transcript

Speaker 1:
[00:00] Today on the AI Daily Brief, the new ChatGPT Images 2.0 model and why it's the first image model for the agentic era. Before that, in the headlines, a big team up between SpaceX and Cursor. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends, quick announcements before we dive in. aidailybrief.ai is where you can see all the things going on in the ecosystem. Check it out, subscribe to the newsletter, come join us on the AI Operators Community, have a grand old time. And with that out of the way, let's get into the headlines. SpaceX has signed a massive new deal with Cursor that adds a pretty meaningful twist to their rapidly approaching IPO. On Tuesday, SpaceX announced in a post on X, of course, SpaceX AI and Cursor are now working closely together to create the world's best coding and knowledge work AI. Now, it had previously been rumored that Cursor would be renting XAI servers for their next training run, but it now appears the collaboration is going much deeper. SpaceX continued, The combination of Cursor's leading product and distribution to expert software engineers with SpaceX's million H100 equivalent Colossus training supercomputer will allow us to build the world's most useful model. The post also announced, and obviously this is the part that everyone focused on, that SpaceX had been granted the rights to acquire Cursor at a $60 billion valuation later this year, and if the acquisition doesn't go through, SpaceX will pay Cursor $10 billion for their collaborative work. The deal potentially solves a number of problems for both companies. By some reports, Cursor has been backed into a corner over the past six months. Reports have suggested that they are making a loss on every Claude and OpenAI token they serve, so much of this year has been focused on developing a state-of-the-art in-house model. And of course, beyond just the training runs, Cursor will need access to a ton of additional compute as they scale up revenue. The company is reportedly in talks to raise $2 billion in venture funding, and even if that round does close, they are still massively resource constrained compared to OpenAI and Anthropic. Going back to the Harness Engineering episode from last week, the challenge for Cursor is, in short, that the biggies are not choosing model vs. harness, they are doing both and. So then here's the logic for why to team up with XAI. The company has access to huge amounts of compute that at the moment aren't doing as much as they could. Elon Musk has claimed the data centers are currently being utilized for massive training runs, but XAI has struggled to generate revenue from their products, and they also haven't released an impactful model in months and at this point have no meaningful footprint in the AI coding space. Cursor would provide a huge data pipeline to help XAI catch up. A joint Cursor XAI coding model could be exactly the kind of product that could help return XAI to relevancy. XAI has already taken a look under the hooded cursor with Musk poaching two senior engineering leaders last month. Now in addition to the Cursor deal, which obviously for our purposes is the biggest part of this news, the IPO disclosure process is also uncovering a bunch of additional details about SpaceX as well. The information got hold of confidential disclosure documents, suggesting that Elon upped his stake in the company last year, purchasing $1.4 billion in stock from current and former employees. SpaceX also plans to award Elon a compensation package with very lofty milestone goals. He could receive tens of millions of shares of SpaceX, tied to market cap achievements ranging from $1.1 trillion all the way up to $6.6 trillion. For context, SpaceX is expected to target $1.5 trillion at IPO, meaning the low end might be easily achieved. However, a $6.6 trillion valuation exceeds NVIDIA's $4.9 trillion as the most valuable company in the world. The documents also discuss a stock incentive tied to deploying 100 terawatts of compute power via spacefaring data centers. Peak energy demand in the US is less than 1 terawatt, giving a sense that this is a science fiction style goal at present. The IPO is currently expected in June, and the debate is on around what the implications are for the entire AI industry. SpaceX will be going first, so theoretically their success or failure could impact IPOs from OpenAI and Anthropic in the fall, and yet I'm just not sure that that's exactly how it'll play out. I think mostly the SpaceX deal is going to be a referendum on how much exposure people want to Elon. Maybe this cursor tie up makes the AI part of the story less of a sideshow, but right now I'm just not sure. In another very big story that broke last night, an unauthorized group has gained access to Claude Mythos, playing right into cybersecurity fears. Bloomberg reports that users from a private Discord group gained access to Mythos on the same day Anthropic announced its preview release. That release was of course intended to be limited to a small group of companies for cybersecurity purposes. When they announced the model, Anthropic told the press that access would be tightly controlled to ensure that it didn't end up in the wrong hands. Bloomberg's source provided screenshots and a live demonstration of the model, implying that the breach hadn't been detected and that the group still had access weeks later. The source said that the group had been regularly using Mythos but hasn't used it for cybersecurity purposes in an attempt to avoid detection from Anthropic. Instead the group has been testing the model on relatively mundane tasks like website design. The source said the group isn't interested in malicious use, they just want to play around with unreleased models. Now in terms of how they got access to this, the source said that Mythos was accessed through a third-party vendor where one member is employed, and it also required a few educated guesses based on information gleaned from the recent Merckort data breach. Basically the member working at the third-party vendor has general access to Anthropix models including pre-release models as