How ChatGPT Pulse Could Change AI

0:00Today on the AI Daily Brief, ChatGPT pulse and the rise of background agents. Before that of the headlines, a new benchmark that focuses on real-world tasks. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI.

0:17All right, friends, quick notes before we dive in. First of all, thank you to today's sponsors, Blitzy, Agency.org, Robots & Pencils, and Notion. And to get an ad-free version of the show, go to patreon.com slash aidailybrief. Welcome back to the AI Daily Brief Headlines Edition, all the daily AI news you need in around five minutes. Man, some days things are just happening, and this is one of those days. We have so many announcements. We have, in fact, companies competing with themselves and their own announcements. Our entire main episode is about ChatGPT Pulse, which is this new background agent that suggests the new paradigm of ambient AI and the future of AI personal assistance.

0:54And yet for me personally, that's not even the most exciting open AI announcement today. Instead, that title belongs to GDPVal, a new evaluation that they say measures AI on real world economically valuable tasks. Evals ground progress, they write, in evidence instead of speculation and help track how AI improves the kind of work that matters most. Now, this is something I've been talking about a lot recently. You might have heard me talk about the idea of needing a utility score or a different way of looking at what a new model means for the unlocking of new use cases. I was thinking about this first in the context of Nano Banana, where I saw all of these people saying, yeah, but it's just better at editing in one part for another part, as though that somehow undermined its novelty because it wasn't better at things like native base generation.

1:41But it turns out that the things that it was better at unlocked a huge amount of very commercially valuable use cases. So you have that phenomenon, plus the fact that our existing benchmarks are so washed, and it's clear that we've needed something else. Well, OpenAI is stepping up to the plate on this one with their introduction of GDPVal, which measures model performance on economically valuable real-world tasks, specifically benchmarked to 44 different occupations. They say that it's called GDPVal because, quote, we started with the concept of gross domestic product as a key economic indicator and drew tasks from the key occupations in the industries that contribute most to GDP.

2:18In their announcement post, they go big. They write, people often speculate about AI's broader impact on society. but the clearest way to understand its potential is by looking at what models are already capable of doing. History shows that major technologies, from the internet to smartphones, took more than a decade to go from invention to widespread adoption. Evaluations like GDPVAL help ground conversations about future AI improvements in evidence rather than guesswork and can help us track model improvement over time. So what is this measure? Well, like I mentioned, it spans 44 occupations across the top nine industries that contribute to US GDP.

2:50The full GDPVAL set includes 1 ,320 specialized tasks, each of which was crafted and vetted by professionals with at least 14 years of experience from each of these fields. Every task, they say, is based on real work products such as a legal brief, an engineering blueprint, a customer support conversation, or a nursing care plan. Unlike benchmarks, they write, which involve synthetically creating tasks in the style of an academic exam, GDPValve focuses on tasks based on deliverables that are an actual piece of work that exists today. They also write that GDPVAL tasks are not simple text prompts.

3:24They instead come with reference files and context, and the expected deliverables include documents, slides, diagrams, spreadsheets, and multimedia. They do say that it still doesn't reflect the full nuance of many tasks, but obviously this gets us much closer in that direction. So what are some of the tasks? Think of an example from a producer, an order clerk, and a manufacturing engineer. For the producer, the prompt and task context is, You're a video producer for an advertising agency prepared to onboard a new project, a 60-second live-action B2B photo shoot. The client has set up a kickoff call for this project on Monday, July 7th, 2025, and set a deadline for final delivery of the video on Friday, August 29th, 2025.

4:02Then there's a whole bunch of additional context, like a huge, huge amount of it, and the so-called Experience Human Deliverable was a plan to get all the work done in a calendar that lays it all out. Now, you might be asking how they grade this, given how comprehensive it is. well, they are going old school with this. Basically, they rely on expert graders, professionals from the same occupations that are represented in the dataset. The graders are asked to blindly compare model-generated deliverables with those produced by task writers without knowing which is AI versus human. Graders offer critiques and rankings, ranking the human and AI deliverables, and classifying each as better, as good as, or worse than one another.

