Developers are looking for creative ways to create AI-powered chatbot assistants
Since OpenAI introduced ChatGPT last year, it’s been nearly impossible to go a day without a company or developer releasing an AI-powered tool. Now, with the release of new models like OpenAI’s GPT-4 and Anthroipic’s Claude, more app builders are testing tools that are more accessible and useful to people.
Some apps provide a native desktop or mobile experience that allows people to talk to AI-powered chatbots outside of the web browser. Their core idea is to earn money by unlocking unlimited access to these bots and providing some quick ideas to users.
However, some developers want to go a step further and are working to better integrate their applications with the system. There have already been some attempts to make these apps compatible with Siri via shortcuts. So users can ask questions to ChatGPT (or any other model) through voice or even text. This gives them an advantage in cases where Siri can’t understand a user’s query, they can simply use the keyboard to type a question for the AI-powered bot.
An example is Short Circuit, an application developed by joe fabisevich, a former Twitter employee turned independent developer. The app allows you to chat with a bot called Shorty. It comes with suggested prompts for workout plans, meal plans, writing fun songs, and discovering fun facts on a topic. It also allows you to use the “Hey Siri, Hey Shorty” command to ask questions through your voice.
Fabisevich said the tools helped him write an app store description, brainstorm app store optimization tips, and write app purchase code. He added that the development team behind Short Circuit is exploring ways to integrate the app with all kinds of automation using Shortcuts.
The folks at MacStories have gone a step further to develop a shortcut called S-GPT, which is built into many parts of macOS and iOS.
The tool can summarize a web page shared via Safari’s share sheet; it can help with time management to understand which days you have a busy schedule; check the text on your clipboard for grammatical errors; and offers to open links from ChatGPT’s response in multiple Safari tabs.
But the feature that stands out the most for its fun factor is that you can ask S-GPT to create a playlist based on a prompt. For example, when you ask “Make me a playlist of 15 rock songs from the early 2000s,” the playlist will be saved to Apple Music. Tools like PlaylistAI have released features like prompt-based playlist generation for Spotify. But the S-GPT effort feels more integrated since it uses Apple Music.
Federico Viticci of MacStories mentions that you can also enter complex queries like “Make me a playlist of 25 soft indie rock songs released between 2000 and 2010 and sort them by year of release, oldest to newest.”
The developers also claim that S-GPT is better than Siri when it comes to back and forth conversations. This is because the tool prompts you to ask follow-up questions if you wish. Having a visual prompt helps in some cases, as Siri often forgets the context of the conversation.
The developers of S-GPT and Short Circuit mention that Siri often only reads long text from a source like Wikipedia, which could be tedious. They argue that it is easier to read slightly longer text on the screen.
GPT-powered templates are more useful than Siri in cases where they need to generate text, summarize text, and present more information on the subject from multiple sources. Fabisevich said by email that the best advantage of these great language models is how they reduce the “cycle of going through multiple Google searches.”
He said Short Circuit users have found use cases ranging from meal planning to finding more information about a bird while birding, coding issues, and Dungeons & Dragons story generation.
However, when it comes to figuring out facts like tomorrow’s weather or the score of the last game, Siri (or any equivalent assistant) is still superior. Fabisevich said that’s why there’s a fact check button in the app, which takes you to Google’s search results page with the answer as the query.
“I still don’t trust the results that ChatGPT gives me and I find myself verifying information that seems counterintuitive or suspicious. Sometimes my intuition is wrong and ChatGPT is right, but this skepticism is what led me to create a fact check feature in Short circuit. Although GPT-4 is better with hallucination, I still think hallucination will be a big problem for large language models,” he said.
Even Microsoft’s Bing and Google’s Bard often fail when asked about current or historical events. So in that sense, current generation AI assistants on phones are less likely to spread misinformation, as they would simply direct you to a web search if they don’t know a topic. AI chatbots also lack speed, as they need to query a server (such as OpenAI) to get answers to a query.
This first wave of GPT-powered bots isn’t exactly trying to replace Siri. But they are trying to make life easier when it comes to asking for suggestions or ideas. In addition to those mentioned, tools such as AnyGPT and MacGPT make ChatGPT easy to access on Mac, but don’t offer things like voice command integration.
With both Google and Apple hosting their annual developer conferences in the coming months, it will be interesting to see how they update their Assistants. Various teams at Apple, including the one that runs Siri, have reportedly been experimenting with large language models.
At the same time, developers hope to gain access to more system-level automation to make use of these large language models. A recent analysis from various companies suggests that both downloads and consumer spending on AI-powered apps have skyrocketed in recent months. So developers would want to take advantage of this wave of generative AI.