Reactions from ranked influencers
4 posts
Our team has today released a huge analysis of the AI economy as seen through 15 million aggregated and de-identified human-AI interactions. Most comprehensive look at how real people are using AI at scale yet! https://blog.google/innovation-and-ai/technology/research/understanding-the-ai-economy/
@Google @GoogleDeepMind ⚠️ 15 million interactions, not 5!!
A very helpful overview of how AI is being used across Google’s surfaces. It's great to have more insight into AI adoption around the world to complement what we see in the Anthropic Economic Index. This work helps us all make sense of AI’s economic impact.
Very excited about the release of the ATLAS 1.0 white paper on AI use and the economy. The report is a broad collaboration between @Google & @GoogleDeepMind and covers how our AI is being used at scale. The data set has 15 million de-identified interactions across a large number of surfaces including the Gemini App, API, and Google's AI mode. This broad coverage gives us perspective on how AI is used both at work and outside of work. The team did a huge amount of work classifying interactions to specific tasks at work and at home. There are tons of interesting findings in the paper (link below) but here are some of my favorites: 1. Yes, white collar work is overrepresented in work-related AI usage, but it's also being used for a non-white collar work. It's used as a hands on collaborator for diagnostics and troubleshooting. Manual and technical trades tend to use the multimodal features more, images and video, e.g., auto techs and industrial mechanics interpreting complex test results, inspecting machinery for wear, debugging electrical wiring. 2. There is super broad diffusion of AI for almost everything. AI use covers occupations that represent 88% of total US employment, white collar work but also farmers and foresters. 3. AI is creating real value outside of work that GDP statistics may miss. This supports @erikbryn's proposal that GDP will be an incomplete measure of AI's positive impact on society. A huge number of AI interactions are in "productive household activities" such as researching things for the home, interior design, help with tools/appliances etc. Interactions regarding legal services are also highly overrepresented. 4. One of the most overrepresented uses of AI for non-work activities is interacting with the public sector. Activities related to interactions with govt services and civic obligations make up a huge number of interactions, and importantly, many take place outside of business hours. 5. Both #3 and #4 suggest to me that AI will be very useful in helping people navigate complexities and overcoming frictions/barriers in their every day lives. So many conversations about AI focus on labor market impacts, but welfare is not just work, and ATLAS nicely highlights that the broader impact of AI is quite significant. 6. We cover a good deal of global diffusion. English accounts for only 1/3 of global conversations and users engage with AI in their native language for both complex work and non-work tasks. Interestingly, non-OECD countries have a lot more image and video usage. The reason it's called 1.0 is because this is only the beginning. We have a ton of follow up projects in the works. https://ai.google/static/documents/GoogleATLASv1.pdf
The long-awaited Google analysis of usage data is out! I enjoyed the paper for many reasons. Most of all, the top-level conclusions are well-chosen. They crisply articulate and evidence points many of us have been making for a while, arguably more cleanly than their OpenAI/Anthropic counterparts. - Workplace adoption is broad (many occupations) but shallow (few tasks) - AI can be used beyond white-collar work, including in physical labor. We give some nice examples in our recent EconEvals paper from @alexwan55 - Household use is a really big deal, especially for more consumer-oriented Google and OpenAI models. Michael Blank at @StanfordGSB has an excellent paper in this direction that more should read. A more modern ATUS would be even better for this type of analysis! Figure 14, the associated analyses in that section, and the guest comment from Diane Coyle were probably my favorite results in the paper. - GDP per capita predicts Google AI adoption. Based on our national exposure work with @_arulm_, I wonder if national AI exposure and/or white collar shares are better predictors, since this seems to be true of Anthropic/Microsoft/OpenAI data. Internet access as a bottleneck is a nice thing to study, aligning with recent work out of the ILO from Gmyrek et al. Europeans should attend to the issue brought about in tracking the economic impacts of AI (see page 51) "Due to conversation logging limitations, we are unable to explicitly classify API calls from European countries and any paid API usage as work/non-work ... and omit European countries from the analysis due to lack of data." - Appendix B does some nice work on validating conversational classifiers. I appreciate that the authors report the task-level classification accuracy is just 22.58%, of course acknowledging there are 18k tasks in O*NET. This point deserves a lot more focus: it would be see a collective effort to build and use standardized open-source classifiers for economic analysis of usage data. Overall, very happy with this long-awaited first work from Google on analyzing their usage data for economic purposes. Look forward to seeing them sustain and build upon this infrastructure, and working towards something more unified with the other frontier labs!