| Hey y'all. As part of our end-of-year newsletters, I've been reflecting on the best AI product design of 2024. But first... Three things you need to know today: • Salesforce says it's winning big pharma clients in a new software competition with Veeva
• TSMC shares hit a record on excitement over AI, with the stock on target for its best year since 1999
• Marvel Rivals is a free-to-play video game hit for NetEase and Disney Putting together a year-end list of the best artificial intelligence products is nigh impossible. Evaluation metrics are bewilderingly abstract and the tech moves at an unreasonable pace. Just look at the big video releases lately: OpenAI rolled out its mind-blowing Sora generator on Dec. 9, and a mere week later, Google announced an upgrade to its rival Veo engine that it claimed is truly the state of the art. Rather than trying to gauge the best AI models, I've been paying attention instead to the best interaction models in AI software. Not long ago, most AI features were driven by prompts typed into a text box. Today, the user interfaces and experiences are far sleeker and more sophisticated, making design a key differentiator in the industry. With each major product launch, you'll hear abstruse engineer jargon — "it's natively multimodal!" — which typically boil downs to how easily and enjoyably you can interact with AI services. When ChatGPT responds on your iPhone, for example, it will reveal its answer word by word with haptic feedback making it feel like it's typing to you. Whereas Google's Gemini shows its replies all at once without any corresponding tactile buzzes. These small flourishes, often overlooked, are going to help distinguish AI apps as their capabilities become commoditized. Below is a list of the AI design trends that caught my attention this year — and the attention of designers I've been chatting with for this newsletter. Invisible assistance: When asked to contribute recommendations to this list, several topflight designers immediately pointed to Granola, a system for automatically converting messy notes and long-winded conference calls into clean summaries and action items. Developers talk a lot about making AI a seamless layer of software, but too often it ends up manifesting as a standalone bot or a bunch of extra buttons. Granola, by contrast, lives in the background. One designer told me it's an exemplar of AI design adjusting to human habits, as opposed to humans having to force themselves to adjust to AI. Player two has entered the game: This fall, Anthropic introduced "computer use," which lets the AI startup's popular Claude models take control of a device to handle humdrum tasks for users. It's aware of what's happening on the screen and can move the cursor and click and type just like we would. It's a fascinating approach to what would normally involve much manual prompt engineering. The Browser Co. and Google, likewise, have been exploring AI agents that surf their respective web browsers for you — imagine asking it, say, to look up and book the cheapest flights to Bermuda and then email the itinerary to family members. This raises fascinating questions about the future of UI design, and whether we should prioritize making it digestible for services that will use AI agents. Real natural language processing: ChatGPT's advanced voice mode is genuinely magical — probably the most stunning AI experience I've had this year. The crazy thing is voice services are nothing new: We've been trying to talk to computers since long before Apple Inc.'s Siri and Amazon.com Inc.'s Alexa became household names. Yet these verbal interactions were incredibly stilted, not so much a fluid conversation, but more a basic (and frequently frustrating) question-and-answer exchange. What's so impressive about OpenAI's dialogue module is that it grasps the incoherence of human speech — the stops and stutters, the tangents, the interruptions — and adjusts naturally and smartly. Memory updated: Give Microsoft Corp. credit for at least trying to build such a risky feature as Recall, a Windows tool for recording screenshots of everything you do on your PC to better inform its Copilot AI. Inevitably, it sparked concerns over privacy and security, and Microsoft is still figuring out how to bring Recall to the masses. But it offered a glimpse of what AI might be capable of if we let it study us intimately. Yes, it's scary to imagine large language models gobbling up our personal data in the same way as social networks and search engines, but there's a very linear relationship between a bot's information and usefulness. There's a world of design possibilities — and unprompted AI assistance — behind the "tell me more" door. Social by design: AI interactions tend to be a one-on-one conversation between a user and chatbot. The makers of workspace app Cove set out to make the experience more collaborative. When a group of people are messing around the same project online, the AI is treated just like another member of the team. Each person's cursor shows up in real time on everyone else's screen, so you can see what they're editing and what they're requesting the AI to help with. No more need to copy and paste from individual ChatGPT threads to a separate message to coworkers; Cove lets you iterate together, with the AI available to everyone simultaneously. Social prompt engineering seems likely to become more popular next year considering the plaudits I've heard about Cove and Anthropic's Artifacts, a similarly slick feature for sharing your AI work instantly with colleagues. Then there's SocialAI, an iPhone app that lets you create a social network entirely of bots with selectable traits (debaters, jokesters, critics, nerds, etc.). It's akin to X and Bluesky, except for whatever you post, you'll get dozens of concise replies from bots instead of humans. One designer told me it's actually a fascinating (albeit super weird) way to chat with a range of AIs at once, for help with ideation and generating unique answers to your questions based on the AI personalities in your feed. Expect these kinds of AI experiences to be everywhere in 2025 — ideally without AI trolls in lieu of human ones. —Austin Carr Attaining a partnership in a venture capital firm used to be a job for life. But a downturn in startup investment, and less independence at bigger funds, has spurred an unusual number of partners to walk away — or be pushed out. Chinese data center operator Yovole is said to be considering an initial public offering on the US markets. Amazon may have another union fight in the new year. Newsweek has created a "fairness" meter as a way of reestablishing trust with its readers. | 
No comments:
Post a Comment