4:39In addition to that, they also created detailed scoring rubrics for their occupations, adding a layer of consistency and transparency. Finally, they also built an automated grader on top of that, which is an AI system that's trained to estimate how human experts would judge a given deliverable. They're releasing that tool right now at evals.openai.com, but they say that it's not yet as reliable as the expert graders, and so are not replacing them with it. A couple really interesting takeaways. First of all, the models are winning or tying industry expert performance at a pace of about a quarter to a half the time.

5:11They find that there is clear linear progress up, with performance more than doubling from GPT-4.0, which was released back in spring of 24, to GPT-5, which was released this summer. Interestingly, and with big kudos to OpenAI for still releasing this, they found Claude Opus 4.1 as the most performant model, meaningfully above even GPT-5 high. Now for OpenAI, this is very clearly just a beginning. They're already identifying areas where there is work to be done. They want to expand it to more occupations and more tasks. They also note that the current version is one-shot, so it doesn't capture cases where a model would need to build context or improve over multiple drafts.

5:48They also note that in the real world, tasks aren't always clearly defined, so there is a ton of work to be done, but man, is this such a productive and powerful area to be working on. I may have been overstating it slightly when I said that this is the AI benchmark that we've all been waiting for, but it is certainly the AI benchmark that I personally have been waiting for. I could go on and just make this just two full mains stacked together. But I did also want to briefly touch on the announcement of a new feature for Meta, which they are calling Vibes. It's a new dedicated feed in the Meta AI app that is exclusively short-form AI-generated videos.

6:22This is the first output of their collaboration with MidJourney and Black Forest Labs. And Nick St. Pierre writes, think reals but AI and interactive. You can create and remix vids, make image and video edits, relight, restyle, add music, and publish. Meta, for their part, calls it a new way to discover and create AI videos. and say that it's designed to make it easier to find creative inspiration and experiment with Meta's AI media tools. Now, if you had to guess, what would you think the response to this was going to be? Maybe we get a little bit of creative enthusiasm, excitement to see all these generations in one spot.

6:55Maybe we get a little bit of skepticism, given that it's Meta, and their job is to keep our attention firmly locked on their apps. Maybe we get literally nothing but scorn in utter derision. And if your guess was that last one, you are by far the most correct here. Railway founder Jake writes, what we do for money, what we choose to build. These say a lot about who we are. I work hard every day because this is not the future I would like to live in. Sam McAllister said, excited to share slop, a new slop trough for you all to enjoy. Bazelord had perhaps the most crisp response, resharing Alexander Wang's post with the simple caption, mega F this.

7:33Bloomberg's Odd Lots podcast host Joe Weisenthal writes, everyone else has already weighed in, but I too feel compelled to say that this all looks like pure garbage. Dean Ball writes, the meta algorithmic video slop in this post is from a division of meta called superintelligence. In what sense is this superintelligence? In what sense is this on the critical path to superintelligence? Shame is underrated and this is shameful. Ruxandra Teslow writes, I've exited my naive techno-optimism era, and I will unabashedly state we should shame this stuff. That's not how we should use AI. To further fry our brains and turn us into walking zombies, we should build real things and avoid these cursed use cases from the 2010s era.

8:11Matthew Iglesias writes, is he really excited to share this? To me, that's the truly scary possibility. Is there a company full of smart, talented, hardworking people who are excited to bring the world more short-form video? Now, even I, who has a sense for how people react to most meta announcements, didn't quite expect the utter vitriol that this thing is getting. Here's my take on it. One, I do think that there was always going to be this question, as Zuckerberg was out recruiting a superintelligence team, around what they were going to use all of that brain and computing power for. It seemed like behind the scenes, a lot of the reasons that people were saying no is that they assumed that they would just be forced into having to use AI to keep people's tension firmly fixed in the meta family of apps rather than trying to do something actually meaningful, at least in the way that they define meaning.

