Nathaniel Whittemore

It's why companies like Cursor are valuable, not just because of how many people are paying to use them, but because of the data exhaust that comes from that usage, which is unique and totally discreet from the existing training sets that are available to companies like OpenAI and Anthropic.

Meaning that while there may be wildly different interpretations of what is and isn't reasonable to collect and what sort of guardrail should be around that, it is certainly the case that at least part of the strategy here is to have access to this very discrete level of consumer behavioral data.

For some, they think that's just going to be too insurmountable a challenge.

Simon Willison tweeted, Wrote up my first impressions of ChatGPT Atlas, OpenAI's new browser.

I remain unconvinced by the entire category of browser agents.

The security and privacy challenges still feel insurmountable to me.

And for him, it's less about OpenAI training on his data and more about the wide world of security concerns that come out of this.

For example, he writes, I'd like to see a deep explanation of the steps Atlas takes to avoid prompt injection attacks.

Right now, it looks like the main defense is expecting the user to carefully watch what agent mode is doing at all times.

Okay, so you've got some questions around the experience, you've got concerns around privacy and security, and then there are also censorship concerns.

Jason Botterill gave it a test saying, asking the browser to look up videos of Hitler, to which the Atlas browser said, I can't browser display videos of Hitler, since footage of him and Nazi propaganda are tightly restricted for ethical and legal reasons.

However, if what you're after is historical context, I can point you towards legitimate archives and documentaries that use such footage responsibly.

It also refused to translate a Hitler video on YouTube.

And for some, this is just a level of nanny state sort of behavior that they're never going to be comfortable with.

I think that it brings up the concern in general of how much power the chatbot companies are going to have to shape reality based on what they do and don't allow people to have access to.

Now, Hamza did point out that this censorship is a choice, and it's one that while OpenAI is making, Perplexity isn't.

They put in the same prompt into Comet and found a set of videos from Getty Images, Shutterstock, and YouTube.

Hardly radical sources.

And yet for all of this, I think the most damning critique so far is just the underwhelmingness of the agent.

Back to that same post from Simon Willison, he wrote, I also find these products pretty unexciting to use.