Author: walfresh

How AI and Purpose-Built Models Could Break Through App Fatigue

In the early days of smartphones, apps were a novelty. There was an app for everything, from tracking your steps to turning your phone into a flashlight. However, as hardware became more capable and business models clearer, operating systems have incorporated these functions over time. As the number of home screens on our phones grew from two or three to five or six, users began to favor more streamlined, web-based solutions instead. This phenomenon, known as app fatigue, highlights the limits of the current app ecosystem, where the question, “Why do I need to download an app for this?” often arises. On our computers, where we have access to those web-based solutions, we are bogged down switching and populating information between different systems, and find ourselves asking, “Why can’t this be automated?” With highly capable large language models and a new wave of AI-enabled hardware coming this year, innovation is poised to rejuvenate the app ecosystem.

The first question to always ask: Do I need to build an app?

During Covid, some friends of mine built a successful business and brand for stock tank pools. They made it easy to set up, added a heating/cooling system, a custom top, and built a strong brand that in Texas is now synonymous with circular metal tanks. As they continue to grow (now shipping nationwide), they need to build a more defensible business. Building an app that could help control and maintain the pool, enable future functionality, and build a network for referrals and community could be an option. However, building an application would require them to hire/contract expensive engineering resources, and in this case, the value created likely wouldn’t justify the required investment. The best bet, and what they’ve started doing, is implementing software on the device (in this case, the heating/cooling system) that allows them to interact with smart home hubs like Amazon Echo, Apple HomeKit, and Google Home while also enabling integration with other devices in the future. The other benefits (network, maintenance, community, etc.) can all be achieved through their website.

The technology isn’t what makes you unique, it is the value to users

Another friend of mine started a company five years ago targeting one of the most time-consuming tasks in real estate – generating property descriptions. They built a model that leveraged available structured data combined with unstructured images and geographic variables, outputting descriptions at scale that could be reviewed rather than written. By building in feedback mechanisms into this time-saving process, they increased their confidence while collecting more data from their customers who were getting value from the product. They developed a specialization in understanding real estate documentation and started to build additional workflows. However, as they focused more on the technology differentiation, they became more of a feature and ultimately were acquired by their largest customer. For a startup, this isn’t a bad outcome. They created so much value for a particular segment that one company saw owning the technology outright as a differentiator.

For companies that have an established product and customer base, the challenge is continued innovation, or put another way, avoiding disruption. AI is making it easier than ever to build and iterate on ideas. What took my friend’s real estate technology company months to build just a couple of years ago could likely have a working POC in a matter of weeks.

Good ideas still need data

The ability to process (and availability of) large amounts of data that could train these large language models was the breakthrough that led to the ChatGPT moment we experienced at the end of 2022. While these models boast broad knowledge, they are expensive to run and still cannot be fully relied on to perform tasks. Instead, they provide immense value as assistants. This is where the need arises for specialized models shaped by specific data sets and more efficient to run, becoming “expert” assistants. This opens doors for companies to highlight their unique insights and capabilities within niche domains. The head of AI for a large private equity firm summed this up nicely, saying, “Data is the differentiator, AI is the enabler.”

To be clear, data isn’t your strategic advantage. There is so much data available on the Internet that no matter how good a dataset you have, the quantity of publicly available data makes it challenging to provide differentiated value with a smaller high-quality dataset.

An organization’s data is more valuable to them than anyone else

To highlight this point, there is a reason there is no established marketplace for data. Companies like AWS and HuggingFace do have data exchanges where one can find (and in some cases purchase) a dataset, but it isn’t a huge market. Data is still contextual and without the context of a system (usually the one it is captured in), it is a lot of numbers and letters. In most cases, extensive work is still needed to introduce new data through data cleaning and transformations.

It’s important to understand this fallacy companies (and governments) fall into with their data and the perceived value they believe it has. The way that one differentiates is what they do with their data. You’ve got user data that is likely only valuable to you, but when adding in additional metadata and/or combined with other datasets, it can create a unique and differentiated offering. It’s not just the data; it’s how you use that data. Every company has data that could benefit from AI and provide value to their company. That data itself just isn’t that valuable.

Apps as APIs for (L)LMs (LOCAL Language Models)

Application data (whether that be security, networking, a calendar/task, or messages) is currently siloed within an operating system, and often a lot of the context is lost once it leaves the application it originated in. As the operating systems continue to evolve, LLMs will enable the capability to drive more interoperability. Some examples of what this might look like in current/future state app ecosystems:

• Workflow and record management systems like Salesforce might actually be valuable if an AI assistant could surface relevant information about a record ahead of a meeting I have scheduled in my calendar and update the record in some way post-interaction.

AI at the OS

One thing that has become evident over the past couple of weeks is that AI will most commonly be adopted at the operating system level (rather than the model itself).

