Jan 27, 2025 5 min read

DeepSink, Don't Trust AI, ChatGPT vs. Gemini Users

Dialog: Exchange Number 15

Thinking Out Loud

China's DeepSeek Shocks the Market: Chinese AI startup DeepSeek has shocked Silicon Valley by building LLMs that perform as well or better than the AI giants' at a tiny fraction of the cost, using lower-end chips. The app version of DeepSeek hit number one on the iOS app store, besting ChatGPT. The efficiency and performance of DeepSeek's AI models cast doubt on the efficacy of trying to keep advanced chips away from Chinese companies and raise questions about whether the US can maintain its leadership position in AI. Many companies and investors have sunk billions into the familiar AI platforms, whereas DeepSeek’s R1 model was built on $5.6 million (million) budget and outperformed OpenAI and Meta in complex problem-solving, math and coding. Investors quickly panicked and caused Nvidia to lose roughly $600B (billion), the biggest single day rout of any company in history, as they worry that Nvidia's AI chip monopoly is now over.

Operator – Like Watching Paint Dry: We wrote about OpenAI's Operator last week. Operator is an autonomous agent that can make bookings and buy things behalf of users. Currently, it exists mostly at the bottom of the funnel with a local emphasis: restaurant reservations, travel, sports tickets, grocery shopping, ride sharing, home services. It's currently available to US ChatGPT Pro users ($200/month). Operator complements ChatGPT by focusing on task execution rather than informational queries. OpenAI plans to integrate Operator into ChatGPT and expand access in the future. Using a remote browser it navigates websites, scrolls, clicks and fills out forms just as a human would. While impressive, watching Operator go through its paces to book a travel reservation is like watching paint dry – slow and tedious. Generally, it would be more efficient to make the booking yourself. Operator typically needs to ask clarifying questions and, eventually, permission to book. We expect that it will get faster and be less reliant on the user over time. Together with ChatGPT, it could become quite disruptive of search in certain contexts. But right now it's more of a science project.

Yahoo Search Is Back, Sort Of: In the early days of search, Yahoo and Google were neck and neck. Then Google eclipsed Yahoo in roughly 2006 and never looked back. Yahoo went through a couple of owners and has now stabilized and is rebuilding under PE firm Apollo. In 2010, Yahoo gave up its own search infrastructure and began using Microsoft-Bing. Fast forward to 2023, Yahoo said it was going to "make search cool again." What that means, we guess, is adding AI Chat to search courtesy of Microsoft Copilot. Yahoo's AI Chat relies on Bing but doesn't seem to directly replicate Bing-Copilot results. Users must be signed in to use Yahoo AI Chat and there are caps on the number of "conversations." It's a nice UI and arguably a necessary addition to Yahoo Search. But it doesn't break any new ground and won't gain Yahoo Search any new users, though it may generate additional volume and engagement from existing users.

News & Noteworthy

iOS 18.3 features bug fixes and enables Apple Intelligence by default.
Gartner: AI sliding toward "trough of disillusionment" but spending still high.
Goldman Sachs rolls out AI assistant to 10K employees.
Perplexity revises merger proposal for TikTok to include partial US ownership.
Operator: will websites of the future have an "agentic interface"?
Google investing heavily in NotebookLM (people and money).
Gemini will be able to control smart home apps more intuitively.
Samsung kicks assistant Bixby to the curb, goes all in on Gemini.
Oracle introduces AI agents to help sales teams be more effective.
First autonomous semi-trucks make customer deliveries.

People Don't Trust Companies with AI

In Dialog surveys with consumers, agencies and local businesses, we've found a high degree of enthusiasm and general trust for AI technology. Yet many US adults remain wary of the impact of AI on jobs and society, according to a 2024 Gallup survey of nearly 6,000 Americans. Compared with 2023, their views of AI have improved, with 56% now saying AI will have a "neutral impact" (equal harm and good). The number saying it will do more harm declined by nine points, while those saying it will do more good increased by three points. Overall, 64% of Americans report being "at least somewhat knowledgeable about AI." We've found that as users become more experienced they're less fearful.

Indeed, since 2023 AI skepticism has decreased among Americans over 30; younger users' opinions are mostly unchanged. However, 75% of respondents still believe that AI will have a negative impact on jobs (job displacement). And 77% say they don't trust companies to use AI responsibly. Even those who believe themselves highly knowledgeable about AI (69%) agree. Areas of highest concern to consumers were: AI use in hiring (85%), driving (83%) and medical advice (80%). On a semi-upbeat note, 57% said their concerns could be alleviated by more transparency in how AI is used by companies.

Dialog Data: Gemini vs. ChatGPT Users

Our recent consumer survey of AI and search users continues to produce interesting findings as we drill further into the data. We previously revealed that heavy AI users tend to think of search and AI as more different compared with more casual AI users, who tend to see AI and search and overlapping or interchangeable. There are also interesting differences between ChatGPT and Gemini users. In our overall sample, just under 40% of respondents said their search behavior hadn't changed since they began using AI. But 35% said they now search more, while 26% use Google less. It turns out that ChatGPT-heavy users are less likely to use search, while Gemini users search more often. One could infer that Gemini users see Google and AI as more complementary and ChatGPT users tend to see AI as more of a replacement for Google.