Happy Friday. I\u2019m back from vacation and still getting caught up on everything I missed. AI researchers moving jobs is getting covered like NBA trades now, apparently.<\/p>\n
Before I get into this week\u2019s issue, I want to make sure you check out my interview with Perplexity CEO Aravind Srinivas<\/strong> on Decoder this week<\/a>. It\u2019s a good deep dive on the main topic of today\u2019s newsletter. Keep reading for a scoop on Substack and more from this week in AI news.<\/p>\n From chatbots to browsers<\/strong><\/p>\n So far, when most people think of the modern AI boom, they think of a chatbot like ChatGPT. Now, it\u2019s becoming increasingly clear that the web browser is where the next phase of AI is taking shape.<\/p>\n The reason is simple: the chatbots of today don\u2019t have access to your online life like your browser does. That level of context \u2014 read and write access to your email, your bank account, etc. \u2014 is required if AI is going to become a tool that actually goes off and does things for you.<\/p>\n Two recent product releases point to this trend. The first is OpenAI\u2019s ChatGPT Agent<\/a>, which uses a basic browser to surf the web on your behalf. The second is Comet, a desktop browser from Perplexity<\/a> that takes it a step further by allowing large language models to access logged-in sites and complete tasks on your behalf. (OpenAI is rumored<\/a> to be planning its own full-fledged browser.)<\/p>\n Neither ChatGPT Agent nor Comet works reliably at the moment, and access to both is currently gated to expensive subscription tiers due to the higher compute costs required to run the reasoning models they necessitate. Perhaps most frustratingly, both products claim to do things they can\u2019t, not just in marketing materials, but in the actual product experience.<\/p>\n ChatGPT Agent is a read-only browser experience \u2014 it can\u2019t access a logged-in site like Comet \u2014 and that severely limits its usefulness. It\u2019s also very slow. My colleague Hayden Field <\/strong>asked it to find a particular kind of lamp on Etsy<\/a>, and ChatGPT Agent took 50 minutes to come back with a response. It also failed to add items to her Etsy cart, despite claiming it had done so.<\/p>\n While Comet is nowhere near as slow, I\u2019ve had numerous experiences with it claiming it has completed tasks it hasn\u2019t, or stating it can do something, only to immediately tell me it can\u2019t after I make a request. Its sidecar interface, which places the AI assistant to the right of a webpage, is excellent for read-only tasks, such as summarizing a webpage or researching something specific I\u2019m looking at. But as I told Perplexity CEO Aravind Srinivas<\/strong> on Decoder this week<\/a>, the overall experience feels quite brittle.<\/p>\n It\u2019s easy to be a cynic and think the current state of products like Comet is the best AI can do at completing tasks on the web. Or, you can look at the last few years of progress in the industry and make the bet that the same trend line will continue.<\/p>\n During our chat this week, Srinivas told me he\u2019s \u201cbetting on progress in reasoning models to get us there.\u201d OpenAI built a custom reasoning model specifically for ChatGPT Agent that was trained on more complex, multi-step tasks. (The model has no public name and isn\u2019t available via an API.)<\/p>\n Even with the many limitations and bugs that exist today, using Comet for just a few days has convinced me that the mainstream chatbot interface will merge with the browser. It already feels like taking a step back to merely prompt a chatbot versus interacting with a ChatGPT-like experience that can see whatever website I\u2019m looking at. Standalone chatbots certainly aren\u2019t going away, especially on smartphones, but the browser is what will unlock AI that actually feels like an agent.<\/p>\n Some noteworthy career moves<\/p>\n If you haven\u2019t already, don\u2019t forget to subscribe to The Verge<\/a>, which includes unlimited access to Command Line<\/a> and all of our reporting.<\/p>\n As always, I welcome your feedback, especially if you have thoughts on this issue or a story idea to share. You can respond here or ping me securely on Signal<\/a>.<\/p>\n\n
\n