What the OpenAI Web Browser Means; Musk Says Grok 4 is No. 1

OpenAI’s embrace of Google Cloud servers isn’t stopping it from launching a direct assault on several of Google’s core products. Its next target: Chrome.

Reuters on Wednesday reported that OpenAI’s web browser is coming soon. It’s been eight months since we broke the first news about OpenAI’s development of the browser, including its hiring of two early Chrome engineers from Google. If you’re wondering why it takes so long to make a browser, the article explained that there are a bunch of security and site authentication features that a browser needs to work properly.

As for what this browser would aim to do, it’s best to think about how companies like OpenAI want web services to work in the future. In essence, they appear to want customers of an app like ChatGPT to tell the AI what task they want to accomplish, such as doing research or handling a transaction, and that launches an agent to go off and do it. 

If OpenAI and other AI providers have their way, all manner of websites will adopt Model Context Protocol, which lets agents have access to their systems. (Please check out our helpful field guide on AI agents here.)

In OpenAI’s ideal future, ChatGPT is the front end of the web and most web traffic becomes bot traffic, handling the errands of OpenAI customers. OpenAI previously launched Operator, a “computer-using agent” that can use a person’s browser to handle some web-based tasks, though early customers said it’s been janky. Presumably, controlling the browser will improve that kind of product.

In any case, OpenAI’s vision of the future scares a lot of web app owners, from content publishers to commerce firms like DoorDash. Those businesses want to control how people experience their services, and many sell ads on their sites. 

We can’t emphasize enough how much website publishers are getting crushed by AI. More people are using ChatGPT, rather than individual sites, to gather information on anything. And as we covered in yesterday’s issue, Google doesn’t give sites the ability to opt out of its AI Overviews or AI Mode features—which summarize web content without requiring people to visit other websites—unless the sites also opt out of being included in Google’s search index altogether.

No word yet on whether OpenAI will promote Natural Language Web, software that helps retailers, magazines and other websites install a customized version of ChatGPT that allows visitors to use conversational language to search for information or products. 

Microsoft in May promoted the NLWeb technology and said Shopify, Eventbrite and TripAdvisor would use it. (The Information this week launched its own chatbot, trained on its articles, so subscribers can ask tough questions about technology and media businesses.)

Who knows if OpenAI can convince people to use a new browser. But if it gets any takers, that will increase the company’s understanding of how people work, shop and entertain themselves online. That could be valuable intel for a lot of different businesses it might launch someday.

Coincidentally, AI search app Perplexity on Wednesday launched its own web browser with similar goals, aiming to get a head start on OpenAI. My colleague Sri says the Perplexity browser, Comet, was developed using Google’s Chromium open source software. We’re guessing OpenAI’s will have similar bones, given the pedigree of the people creating it.

Still, we’re left with one question: If the web as we know it dies out, how important will a web browser actually be?

Here’s what else is going on…

Musk’s Grok 4 Hyperbole

There’s no doubt that Grok’s shocking-yet-not-shocking display of stupidity on Tuesday, when the chatbot embraced Adolf Hitler, overshadowed xAI’s launch of the latest Grok AI model on Wednesday night.

In an hourlong livestream that started around 9pm Pacific Time, Tesla and xAI CEO Elon Musk said the new model was the “smartest AI in the world” and “smarter than almost all graduate students in all disciplines simultaneously.”

Grok 4 and its bigger and better cousin, Grok 4 Heavy, are reasoning models that provide better answers when they use more computing power to tackle a complex question. OpenAI pioneered this type of model last fall with o1 and has followed that up with o3.

Musk said the Grok 4 models performed 12 to 17 percentage points better than state-of-the-art models from OpenAI, Anthropic and Google in taking Humanity’s Last Exam, a closely-watched evaluation of how models handle tough problems in math, science and other fields. It also did extremely well on the Arc-AGI tests, which measure models’ abilities to solve visual puzzles efficiently, in terms of their cost. 

Standardized benchmarks are one thing, but we’ll hopefully see more proof in the pudding when people actually use Grok 4 in its chatbot form.

Musk added that Grok 4 was a “super genius child that will ultimately outsmart you but you can instill the right values….encourage it to be truthful, honorable.”