part of an evaluation contract. Anthropix responded to the report by stating, We're investigating a report claiming unauthorized access to Claude Mythos' preview through one of our third-party vendor environments. Anthropix added that they have no evidence that access went beyond the third-party vendor's environment or that it's impacting Anthropix systems. Now the discussion of this on X has been extremely breathless and overwrought, which is perhaps understandable given the way that Anthropix has chosen to promote this model. Coincidentally, Sam Altman had some pretty pointed comments about the way Anthropix had introduced Mythos. In a podcast interview that came out earlier this week, he said, If what you want is control of AI because we're the trustworthy people, I think fear-based marketing is probably the most effective way to justify that. That doesn't mean it's not legitimate in some places, but it is clearly incredible marketing to say, we have built a bomb, we're about to drop it on your head, we will sell you a bomb shelter for $100 million. You need to run it to access all your stuff, but only if we pick you as a customer. Wow, gloves are off. And it seems like we might be getting OpenAI's different approach to that in not too long. Lastly today, Google has released a big new upgrade to their Deep Research agents. The agent is now available in two flavors, the standard version and a state-of-the-art version called Deep Research Max. The agent now features MCP support to connect to third-party data sources for the first time. As part of MCP support, users can define arbitrary tools rather than relying on the agent to figure it out. The agents can now also output charts and infographics within their report, tapping into the nano-banana models for image generation. Both the normal and max versions produce a pretty significant bump to relevant benchmarks, with the max version now state-of-the-art compared to GPT 5.4 and Opus 4.6. Interestingly, the agents are still just Gemini 3.1 Pro under the hood, the same as the previous version of Deep Research. This means the entire improvement was driven by harness upgrades and additional inference rather than a more advanced model. The agents are only available through the API, so they are designed to be used in professional workflows. Google said that Deep Research Max is designed to consult significantly more sources and identify critical nuances that are overlooked by other agents. They wrote, The result is a nuance report that draws from authoritative sources like SEC filings and open access peer-reviewed journals, lays out information well and transforms dense technical data into actionable stakeholder ready formats. A small upgrade on the surface but one which could be extremely valuable to people who have a deep research use case. For now though of course that is not the new model that everyone wants to talk about today, so that is going to do it for the headlines. Next up, the main episode. Welcome back to the AI Daily Brief. Image generation models are kind of an interesting phenomenon in AI. They are in many ways the simplest quickest way to understand the power of this new generative medium. In fact, I think for many of us, it was in fact image generation that was kind of our gateway into the space. As impressive as the initial ChatGPT was, what really caught my attention all the way back at the end of 2022 and the beginning of 2023 was the absolute feeling of being a wizard when I was creating images of Hemingway in Paris in the 1920s with the Midjourney model at the time, which is about 100 iterations ago at this point. And yet to some extent, there's always been a little bit of a gap between how cool AI image generation was and how useful it was, at least for many people in many use cases. Which is not at all to say that AI image generation models have just been novel rather than useful. I personally have, I don't know, a half dozen use cases that I use them for literally every single day. But if you think back at the sequence of model releases, the big moments have been general consumer viral moments like OpenAI Studio Ghibli moment last year. Now, when it comes to this new OpenAI model, what I want to argue with today's show is not only that we are getting to a capability set that unlocks more use cases, but also that increasingly it is clear that the power of the image generation models is going to be in their integration with other systems, not just what they can do stand alone. Now, when it comes to this new ChatGPT image model, there has been speculation for a couple of weeks now that this model was live in the world being tested on Arena. Many users pointed to very impressive generations from LM Arena that included things like handwritten notes, layouts of YouTube page, and a simple kind of janky iPhone style image of a retail store. What people were noticing about these things is how little they felt like AI images. They just seemed like a random iPhone photo or a screenshot. And people also identified that they seemed to have good world knowledge. They weren't just making stuff up in their images. They were actually bringing what the model knew into their ability to create. Yesterday, as I mentioned at the end of the show, OpenAI teased that the new model would be coming in the afternoon. And indeed, on Tuesday around 3 p.m. Eastern, we got the new ChatGPT Images 2.0. From a sheer quality standpoint alone, there is absolutely no denying that the model is fairly stunning. And that seems to largely be consensus. Arena announced that not only did GPT Image 2 take the number 1 slot in their ELO score human preference board, it absolutely dominated. The number 2 through 15 image generators are all clustered basically within 100-130 points of each other. Number 15, Flux 2 Dev, had a score of 1149, whereas the previous leader, Nano Banana 2, had a score of 1271. GPT Image 2 came in over the top with a 1512. Arena points out that that is a record-breaking 242-point lead in the text-to-image category and the largest gap they've ever seen. In their announcement post, OpenAI gets into a lot of what makes this