• This trend started at NVIDIA’s GTC conference this past March, where CEO Jensen Huang laid out a vision for their integrated management system (NIMS). One large model (NIMS) sits on top and directs/schedules resources to more specialized models underneath it and then determines whether the output/result is satisfactory for the broader request.

• While Meta has a limited operating system, they announced their free “Meta AI” assistant, which has all but replaced search within their applications.

• Google, at their I/O conference, showcased a number of AI-powered features within their Workspace, Android, and Chrome operating systems. By having context of your calendar, your inbox, and the general user interface, you can get suggestions and accelerate your existing workflows to be more efficient.

• Microsoft’s Build conference was all about “Copilot,” which is an operating system that works across an enterprise providing stable APIs for developers, seamless integration into a company’s existing applications and data, and with their Copilot PCs, enablement for users on their local devices.

• Apple Intelligence will focus on specific use cases that leverage their strength as an integrated provider of hardware. On the surface, the capabilities seem to be evolutionary to anticipated actions (like Apple Maps always trying to guess where I’m going when I get in my car based on time of day). But these are functions that will enhance the experience and actually be used. Notably for developers, they will have access to an “App Intents Toolbox,” which will allow them to define actions to the operating system.

All of these announcements diminish the value of the user interface of a specific application (which was already happening with “App Fatigue”) and stress the importance of having differentiated resources.

The race to becoming an AI Aggregator

Ahead of the Apple WWDC, many were expecting OpenAI to essentially replace Siri, but ChatGPT wasn’t even mentioned until the final five minutes of the presentation. When Apple Intelligence can’t fulfill a request, a prompt to ask ChatGPT will pop up. I anticipate other third-party models to become available, giving users a choice of model provider just like they have with browsers. It is easy to see how this could evolve (or revolve?) to applications more broadly.

While Microsoft and Google were originally focused on having the best model, AWS’s strategy was (and has always been) to provide the right tool for the job (specifically for builders). For the majority of use cases, the “best” model isn’t always the most logical. We’ve seen both of the other hyperscalers shift to offering a choice because, at the end of the day, whether you’re using a managed service, a particular LLM, or running your open-source model, all roads will lead back to the cloud. Ultimately, what is important to them is where the data is stored and applications are built. The analogy I often use to explain what AWS does is “building your house on the internet,” and they are the Home Depot/Lowes. A builder can buy a shed or build one, add in all the plumbing or have a specialist do it, and on and on. Hyperscalers like AWS, Microsoft, and Google benefit from the scale of their users and can provide some differentiation in what they offer.

Going back to my hardware store example, when I wanted to purchase a wood-pellet grill, the brand I was familiar with was Traeger, which was only available at Home Depot (like OpenAI is only available through Microsoft), but I ended up purchasing a Pit Boss from Lowe’s (which was closer) that provided the same utility (which would be like Anthropic on AWS).

If it’s all back to cloud and large LLMs, what’s left?

When IBM first invented computers, the prediction was that there would only be five computers in the whole world. What we’ve seen is a wide range of computers, from hyperscaler data centers all the way down to small chips heating and cooling devices for a pool. A big driver of this was the introduction of the Windows operating system that made computers more accessible through a windows, icons, mouse, and pointer interface. Apple came along and made computing even more ubiquitous with pocket-sized devices that minimized the need for a mouse and windows with finger gestures. Google introduced the search bar, which is gradually phasing out icons.

Just like computers, the future of AI will come in all shapes and sizes and different functionality. In another post, I might dive into why AI isn’t going to take over the world, but to keep it brief: Humans are and will always be the agents making decisions. AI and other technology that has come before only accelerate the time to make the next decision. There will always need to be a human in the loop as “the pointer” to the next action needed (even if that is just reviewing/approving suggested actions from an AI). Microsoft held on to their WIMP interface for too long. They saw the mobile revolution as devices that were an extension of your PC that you could take everywhere. Mobile (and search that followed) was much more about the everywhere part than the PC.

It is this everywhere point that is important to keep in mind. There certainly will be a need for the large language models that we see today that know a lot of things about everything, but as we move forward, there is a need for models that know everything about ONE thing. These models will give users the confidence needed to make the next decision in their workflows.

AI can make it possible for an application to do anything, but for organizations looking to build something, it is the “thing” that matters. Rather than thinking about how AI can replace these things, focusing on the human doing the things and how your app can augment will be the key to innovation and sustained success.

So I should build an app?

Rather than asking, “Should I build an app for that?”, some better questions to ask might be:

• What is the decision point I’m trying to accelerate?
• Why would a human want an answer from me and not Google or ChatGPT?
• What data is there (or could be collected) that will differentiate the experience?

If you’re interested in leveraging AI within your organization and want to talk more specifically about what you’re doing, I’d love to help you. Please don’t hesitate to reach out, and we can set up some time to chat.