8:56The flip side is all of the distribution platforms are going to have to reconcile and deal with AI content in some way. As the cost of AI content production comes down dramatically and it becomes easier than ever to produce video content, it is just going to absolutely flood the channels, meaning that the power of the discovery algorithms gets even more powerful. I tend to think that consumer demand might force, at least in the short term, some sort of explicit divide between AI and non-AI content, although ultimately you got to think that it's likely to blend. So I guess on that front, at least having it segregated in a different experimental part of the app means that people can avoid it if they don't want it.

9:35The other thing is, it feels completely inevitable to me that this new set of creative tools will generate some new social platform. It's kind of the way that it always works. It is very rare that the platform from the last era of technology is the platform for the new era when it comes to social content, although I'm not sure yet that just Instagram Reels or TikTok, but for AI video is the final form. Entrepreneur Eugenia Kudia writes, Yes, the next huge consumer AI winner will be a social platform. No, it will not be this. As if to make the point that everyone is going to have to deal with this, we also got news that Spotify was taking a bunch of actions to tamp down on AI on its platforms as well.

10:13In a post on its website on Thursday, Spotify said that it had removed 75 million what it called spammy tracks. They wrote, the pace of recent advances in generative AI technology has felt quick and at times unsettling, especially for creatives. At its best, AI is unlocking incredible new ways for artists to create music and for listeners to discover it. At its worst, AI can be used by bad actors and content farms to confuse or deceive listeners, pushing slop into the ecosystem and interfere with authentic artists working to build their careers. That kind of harmful AI content degrades the user experience for listeners and often attempts to divert royalties to bad actors.

10:48Now, part of what's going on here is that this is not Spotify saying that people can't experiment with AI music, but people right now are using a flood of AI to try to game the system and capture ad revenue. Spotify has introduced new policies around impersonation, AI voice clones, and are exploring a number of other protection tactics as well. I think it's an interesting first step, but I continue to believe that we are likely to see, again, at least in the short term, entirely segregated experiences for AI-generated music. And certainly you got to think that these companies are all going to be watching the responses to things like MetaVibes and these announcements from Spotify to try to understand what consumer pressure is going to push them to do.

11:24In any case, it continues to be a fascinating time. That's going to do it for today's Not So Brief Headlines. Next up, the main episode. This episode is brought to you by Blitzy, the enterprise autonomous software development platform with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Enterprise engineering leaders start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and precompiles code for each task.

11:55Blitzy delivers 80 % plus of the development work autonomously while providing a guide for the final 20 % of human development work required to complete the sprint. Public companies are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Blitzy is providing a limited-time, 30-day free proof-of-concept for qualifying enterprises. The team will provide a 5x velocity increase on a real development project in your org. Visit blitzy.com and press book demo to learn how Blitzy transforms your SDLC from AI-assisted to AI-native.

12:31That's blitzy.com. Shape the future of enterprise AI with Agency. A-G-N-T-C-Y. Now an open source Linux foundation project, Agency is leading the way in establishing trusted identity and access management for the Internet of Agents. The collaboration layer that ensures AI agents can securely discover, connect, and work across any framework. With Agency, your organization gains open, standardized tools and seamless integration, including robust identity management to be able to identify, authenticate, and interact across any platform. Empowering you to deploy multi-agent systems with confidence, join industry leaders like Cisco, Dell Technologies, Google Cloud, Oracle, Red Hat, and 75-plus supporting companies to set the standard for secure, scalable AI infrastructure.

13:14Is your enterprise ready for the future of agentic AI? Visit agency.org to explore use cases now. That's A-G-N-T-C-Y dot org. Today's episode is brought to you by Robots & Pencils. When competitive advantage lasts mere moments, speed to value wins the AI race. While big consultancies bury progress under layers of process, Robots & Pencils builds impact at AI speed. They partner with clients to enhance human potential through AI, modernizing apps, strengthening data pipelines, and accelerating cloud transformation. With AWS-certified teams across US, Canada, Europe, and Latin America, clients get local expertise and global scale.

13:51And with a laser focus on real outcomes, their solutions help organizers work smarter and serve customers better. They're your nimble, high-service alternative to big integrators. Turn your AI vision into value fast. Stay ahead with a partner built for progress. Partner with Robots and Pencils at robotsandpencils.com slash AI Daily Brief. Chatbots are great, but they can only take you so far. I've recently been testing Notion's new AI agents, and they are a very different type of experience. These are agents that actually complete entire workflows for you in your style. And best of all, they work in a channel that you already know and love because they are purpose-built Notion super users.