model different and what it can do. They write, or more accurately generate, in an image, an announcement post that argues that, quote, this model is a step change in detailed instruction following, placing and relating objects accurately, and rendering dense text, with the ability to generate across aspect ratios. They say that it has better composition and visual taste, meaning it feels less AI generated, it has, as people were speculating, more world knowledge, and the ability to actually reason and think. They write, when a thinking model is selected in ChatGPT, Images 2.0 can search the web for real-time information, create multiple distinct images from one prompt, and double-check its own outputs. Now, this was, of course, the big unlock from NanoBanana 2, but it seems like, as we'll see with some of the early examples, that this model takes it to the next level. In terms of the capabilities they highlight, right at the top is Greater Precision and Control. They point out that it can do small text iconography, tiny UI elements, dense compositions, and it can do so at significant resolution up to 2K. The practical effect, they say, is instead of getting something vaguely in the neighborhood of what you meant, you get something you can actually use. One example they gave of this is this pile of rice with a tiny kernel in the middle that has a little bitty GPT Image 2 written on it. This model is multilingual, which they say not only helps with translation, but can also quote generate visually coherent outputs where language is a part of the design itself, from posters and explainers to diagrams and comics. Some of the other things they point out are much enhanced stylistic sophistication and better realism, including they point out adding tiny flaws that add realism. And they also discuss this idea of real world intelligence. Pointing out that it doesn't just mean it's cool, it unlocks a set of use cases like explainers, maps, educational graphics, and visual summaries, where, as they put it, correctness and clarity matter just as much as aesthetics. Now, one last thing they note on the utility side of things is that they can now generate more flexible sets of aspect ratios, which goes to just give people more fine-grained control. Now, in terms of community response, the first thing that I noticed is that a lot of people, including by the way Sam Altman himself in a recent interview, came into this new model effectively feeling like for all intents and purposes, image generation was solved, or at least that if it could get better, it was just an incremental better that didn't really matter practically. For many, that perspective has now been blown out of the water. Ethan Malik writes, I didn't think that better image generators would be a big deal, but it turns out that there is a quality threshold I didn't expect where you can now get text slides and academic papers. And as people dug in, there were a few themes that I saw over and over and over again. The first was how much less AI-ness a lot of these photos had, that you really could get not just pretty images but very realistic looking images, including things like not-so-great regular photography. Pietro Sorano shared the output of a photo of a computer screen displaying a Spotify playlist at night, adding it's an insane model and a true imagination engine with an incredible level of realism and small details. Now while the realism is impressive, a lot of people jump straight to the implications of the massively improved text and detail handling. It unlocks things like entire comic panels. And by the way, the ability to generate multiple images and to keep character consistency makes larger editorial generation along these lines much more possible as well. The detail in text was showed off in all sorts of different ways. Iman Mastok created the periodic table of the original 151 Pokemon. Chris Cascianova did a Where's Waldo style illustration, placing herself in a densely crowded New York City scene, while others like Nick Duns took messy handwritten photos, asking ChatGPT to quote, get rid of the creases and make it a scan with both of the generated outputs, not only perfectly capturing all the information on the pages, but even preserving Nick's handwriting. Other people are experimenting with all sorts of other use cases, taking an image of a house and turning it into a generated floor plan, improving the visual quality of graphs, making technical diagrams, brand kits, combined styles and more. One of the craziest tests showing off Image 2's world knowledge came from entrepreneur and content creator Riley Brown. He asked the model to create an image of a specific book, including a barcode which would actually take you to that publication, and it actually worked. He used a barcode scanner on his phone to test the image, and sure enough, it actually took him to that specific publication. Testing to make sure it wasn't the ISBN number, he even covered that part up, just leaving the barcode, and it still worked. Still, maybe the most common thing that I saw explored was UI and software designs. And this gets into what I think is actually really important about not just this model, but the context into which this model is coming. The short of it is, I think this is the first image model whose biggest impact is not going to be stand-alone viral moments like Ghibli, but has the potential to actually be integrated quite quickly into the agentic stack. Prins on X writes, Images 2.0 is the first model I have ever tried that feels ready for real enterprise workflows. It's a reasoning model, which means it will search the web, use tools, and think about your request before generating the image. It is able to generate huge volumes of text without a single error. It can keep the image sharp and consistent in between generations, giving you the ability to make additional edits to any image of your liking. The example they gave was asking the model to generate an organizational chart of a public company based on a template. And yet if Prins was thinking about general knowledge worker and enterprise usage, where other people went