposed $250 million settlement over allegations about its marketing of enhanced Siri, while denying wrongdoing. As of this writing, the settlement still awaits final court approval.
That question reaches well beyond Apple.
The AI race was already moving quickly before that June announcement. Companies were developing new models, experimenting with tools, and trying to turn impressive demonstrations into useful products. Siri became a particularly visible example of how difficult that last step could be.
By September 2026, Apple had begun rolling out Siri AI in beta, including personal context, onscreen awareness, and actions across apps. Along the way, Apple and Google announced a partnership to base the next generation of Apple Foundation Models on Google’s Gemini models and cloud technology.
That path involved delays, new partnerships, and a changing approach to delivering the original promise. It is a useful place to begin understanding the wider frontier.
I want to make sense of that frontier with you. Three developments help explain how we got here: what AI can do, how it works through a problem, and what it takes to make the result practical.
Doing
Much of the early excitement around generative AI centered on a conversation. You could ask a question, draft a paragraph, or get help writing code. Using the answer often meant carrying it into another tool yourself.
Tools and automation already existed. What has changed is the growing effort to let AI work more directly inside software.
In October 2024, Anthropic introduced a computer-use capability in public beta. It allowed a model to interpret a screen and use a cursor and keyboard. Other systems have pursued similar goals, moving the interaction closer to carrying out a task across several steps.
That creates useful possibilities. An assistant might gather information from several pages, organize it, and prepare a document for review. It also creates responsibilities. A mistaken answer is one problem; a mistaken action in an account, file, or payment screen can be another.
The meaningful question is how reliably an assistant completes a task, how clearly it shows its work, and when it asks a person to take over. A convincing demonstration gives us a reason to investigate. Everyday usefulness takes more evidence.
Reasoning
As AI began taking on longer tasks, the ability to work through a problem became more important.
In September 2024, OpenAI introduced o1, a series of models designed to spend more time reasoning before responding. This helped bring attention to an approach in which a model uses more computation on a difficult problem rather than always producing the quickest possible answer.
Some AI systems can also use tools to run code, inspect results, and revise an approach. Those abilities depend on the model, the tools it can access, and how the system is built. A longer pause alone tells us very little about whether an answer is correct.
For the reader, the practical interest is straightforward. Can these systems help solve a difficult problem more consistently? Can we check their results? Where do they still make mistakes?
Progress here matters, especially in software and engineering. So does remembering that a carefully explained answer can still be wrong.
Pragmatism
As these capabilities develop, practical questions become harder to put off.
What does it cost to complete a useful task? How long does it take? What information must the system access? How much supervision does it need? How much energy does it consume?
A model that performs well on a benchmark may still be too expensive or unreliable for a particular job. A less ambitious system may be valuable if it does that job consistently and fits into the way people already work.
This is where the race becomes especially interesting. Improvements in capability have to meet the realities of cost, privacy, infrastructure, and trust. There is still room for experimentation, and plenty of uncertainty about which approaches will last.
That is the story I want The Engineering Frontier to follow.
Announcements can help us see what might be possible. Understanding what follows takes patience: looking at what actually ships, examining the evidence, and listening to the people using it.
You do not need a technical background to take part. If you have wondered why a promised feature took so long, whether a new AI tool is genuinely useful, or what any of this means for your work and daily life, you are already asking the right kinds of questions.
My aim is to give those questions useful context, in language we can all follow, without losing the details that matter.
Thank you for being here at the beginning. I look forward to exploring this with you.
A.R.P.