Is a "recent orders" or "reorder" feature not enough to meet this demand if you frequently order the same thing from the same place? I don't quite understand the need for a middleman LLM here.
When you're asking for suggestion, is that based on a random suggestion to try new dishes or a pre-selected range to stick to what you know and like?
UberEats should have a MCP service that is discoverable by agents - lets call her Siri. I tell Siri what I want, and Siri is checking via that MCP API what restaurants, what type of food, prices, ratings etc. Then Siri is free to generate whatever UI is needed to present the order. This could be generated HTML, but also be read back to me with voice, or even displayed on my Apple TV if I was just watching a series.
I am a professional frontend/UI dev working as a freelancer, and I myself think there will be no more market for creating UIs for the product - the UI is something that will be generated on the fly by the LLM agent for the end-user in what ever way and context the end-user specified. All you need is access to the data and services - MCP just entered the room.
And there, it has a database of sorts, running god knows where on god knows what, and you have a single-user web app that you may or may not ever open and see (since your favorite AI agent can just read and update it for you anyway).
Today it’s still something the user has to suggest / request.
Tomorrow? If they put a bit of effort these models will reach for building these one-off ephemeral or long-lived hyper-custom apps the same way they now reach for a python sandbox for basic arithmetics and statistics.
I wonder how open these services will be to data storage if they do become a "universal app". If I used LLM to make an app for tracking training for endurance running, and I slapped an 800kb .fit file in there each day, the storage for all of these files would probably be fine, but without condensing each run down to a textual summary like Garmin/Strava currently does ("You ran 10km in 35:00, at a pace of 3:30/km. This improved your speed/{vague motivational LLM quote}"), wouldn't the context become enormous after a pretty short duration, causing huge costs? Genuinely curious.
Harnesses have been developed for this, and its also fairly trivial to do large scale verifiable-rewards post-training for "personal assistant" style things. eg., invent a booking website, invent a PA GUI-aware harness, and reward the model when a successful booking takes place. Do that 1bn times.
However, about 1 year ago, AI coding agents were laughed. Now the vast majority of code is written by them.
Intelligence getting exponentially cheaper per token. At some point, there will be an inflection point where "it just works".
I may be completely out of the loop here, so forgive me if I'm missing details, but I always have and continue to believe that the coding prowess of LLM agents more-or-less comes down to the sheer abundance of data it has trained off of on the web, and that code itself is ideal for LLM by the way it is discrete and syntax-bound, same as the way LLM can write very well provided they've read more-or-less the same sequence of text before. For the speed advantages it provides in coding, and for the accuracy it has for well-established information, LLM is a very useful technology.
I just don't follow how they are going to do the same thing with such granular telemetry data, such as datasets from a day of nutrition or a workout. The values are continuous, minute, and are so parallel to noise in the way they are provided. Take running. If you dumped a .fit file from a Garmin or other watch with a series of heart rate values, paces, elevation levels, GPS co-ordinates, respiration rate, vertical oscillation etc, and you trained on a bunch of files like this, the values are so discrete that I can't understand how it will even tokenize it.
I've seen second-hand through friends in running with apps like Runna that it produces plans that aren't suited for them at all. There's simply too many variables. It can give general advice about things, but it doesn't have the level of detail or understanding necessary to give good guidance. Strava "AI" and Garmin "AI" just make a 2-3 line summary from the average data from the entire run or workout.
Perhaps this is something that is already being worked on, and I would be happy if it was, but I just can't personally see how LLM can work with such datasets in a detailed way, beyond the useless summaries Garmin and Strava give "You ran 10km in 35:00, with a pace of 3:30/km. This shows that your speed is improving". I would love an explanation/links on how LLM can get around this if anybody is willing to spend the time providing it. Happy to have my mind changed.
For example, I use LLMs to create 3d models. I don't tell them to recreate the 3d vertexes every time, they have a python script that programmatically creates the shapes, and with a freecad mcp, they get views in images, also they can create other scripts to check for collisions and other issues.
They can also get a histogram, discard noisy values, check outliers, etc, etc, using scripts or by running a traditional ML value to get later the refined data.
Opus 5.5 is quite decent currently, but expensive and not enough for many challenges, but I see the progress happening.
The problem with these Garmin and Strava AI summaries is that most of them just point an LLM at a heap of data and ask them to offer a summary. They also usually use older models since the feature is not something particularly well curated. I do financial modeling with LLMs and use the LLMs to build/verify/iterate on models.
If you're curious on what it takes to run long-running tasks Anthropic has https://www.anthropic.com/research/harness-design-long-runni... out.
Maybe it's a skill issue on my part, but whenever I've tried to use LLM apps over any sort of medium-to-long-term time scale for tracking something (ie. endurance running training), within about a month it begins hallucinating, recalling the wrong things, linking the wrong bits of data to other bits of data, and it does all of this even when pointed to a basic spreadsheet or simple text file to keep it onboard with what happened where. It just seems to look at these files, and then look at previous conversations and confuse itself.
Every other month, some brand new shiny model is released and it's exactly the same deal every time. I think the underlying technology simply can't deal with these sorts of goals. A whole range of apps for lifestyle tracking don't seem under threat, in my opinion, if this is the best its capable of.
Many other things I can't imagine wanting an AI to do for me, especially anything involving payments, bookings.