was much more focused on one specific use case. Mark Crutchman writes, Some of you were disappointed that we only, quote unquote, get an image model from OpenAI today, but you need to see the big picture. GPT Image 2 can generate mockups of websites, which Codex can then turn straight into working code. Choi Arakees goes farther with this, saying, The Codex plus GPT Image 2 pipeline is completely broken. This is the single most disruptive AI workflow I've seen this year. Stop thinking of AI as just a text generator. The real magic happens when you chain the models together. Now Image 2 is coming into a moment when OpenAI has just also announced that there are now 4 million Codex users up from about 200,000 at the beginning of the year. And already less than 24 hours into this, we are seeing people sharing all of their production pipelines from Image 2 UI to Codex. Peter Gostev from Arena writes, GPT Image 2 plus Codex or how to make Codex not suck at UI. Step 1, generate a UI image. Step 2, get Codex to implement the UI based on it. Step 3, get Codex to iterate until it aligns with the image as much as possible. Codex is bad at initial UI but very good at implementing a reference design so this is your way out. Iterate with the image model first and then Codex will do a good job. In many if not most people's estimation, Codex's biggest limitation has been UI. It's certainly one of the reasons for me that Claude Code has remained my primary driver. Although they are obviously different products, I don't think it's unreasonable to compare the combination of the new Codex app plus GPT Image 2 with the new Claude design feature released by Anthropic. As we discussed in that recent show, Anthropic doesn't have a native image generator so the way that they're creating those designs is a little bit different, and it seems pretty likely to me that there are going to be certain types of UI implementations that simply will not be possible with Claude design and Claude code, but that will be with the integration of GPT Image 2. What's more, people are really excited for when we get the next base model with this as well. Simon Smith writes, the image generation to Code Workflow and Codex is going to be spectacular when we get GPT 5.5. I tried it with 5.4 and it's already pretty good. OpenAI is bringing the pieces together. Of course, even if OpenAI weren't bringing the pieces together, there are plenty of entrepreneurial people out there who will do that for them? Something big is happening author Matt Schumer dumped Image 2 into the general agent that he built, leading to it generating slide decks and apps that, in his words, look like they were designed by pros. Leon Lin already posted a new skill that takes advantage of GPT Images 2 to GitHub to make the integration between Images 2 and Codex even smoother. Now I will say that if the vast majority of people's experiences so far have been positive, that wasn't universally the case. Boyantungus writes, I tried making an infographic using GPT Image 2, lots and lots of visually unacceptable artifacts. Someone did suggest that his settings might have been set on low, but obviously that's still going to be an issue in terms of the actual utility of the thing. Speaking of which, journalist Sharon Goldman tested it by asking the model to create an anatomically correct labeled image of the human thorax to be reviewed by her sister, who is a professor of anatomy at a med school. It looked great, but her sister pointed out that there was an extra set of veins, labels pointing to the wrong parts, and some issues with where things were placed. And while obviously this is still a major improvement from what we had before, there are use cases like this one, where the tolerance for mistakes is not 5% but zero. One of the things that I think will be really interesting to see is how many of the new use cases that get unlocked by this new model actually get deployed in practice. For example, one of the things that it can do now is much better, more richer editorial layouts. And yet, is there a group of people who actually need to create editorial layouts who will be willing to trade the controls that they lose in terms of their existing processes for the speed or quality that this new approach represents? I don't think the answer is going to be clear cut there. Another example is precision marketing assets. Image 2 we can already see does an awesome job with things like visual Instagram ads, but will it be people who are creating existing Instagram ads using their own dialed in workflows with still even more fine-grained specific controls, or will the unlock be more about the democratization of the ability to create that type of image or asset from two other types of people? I think overall we're still figuring out what it really means and where the value lies in having reasoning over images. I think we're still figuring out where the line of controllability needs to be to make these skills useful, not just novel. By far the use case that I will be paying the most attention to in the immediate term is this UI codex type of integration. In my first tests, it did make a big difference in terms of the quality of what I could get out of codex when it comes to UI design, but it was still more in the realm of reference images than it was actually about implementing a specific already designed UI. Maybe one last thing to note is that the team at OpenAI is very clearly teasing that this is one of the first examples we've seen of what you can do when you have more resources to throw at a model's training. Greg Brockman doesn't talk about people who have argued in the past that we've hit a pre-training wall, but he does say really incredible what you're now able to create with a little bit of compute. Sounds like from what most folks are thinking and hearing, we might not be all that far from getting to see what that little bit of compute does even outside of images as well. For now, this model has plenty of new capabilities to go play around with and I am excited to see what you do. But that's going to do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace.