This week we have a new enterprise agent from an unexpected source, more data that AI is providing value but is also impacting hiring, AI browser battles, and some Anthropic updates.

A New Enterprise Agent

DeepL (a German company) announced their general-purpose enterprise agent. This was a surprise (at least to me) because prior to this, DeepL was focused primarily on machine translation, not agentic AI. Their approach is based on letting AI control a computer through a browser (first popularized by OpenAI’s “computer use” mode of ChatGPT). The theory is that, since a human can do almost anything through a web browser, if an agent can control a browser, it can too.

Unlike OpenAI which is oriented towards consumer use, DeepL says their agent is specifically targeting tasks in the enterprise. They claim that their agent provides “clear productivity gains” for “the type of busywork that is essential to everyday work (but not very exciting for the people doing it).” That’s a benefit, sure, but not exactly a strong ROI.

AI and Browsers

DeepL’s announcement rests on the belief that a general purpose agent might come through AI using a browser. If true, that’s a big deal…and there are some besides DeepL who think combining AI and a browser might be the future.

Perplexity is one. Remember how they recently offered to buy Chrome from Google? They were anticipating a court ruling might force Google to divest the browser. Turns out, the judge was more lenient and said they don’t need to. So Chrome will remains Google’s.

Atlassian is another. Like Perplexity, they wanted to buy a browser…and succeeded, acquiring The Browser Company which is building an AI-powered browser from the ground up.

AI Delivers Value…

In the ongoing debate about whether AI is delivering value, a new report says yes, at least for some deployments. The debate will continue because similar to the (flawed) MIT report from a few weeks ago that said 95% are failing and 5% are super valuable, some companies are seeing a big impact, while others aren’t.

…and is Reducing College Hires

This week came another analysis showing that entry-level hiring is decreasing, and while we can’t say 100% it’s because of AI, it seems increasingly likely.

Anthropic Wants Your Data for Training…

Until now, Anthropic has been alone among the LLM providers in promising that it wouldn’t train its models based on your prompts. They are now giving (non-business) users the option to train Claude on your interactions. This brings them on par with the others in the industry. While this is a step away from privacy, I think it’s simply an acknowledgment that using real data in this way is an important method to improve their models.

…and Settles a Copyright Lawsuit

Anthropic also settled a copyright lawsuit that claimed they trained Claude on pirated books for a reported record sum: $1.5 billion. As I understand it, this is a settlement offer that still needs to be accepted, but if it does go through it could set the tone for the many other lawsuits related to copyrighted works. Pretty much every AI company is the target of at least one copyright lawsuit. Their rulings could have a significant impact on training LLMs.

---

---

Two observation this week. One is collective; there is still a great deal of uncertainty about generative AI. It’s simply too early to know how this will end. The early data is bifurcated (in the case of value), correlation not causation (in the case of jobs), and uncertain (legal challenges).

I believe AI will be transformational, but it’s harder to predict exactly what will be transformed. It may replace some things, but it’s not on track to become some kind of universal intelligence that takes all jobs. It’s a tool that is good for some things, and not good for others. It doesn’t do everything. We shouldn’t try to make it do everything, nor should we claim that it does.

Which brings me to my second point: is computer use is an avenue to a general-purpose agent, as DeepL believes? It is true that you can do a lot of things through a browser. But there are a lot of things you can’t do through a browser, and generative AI is not you.

Browsers Can’t Do Everything

Yes you can do a lot with a web interface. But we live in a real, embodied world. Yes, we have put a lot of things behind screens; we even hold meetings without being in-person. But no matter how realistic those experiences become, they are still a façade, a replica, an imitation of the real world. 

There are many things that you can’t do through a screen. You can’t hug someone. You can’t feed a baby. You can’t assemble furniture or a LEGO set. You can’t walk your dog or pet your cat. You can’t draw, paint, sing, or play an instrument. You can’t go for a run, a ride, or a swim. You can’t sit on the beach or hike in the woods. You can’t play pickleball or golf. You can’t, as my kids are fond of saying, touch grass.

These activities may not all drive economic value for companies. But that doesn’t mean they have no value. Doing things, and doing them together, has immense value, and nothing physical was ever built, shared, or used through a screen.

Generative AI Isn’t You

Just because you can do stuff through a browser doesn’t mean AI can. These models are great at many things but that doesn’t mean that they can do everything. They don’t reason; you do. They don’t have a model of the world, you do. They don’t have goals and purpose; you do. Therefore it seems optimistic to expect that a general-purpose agent will be capable and reliable enough to “just do it.”

Kudos to the companies releasing general-purpose agents to the market to help relieve people of busywork. But scrutinize them carefully: are they really general-purpose, or are they only handling simple or repetitive tasks? Summarizing a few documents is a far cry from designing a building, writing Shakespeare, or building an iPhone.

General-purpose does not mean universally capable. It’s not even close.