transcript

Speaker 1:
[00:00] Today, we're discussing how knowledge workers in general, but everyone else too, should be using Opus 4.7 and the new Codex app. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. Now, today is probably my favorite type of show, when we get a whole slew of new goodies, and get to dig in and see what they can do for us, how our capabilities have changed, what new use cases become unlocked, and what the patterns are telling us about where the world is going. Now, yesterday, we got not one but two big releases, one model and one harness. The model, disappointing to some, was not mythos preview or anything related to it, it was Opus 4.7. You could feel around the communications that Anthropic knew that there was going to be some amount of disappointment that this wasn't mythos and so it was going to have to be fairly impressive on its own right. Now, on the other side, from OpenAI, we got a new iteration of their Codex application. It adds a whole bunch of new capabilities and is making some very different bets as compared to how, for example, Anthropic is looking at its Cloud Desktop app. So what we're going to do today is discuss all the new things in both of these releases, get some of the first reactions, and then specifically dive deep on what you as an engaged AI user or knowledge worker or entrepreneur should try with these new releases. By the way, if you want to follow along this episode, you can go to play.aidailybrief.ai, it's where I keep companion experiences, and there is a whole website slash slide presentation that has all the information that I'm going to share here, including some of the ideas for what you should do. So let's talk first about what is new in Codex. Certainly one that people are talking about quite a bit, is that Codex now has computer use on Mac. Codex can see, click, and type across any app on your computer with its own cursor. Multiple agents can work in parallel in the background without interfering with what you're doing. And Codex can now use apps that don't have APIs. Now one of the big ideas that you're going to see is that Codex, which was nominally designed as an app for coding, is very quickly becoming not just for coding. Yesterday I tweeted that the problem with the term vibe coding ended up not actually being that all coding became vibe coding, but that all knowledge work is becoming coding work. And you can see that very much on display in terms of where the Codex app is going. Another new feature is the in-app browser with comment mode. Basically you can now load a page inside Codex and click directly on elements to give the agent precise context. This is really useful for things like front-end iteration, bug reporting, and basically any workflow where pointing at the thing is faster than just describing it. Native image generation now lives in Codex with GPT Image 1.5, meaning that you can generate mockups, edit images and create variants all inside the same thread as everything else. This pairs really well with the new rich file previews and artifacts beyond codes. PDFs, spreadsheets, slides, and documents now render inline in the sidebar. Codex produces these as artifacts that can be downloaded and interacted with not just as code. One thing that's really clear from the New Codex is that they are definitely taking lessons from OpenClaw to heart. Pash from OpenAI writes, Biggest lesson from OpenClaw is that a good teammate doesn't start from scratch every time you check in. They remember what was decided, what's still open, and proactively help you. Today we launched Heartbeats in Codex, automations that maintain context inside a single thread over time. Instead of each run starting fresh, Codex wakes up in the same conversation, with the history and context it needs already in place. You can also have it schedule its own next steps. Think about the overhead that quietly accumulates every morning, scanning Slack channels, catching up on email, piecing together what moved overnight. With a heartbeat, you offload that once and wake up to a brief already waiting in a pinned thread. Now, Pash suggests turning Codex into a Chief of Staff, which is something we'll come back to in a little bit. So to summarize, you've got here automations that resume existing threads, which establishes this whole new monothread pattern, which we're going to talk about in just a minute. And Codex also has project list threads. Flavio Adamo writes, The most under-readed feature in the New Codex is chats without a project. Before this, I was literally using a project called Trashcan as a home for every random thought or personal tasks. Basically, this means you can just dive in without having to pick a repo first. This is what led Jason Liu to call it the New Notes app. There are also a whole bunch of daily use quality of life improvements in Codex including a Mac OS menu bar and a Windows system tray with pinned in recent threads, a global hotkey to bring up a mini Codex window from anywhere on your Mac, tabbed terminals inside each thread so you can run build, servers and tests in parallel, slash compact as a standalone command, and a theme picker for the command palette. Now, one note on the computer use thing that so many people are excited about, that is Mac only right now although they say Windows is coming. People's first impressions are good. Riley Brown from the Vibecode app writes, This is exactly what I was hoping for. Full permissions, no co-work like feature which limits agents' abilities, just Codex. If you ask for a coding task, it writes code and gives you a preview. If you ask for a presentation or doc, it gives you a presentation or doc. Organized by project on the left sidebar, easy to create skills, easy to app mention skills and plugins. Now, this pattern of not breaking things into different UIs for different use cases is something we'll come back to as well and is a major differentiation between the way that codex is evolving and the way the cloud desktop app is currently set up. Commenting on computer use, Ari Weinstein writes, This is the first time I've ever seen an LLM operate a GUI as fast as a person and it's surreal. Aaron Levy from Box gets that this is very clearly not just codex as an update for developers, but is thinking about how knowledge workers in general will work in the future. He writes, The new codex is another jump in what agents will look like for knowledge workers. Agents that can code, work with tools, and use computers can begin to execute long-running tasks in the background for all areas of work. This can mean drafting reports, setting up data rooms for a merger, reviewing contracts, helping onboard clients, generating marketing assets, processing invoices and more. So a couple of things that I wanted to double-click on. Nick Bauman on the codex team wrote an interesting post called My Codex Threads Are Alive, and the big statement from Nick is that he has become monothread-pilled. Nick writes, The most useful codex thread I have right now is the one I've been using for the last three weeks. Every hour it checks my Slack, gmails and PRs I wrote or am watching. It turns the noise into clean signal I can act on. My codex usage has shifted from starting lots of short-lived chats to keeping a small number of threads alive around recurring workstreams. I still start fresh threads constantly, but some work should not reset every time I ask a question. So the old mental model of AI assistance is that you either may start fresh for every task or B, maybe create a project folder where context can live around a set of tasks, but where you're still frequently starting fresh, just hopefully relying on the context that's stored in the project to have the new thread be up to speed. Now this paradigm of every question being a new chat and every project being a new conversation was to some extent forced on us by technical limitations. It was a byproduct of the fact that long threads used to degrade. Context got muddy, the agent lost the plot and you were better off starting over. One of the key pillars of my work when I'm working on complex projects with Cloud or ChatGPT is the handoff documents I have the AI create as I start to see the signs of them running into the end of their context window. However, the Codex team has now shipped compaction improvements that weaken that assumption. About a week ago, engineer Anthony Kroger wrote, I literally never worry about context windows using Codex. It can compact like three times and the model still remembers the details somehow. Back even before this new release, Nick Bauman again wrote, So much coding agent design is built on the assumption that breaching context windows and compacting context yields progressively worse results. When you drop this assumption, the product direction it opens up is very exciting. He continues in his new post, put simply, with good context compaction, a thread's value increases over time. I've talked in the past about how we need some sort of benchmark, for new models or new product releases, that isn't about performance on standardized tests, but about the new use cases that get unlocked by any new release. Nick is basically talking about exactly that. He writes, my own version of a mono thread is a work teammate thread. My work is noisy and spread across Slack, Gmail, Gcal, GitHub, files in an Obsidian Vault, and a bunch of other Codex threads. I need something that can filter the noise and tell me which few things are worth caring about. I use one thread to check those places, remember the current priorities, and tell me when something needs my attention before I would have found it myself. I run this as one main teammate thread plus a few long-lived sub-agent threads. The main thread handles orchestration and judgment. The sub-agent threads keep depth in their specialties. The main thread can also spawn new sub-agents for new work streams as they appear. The main thread wakes up, checks the current priority, reads the smallest useful live signal, uses a specialist sub-agent thread only if that lane matters, and then decides whether to notify me or stay quiet. Now, what's super interesting to me about this is that this is basically an alternative architecture for the project manager open clause and chief of staff open clause that I built as part of my first experiments with that system. This is, of course, a radically simpler implementation of that. And speaking of open clause, part of how Nick gets value out of these monothreads is thread automations. He writes, A thread automation is an interval trigger on an existing codex thread. It is not just a scheduled prompt because the automation runs in the same thread with context and corrections already there. That makes the natural prompt very simple. Keep an eye on this for me. If a thread checks Slack, Gmail, GitHub, Docs, and Calendar on a schedule, it accumulates examples of what you care about. It sees which asks you act on, which drafts you edit, which updates you ignore, and which sources usually matter. Over time, the useful behavior is not a bigger summary. It is a short interruption when something actually matters. Now, Jason Liu from OpenAI takes this a step farther, actually creating a recipe for a personal chief of staff. The Codex Chief of Staff takes advantage of a local folder, Vault, which is the durable memory layer and the working folder that Codex opens up and interacts with. The Vault has a small agents.md file that tells Codex how the Vault works. The principles that Jason shares are a projects folder that gets one note per active project or workstream, and a notes folder that gets scratch notes, drafts, and one-off captures. The agents.md file creates a number of instructions around how to work, like preferring to update existing notes over creating new ones, and keeping facts