“Honorable” isn’t exactly the word people used to describe Grok 24 hours earlier.

Why Hitler?

To catch you up, Grok, which is a core feature of the X social media app, debased itself on Tuesday by embracing Hitler in a series of public posts on the app in response to questions from users. The company acknowledged the problem and appears to have deleted some of those hateful responses, including the main one here, but they understandably went viral and even annoyed some of Musk’s most ardent fans. 

We’ve been here before. Grok, which began as an “anti-woke” effort by Musk, has had a series of public embarrassments this year, such as when it censored facts about Donald Trump and Musk himself or claimed there was a “white genocide” in South Africa.

As for exactly why Grok got “misaligned” about Hitler, as AI researchers would call it, we turned to someone who has worked closely with Grok models during their training. This person, who requested to not be named, suggested that there’s probably no problem with the “pretrained” base version of Grok that was trained on all manner of content to learn how the world works. 

Most likely, they said, the company had released a version of Grok on X that lacked some of the basic controls that AI model developers implement after the pretraining phase. Those controls come from a process known as supervised fine-tuning in which companies develop content guidelines to handle inflammatory topics.

Grok Goes ASMR

Even if the base model behind Grok was trained with problematic content that, say, adored Hitler, the supervised fine-tuning and other techniques AI developers use in the “post-training” phase would have smoothed out the Hitler-loving kinks, said the person familiar with how Grok models are developed.

In other words, it’s not a tech problem. That’s good for xAI, because it has a much bigger problem: how to turn Grok into a big business. We have yet to hear a smart answer to that question. For its part, Grok’s answer to that question is here.

Grok is trying to ape ChatGPT by selling subscriptions, and it could also sell ads (or include affiliate marketing links) for its free tier. First, though, it needs people to actually download the dedicated Grok app, or use the dedicated Grok section of X, and that could be a tall order. (As a reminder, xAI merged with X earlier this year.)

It could try to sell its models through an application programming interface, but that could be a race to the bottom on price.

Perhaps xAI is banking on romantic role-play. In the Grok 4 demo, researchers showed off a “natural voice” assistant version of the chatbot that went full ASMR. 

“It’s just you and me having a quiet chat,” the female Grok voice said in the demo in a soft, hushed tone. 

You get the idea.

Deals and Debuts

See The Information’s Generative AI Database for an exclusive list of private companies and their investors.

Microsoft announced its philanthropic Elevate program, which will spend over $4 billion in cash and AI technology to fund nonprofits and AI literacy.

Groq has talked to investors about raising between $300 million to $500 million at a $6 billion post-investment valuation, The Information reported.

MaintainX, which uses AI to help companies monitor their equipment, raised $150 million at a $2.5 billion valuation in a Series D funding round from investors including Bessemer Venture Partners, Bain Capital Ventures, D. E. Shaw Ventures, Amity Ventures, August Capital, Founders Circle Capital, Sozo Ventures and Fifth Down Capital.

LangChain, which develops software for creating AI applications, has raised $100 million at a $1.1 billion valuation in a funding round led by IVP, Forbes reported.

Augmentus Robotics, which develops robots for welding and surface finishing, raised $11 million in a Series A extension funding round, led by Woori Ventures, with participation from EDBI, Sierra Ventures and Cocoon Capital.

XerpaAI, which uses AI to help companies with marketing, raised $6 million in seed funding led by Ufly Capital.

Cobionix, which develops robots to perform ultrasounds, raised $3 million in funding led by TitletownTech.

OpenAI announced that it has completed its acquisition of Jony I’ve io hardware startup.

Perplexity launched Comet, an AI-powered web browser available with the company’s most expensive subscription.

OpenAI could release its new AI model, which will be free for users to download, as soon as next week, the Verge reported.

Andreessen Horowitz relocated its primary headquarters from Delaware to Nevada, citing Delaware courts increasingly scrutinizing the decisions of company directors.

Thank you for reading the AI Agenda Newsletter! I’d love your feedback, ideas and tips: [email protected].

If you think someone else might enjoy this newsletter, please pass it forward or they can sign up here.