14:30Notion's new AI agents completely expands the range of what Notion can do. It can now build documents from your entire company's knowledge base, organize scattered information into organized reports, basically do tasks that used to take days, and get them complete in minutes. These agents don't just help with work, they finish it. Getting started with building on Notion is easier than ever. Notion agents are now your very own super user to help you onboard in minutes. Your AI teammates are ready to work. Try Notion AI for free at the link in our show notes. Welcome back to the AI Daily Brief. Today, we are talking about the latest product release from OpenAI called ChatGPT Pulse.

15:06We're going to talk about what it does, why it represents a paradigm shift for the user interaction mode of AI, and check out the first responses, after which I will give you my hot take on where I think this all goes. Now, taking a step back, people have been eagerly awaiting a set of new product releases from OpenAI. First of all, the company's new CEO of applications, Fiji Simo, is now in place. But then more specifically than that, a few days ago, Sam Altman tweeted, over the next few weeks, we are launching some new compute-intensive offerings. Because of the associated costs, some features will initially only be available to pro subscribers, and some new products will have additional fees.

15:42Our intention remains to drive the cost of intelligence down as aggressively as we can, and make our services widely available, and we are confident we will get there over time. But we also want to learn what's possible when we throw a lot of compute at today's model costs at interesting new ideas. Now, since then, people have just been speculating around what we might get. Continuing to be high on the list is Sora 2, as well as the reasoning model that recently won the ICPC, But last night and earlier this morning, we got our first hints around what was coming as people started to notice a ChatGPT feature that would surface personalized insights daily.

16:14That feature is called Pulse, and Sam Altman calls it his favorite feature of ChatGPT so far. Here's how Sam describes it. He says Pulse works for you overnight and keeps thinking about your interests, your connected data, your recent chats, and more. Every morning, you get a custom-generated set of stuff you might be interested in. It performs super well if you tell ChatGPT more about what's important to you. In regular chat, you could mention, I'd like to go visit Bora Bora someday, or My kid is six months old, and I'm interested in developmental milestones. And in the future, you might get useful updates.

16:45Think of treating ChatGPT like a super competent personal assistant. Sometimes you ask for things you need in the moment, but if you share general preferences, it will do a good job for you proactively. This also points to what I believe is the future of ChatGPT, a shift from being all reactive to being significantly proactive and extremely personalized. And those were a lot of the messages that were reinforced in the announcement blog post as well. Pulse is a once-a-day, everyday push that creates a personalized feed of highly relevant, or at least hopefully highly relevant, information. One thing that's interesting is that throughout the announcement post, they really lean on the language of research.

17:20Pulse, they write, is a new experience where ChatGPT proactively does research to deliver personalized updates based on your chat's feedback and connected apps like your calendar. You can curate what ChatGPT researches by letting it know what's useful and what isn't. ChatGPT can now do asynchronous research on your behalf, etc., etc., etc. It's very clear that the action that they are pushing with this is research, even if the output is this daily update. One other thing that they emphasize very strongly in both their announcement posts as well as the press coverage around it is that this is not meant to be some attention-capturing infinite scroll thing.

17:51They want to give you what's valuable, not get you stuck there. Initially, this is only rolling out for pro users, following that pattern that Sam had promised a couple of days ago. Fiji Simo tweeted, AI should do more than just answer questions. It should anticipate your needs and help you reach your goals. That's what we're beginning to build, starting with ChatGPT Pulse. ChatGPT head Nick Hurley says that there is more like this coming in the future. He writes, this is just the first step. Over time, we want ChatGPT to evolve from answering questions to proactively helping you make progress.

18:22Greg Brockman used the term that I think is particularly relevant for this, calling it a background agent which delivers updates to you every day on topics of interest. And this is an important part of the paradigm shift this represents. With this release, ChatGPT is very self-consciously experimenting with a move from reactive to proactive. Now, background agents are a big theme right now across a number of different domains. It's a big point of conversation, for example, in the agent decoding space, where whereas six months ago, all of agent decoding was you sitting in an IDE and telling it what to do and interacting as it brought code back, now agent decoding frequently involves spinning up background agents that just go off and do things on their own with much less human interaction.