separate from guesses and more. From there, the chief of staff interviews you to get a sense of who you are, what are you responsible for, who matters, what are you worried about missing, which Slack channels, email threads, docs, repos, and meetings matter. What do you not want to be interrupted about? Now, if you've tried the personal context portfolio I released a couple of weeks ago, you could, of course, just transport that over there and not even have to do the interview step, although there is value, of course, in having a follow-up interview even after you've given Codex all of your personal context. From there, Codex proposes the 3 to 7 project notes to create, the smallest useful agents.md improvements, and which plugins or connectors to install. Those common plugins might be things like Slack, Gmail, Drive, Calendar, GitHub, or more. Now, there's more in here, but the one last piece that I wanted to point out, harkening back to the classification of everything, is the idea of the core loop being an every 15 minute chief of staff heartbeat. Every 15 minutes or at whatever interval you want, the thread wakes up, and like Nick Bauman's MonoThread, checks whatever sources you gave it access to like Slack or Gmail, looks for pending asks, blockers, or decisions. It notices how your priorities seem to be changing, and it keeps interviewing you over time. As it does so, it uses your answers to improve the heartbeat prompt, agents.md, and project notes. So I think if you're going to try just one thing with Codex, it would be this MonoThread slash chief of staff idea. But I've also put on this companion site a ton of other use cases that I think are worth trying and that are enabled by this new set of features. So one category of these is around recurring reporting and monitoring. Basically, anything where you have some sort of frequently repeated reporting need, where you have to look at a bunch of sources, aggregate them, pull out the most important signal and do something with it, is really well suited to the new features of the Codex app. That could be a morning brief that pulls Slack DMs, unread emails, Notion updates and calendar. It could be a weekly customer health check that looks like channels like Intercom. And you can probably think about a half dozen more of these recurring monitoring type situations that you interact with. Some other ideas to take advantage of the new computer use, for those of you on Mac, are things like legacy system data entry. If you have some old vendor portal or ancient ERP or accounting software from a decade ago, the computer use features could drive those systems now and make your life significantly easier. You could also try moving data between systems that don't integrate. One example that some people have given is moving from granola to obsidian vaults. There are about a dozen different ideas there of other Codex use cases worth trying now. But let's move on to Opus 47. The biggest knock on Opus 47 is not about what it is, but about what it is not. For the last couple of weeks, we've been hearing about just how powerful Anthropics Mythos Preview model is, and this is not that. Still, it does seem to represent a pretty meaningful capability jump, and if it weren't for knowing that Mythos Preview was out there, my instinct is that people would be pretty stoked about this. Of course, some people are. As they often do, I think Leighton Space nailed it, calling it literally one step better than 4.6 in every dimension. If you look at just the agentic coding chart, you get a sense of what 4.7 is about. 4.7 low is strictly better than 4.6 medium. 4.7 medium is strictly better than 4.6 high. 4.7 high is now better than 4.6 max. Now that's reflected in the overall coding benchmarks, but you see the same pattern in other benchmarks that matter for knowledge workers as well. The Finance Agent jumps from 60.1 to 64.4 percent. Office QA Pro from 57.1 to 80.6 percent. OS World Computer Use 72.7 to 78 percent. Basically, you can see that these are in many cases not just incremental changes, they're pretty meaningful. People's first experience with this seems to validate the benchmarks. They made about 20 percent more money on the Vending Bench 2 test, and many people's first tests with this around visual and design tasks are really positive as well. Mike Taylor writes, Opus 4.7 has the distinct honor of making the best PowerPoint I've ever seen in an LLM. Adam Dot New writes, Opus 4.7 appears to be state of the art at Agentic CAD design. This Week in AI argues that the leap in design sensibility between 4.6 and 4.7 is really significant as well. Now, I did dig into this because front-end design and website design is one of my most frequent use cases, and I wanted to test not only its design capabilities, but its reasoning around design. So I gave both 4.6 and 4.7 the task of redesigning, the kitschy and fun but ultimately kind of challenging AI Daily Brief website that's currently in its terminal theme into something different. 4.6, which is a good designer, did a good job. Although if you've used Claude out of the box for design, it is going to feel very Claude to you. The font choices at this point are getting extremely predictable, as are the color palettes. I was able to push it to do another direction, which was a little more in line with the terminal theme, and again, it did a totally fine job. What I would say about my interaction with 4.7 on this is that one, it certainly had more variety in terms of the visual approaches it was proposing, and when I slowed it down, it could actually do some thoughtful reasoning on the ways to set up the site, but it certainly wasn't a panacea. Based on my first experience, the band of what I'm able to get out of 4.7 is a meaningful upgrade, but I almost have to slow it down and make sure that it uses its full reasoning capabilities before it just rips out to design something that looks good