19:01And indeed, this is actually a slightly different access for agents than sheer autonomy. We've got general purpose agents now, like Manus and GenSpark and even ChatGPT Agent, and these are some of our first paths into much more autonomous AI experiences. These tools have much more autonomy to figure out how to accomplish a goal and can take on much more comprehensive tasks. But they are still reactive to the prompt. You're still describing what you want it to do. It's just that what you can have it do can be much more comprehensive and complete than perhaps the way that you would prompt a chatbot like ChatGPT or Claude.

19:35We've also been having this conversation about ambient versus reactive AI in the context of devices. Some builders, like Avi Schiffman and Friend.com, are really all in on the idea of ambient AI. The friend pendant, for example, experiences the world alongside you and can interact with you in the world as such. Meta's Ray-Bans, meanwhile, while yes, still having more ambient awareness of what's going on around you, aren't necessarily as proactively taking actions, instead waiting for you to control the mode of interacting through the device. The point is that these really are quite different paradigms of interacting with AI.

20:10And I don't think right now it's super clear exactly where people are going to want to end up. Of course, the only way to learn that is by releasing products that do different things and seeing where things all shake out. Now, another piece of this is that it really puts a fine point on the importance of context and memory, particularly as lock-in. One of the biggest reasons that people still aren't counting Apple out of the AI space, even after all of their stumbles and seeming lack of action, is that the iPhone is just a big context machine. It has more information about you and your life as you live it than just about any other single source.

20:44We've seen a wave of recent email assistant tools like the one that was just announced this week from perplexity because, again, your email inbox is a place that is just chock full of context. Pulse seems, among other things, like a way for ChatGPT to take advantage of all that context. Pulse co-creator Andrew Chen wrote about this saying, Last week I asked about bike maintenance and Pulse remembered I'd been looking at weekend activities so it found me a bike route I didn't know existed near my place. When I was debugging Pulse code, it surfaced fresh RL papers from Arvix because it knows what's relevant to my work.

21:15I think memory and context are going to be huge sources of moat and defensibility in the future. And so it's smart, I think, for OpenAI to try to experiment with features that more explicitly leverage that memory and context to try to create experiences that would make it very hard to switch out of that environment. So what are people's first impressions? Well, a lot of them are positive. Mari Haines writes, ChatGPT Pulse looks super helpful. It proactively shows you things you might be interested in based on the conversation. Simon Smith writes, first impression, this is going to be insanely useful.

21:45It's like a newsfeed tailored to recent conversations. It reminded me of Perplexity's Discover feature, only way more personal. Like, you don't just get content on broad topics like artificial intelligence, you get content on very specific things you've been discussing with ChatGPT. For example, I was chatting with it the other day about AIs competing with each other, so it presented some context on that. It even finds content for things you've mentioned in passing but are genuinely interested in. It noticed I've been asking about Toronto's crash in condo market, for example, so found me an article on that.

22:11He also complimented the onboarding experience, and concluded ultimately that this feels like the most personalized news feed you can imagine. It makes me want to dump even more information and context and app connections into ChatGPT so I can get an even better daily feed. Sai Kambapati thinks that this is the start of a trend, saying, ChatGPT Pulse is going to spark a new wave of proactive technology. You're going to see a lot of apps building this out over the next few months. Olivia Moore from A16Z answered some questions that I had about how much this is going to be slanted personal versus professional.

22:41She wrote, at least on my testing, it surfaces mostly professional topics I've been discussing versus personal. This is in contrast to the launch videos which showed more of the personal use cases. I'm curious if that's a deliberate decision not to spook users at launch. The model seems to be, identify trends across topics you care about, summarize them with key insights, point you towards more resources to unpack them further. She also noted that despite the fact that she had connectors set up and had given it access to calendar, it didn't bring in anything from her meeting schedule. There were some folks who were a little bit more skeptical.