but isn't all that well considered. Now, there are a few areas where there seem to be some regressions as well. On one long context retrieval benchmark, the score between 4.6 and 4.7 dropped from 78.3 percent to 32.2 percent, although Cloud Code creator Boris Czerny said that that benchmark is being phased out because they believe that it overweights distractor stacking tricks and doesn't reflect real applied reasoning. Now, with the new model, the team in Anthropic suggests that there are some tweaks to how you want to interact with it to get the most out of it, and that might break patterns from how you've used models like 4.6 in the past. Cat Wu, who is one of the leaders of the Cloud Code team at Anthropic and co-creators of it, gave a few tips. One, she suggested to delegate, not micromanage. Basically, she said treat the model like a capable engineer that you're handing a task to, not a pair programmer that you're guiding line by line. Progressive clarification across multiple turns can actually reduce quality on 4.7. Relatedly, she suggests putting the full goal, constraints and acceptance criteria right front. With every user turn adding reasoning overhead, it makes more sense to give the model everything it needs up front. She also said that Opus 4.7 is better at self-verification than any previous Cloud model, but that you have to tell it how to verify and build the verification loop in. Cloud Codes Boris Cherny also shared a few tips. For example, he talks about a new way to configure the effort level. Boris writes, Personally, I use extra high effort for most tasks and max effort for the hardest tasks. Max applies to just your current sessions. Other effort levels are sticky and persist for your next session also. So, what are some things that you should try outside of just updates to your coding with Cloud Code? One thing to check out is that there seem to be fairly big vision improvements, which means that for things like taking whiteboard photos for meetings and translating them or trying to interact with dense dashboard screenshots, this model should be much better. It should also be able to better pull chart images from PDFs, 10Ks research reports and things like that. And it should be able to better reason over screenshots as well. Think about for example looking at the onboarding flow from a competitor and comparing it to your companies and asking what the competitor is doing better. Maybe even a bigger thing to try is longer harder tasks. Everyone from the Anthropic team really emphasized that this model is all about less babysitting and more real delegation. So what does this open up? Well, you should try things like end-to-end research projects. Instead of summarize this article, get it to research the state of a topic, using a bunch of URLs, the internal notes, and outputting a significant product on the other side. You can also do extended reasoning tasks like legal argument construction, investment thesis development, or strategic option analysis that previously you might have had to break into pieces because the model would lose the thread, but which now can be done in one pass. Full deliverable production, complex data cleaning, cross-functional synthesis, multi-step analysis with verification, basically any harder reasoning tasks that you might previously have tried to break into smaller pieces, you should at least go try to see how 47 handles them natively right now without chunking them into those smaller parts. Now, one more thing that I wanted to point out is a slight difference, at least right now, in the UI design philosophy between the Codex app and the Cloud desktop app. And remember, we got an update for the Cloud desktop app just this week, so this is about as good a comparison as you could ask for right now. In Cloud desktop, you toggle between different experiences for Cloud Chat, Cloud Cowork and Cloud Code. On Codex, it's just all one thing. Again, I read this before, but what Riley Brown said, this is exactly what I was hoping for. Full permissions, no Cowork-like feature, which limits agent abilities, just Codex. If you ask for a coding task, it writes code and gives you a preview. If you ask for a presentation or doc, it gives you a presentation or doc, organized by project on the left sidebar. So the bet on the OpenAI Codex side is that the agent is smart enough that the interface should basically disappear. The implied thesis is that switching modes is friction. And frankly, it harkens back to the original ChatGPT interface, which is like one text box infinite capabilities. On the other hand, Claude, at least for now, is betting that these three different modes of working are different enough that collapsing them into one interface creates compromise. It's closer to the way that native apps are designed now, i.e. you don't write documents in your email client. The good news for you as users is that if you have strong preference towards one or the other, at least for the moment, you have a choice for whichever is better for you. Overall, given that this was not the release of Mythos or OpenAI Spud, these things taken together still represent a pretty significant set of upgrades and new features that are going to take us some time to really integrate into how we work. For those of you who want to spend the weekend building and trying things, again, if you go over to play.aidailybrief.ai, the last slide is going to be 11 things that you can try right now, using these tools to see how much you can get out of them. I know for me the one that I'm going to experiment with is the MonoThread approach and the Codex Chief of Staff, which should be especially interesting to see how it compared to the version of that that I originally created in OpenClaw. For now though, that is going to be our AI Daily Brief for the day. I appreciate you listening or watching as always. Have tons of fun this weekend, and until next time, peace!