GPT-5 has now been out for more than a week. And a lot has happened  in that short time. Let’t take a look.

The Reaction to GPT-5

In the aftermath of GPT-5, two themes have emerged.

• The most prominent theme is that GPT-5 was a major disappointment. The progress was insignificant, or even a step back, and may even cause the “AI bubble” to burst. However, some of this was because the automatic router (which selects which model to use) wasn’t working great. It has since been improved.
• The second is similar to what I said last week, that GPT-5 is a step towards “AI for everyone” – a commercially successful application. And that’s the second emerging theme. Conor Grennan agrees (although he’s more effusive than me). SemiAnalysis makes the case that it’s a setup for OpenAI’s future: a business model for serving ads.

Ads are definitely coming to AI. Just this month, Google indicated that more ads are coming to AI Mode. Elon Musk said he’ll be adding them to Grok, and Meta envisions a future where ads are AI-generated and personalized for maximum effectiveness.

There were responses to GPT-5 from OpenAI, X.ai, Anthropic, and (sort of) Google.

OpenAI: GPT-4o is Back From the Dead

GPT-5 was supposed to be amazing and make all of the other models obsolete. But poor Sam Altman, he just can’t seem to make everyone happy. Turns out, we humans like AI that compliments us, that builds us up. AI that makes us feel smart. AI that likes us. AI that is sycophantic. OpenAI received massive criticism after an update turned GPT-4o into a suck-up, they became fearful of a repeat with GPT-5. So they made GPT-5 more matter-of-fact, less empathetic, and less flattering.

And many regular users hated it.

Those “hooked” on GPT-4o preferred a model that is worse on every benchmark because it made them feel good.So many people complained, that they brought GPT-4o back for those paying $20/month.

Sam Altman was conciliatory (he gave 4o back) even though he felt bad about it:

“It’s so sad to hear users say, ‘Please can I have it back? I’ve never had anyone in my life be supportive of me. I never had a parent tell me I was doing a good job.’ There’s young people who say things like, ‘I can’t make any decision in my life without telling ChatGPT everything that’s going on. It knows me, it knows my friends. I’m gonna do whatever it says.’ That feels really bad to me.”
– Sam Altman, OpenAI

My first reaction was horror that people are relying on a chatbot for emotional support and life guidance. But upon a closer look, the truth is more nuanced. My advice: don’t oversimplify the situation.

• Some users have a perfectly healthy relationship with AI and simply prefer 4o.
• Some users wanted the life advice that 4o gave, and they found helpful.
• Some users are unhealthily addicted to 4o’s sycophancy.
• And some “at-risk” users were distraught. Don’t assume all the noise is coming from just one of these groups.

“Fostering dependence is a normal business practice in Silicon Valley. It’s an aim coded into the basic frameworks of social media — a technology that has socially deskilled millions of people and conditioned us to be alone together in the glow of our screens. Now, dependence is coded into a…chatbot. Rather than simply lacking the skills to bond with other human beings as we should, we can replace them with digital lovers, therapists, creative partners, friends, and mothers. As the resulting psychosis and social fallout amassed, OpenAI tried to pump the brakes a bit, and dependent users lashed out.”
– Kelly Hayes, Some Thoughts on Chatbots, Addiction, and Withdrawal

xAI: The World gets Grok-4 for Free

To entice people away from GPT-5, Elon Musk’s xAI decided to make their model, Grok-4, free to everyone. Although it’s not unlimited access, these models are expensive to run and that could be a massive cost to xAI.

It’s the classic Silicon Valley playbook…when the market is young, don’t worry about profits, just get market share. Get people using your product now so you have enough revenue to squeeze out the competition. It worked for Amazon, Facebook, and many others. xAI is trying it again, as are others:

• Google is giving their AI suite to college students for free
• OpenAI is offering GPT-5 to the US Government for $1
• Anthropic followed suit, giving them Claude for $1

Incidentally (or not), it’s the same model drug dealers use. Give to new customers for free. Charge them once they’re hooked.

Anthropic: More Capable Coding

Anthropic increased the context window (how much text you can send to a model) for Claude 4 by a factor of 5, Now coders can send it up to 1 million tokens, or 75,000 lines of code, making it capable of handling significantly more complicated tasks.

Google: Smaller Models

Google released Gemma 3 270M (Gemini is the name of their closed models, Gemma is the name of their open models). This is a really small and efficient model designed to run on devices (like phones) for specialized tasks, not general use (like Gemini or GPT-5). On the benchmark they chose, it’s twice as good as Qwen 2.5 at half the size.

In Other News

World Models

In a sign of things to come, Google announced their Genie-3 world model (homepage, blog post), a significant upgrade to Genie-2. What is a world model, you ask? Think of it as a created world (like a video game) that you can move around in, but instead of being created ahead of time, an AI generates each new scene based on the direction that you move. Check out one of the links and realize that yes, all of this was created by an AI in response to user actions! Here are single frames of some of their demo worlds:

Genie-3, quadruples the image resolution of Genie-2 but the really big deal is the interaction time is increased from around 10 seconds to 1-2 minutes. The model is (usually) able to “persist” objects; you can look at something, then look away and look back, and the scene stays as it was. Because the AI model needs to re-generate each scene from scratch, persistence has been a big challenge.

This tech will be very useful to create training environments (for humans and for robots) as well as entertainment. The image quality (720p) and frame rate (24 fps) and interaction time (~2 minutes) are not good enough yet to rival today’s video games, but Genie-3 comes only 8 months after Genie-2! Both are still considered research projects and are not yet available to the public.

The New Browser Wars

Perplexity made an unsolicited offer to buy Chrome, Google’s browser. A Really Big Offer. $34B. To put that into perspective, Perplexity’s entire valuation is only $18B! Perplexity is already working on their own AI-powered browser, and Google says Chrome isn’t for sale, so what’s going on?

• There is a lawsuit in progress that claims Google has a monopoly
• If they lose that suit, they may have to divest Chrome to break up that monopoly
• Chrome has the largest market share of any browser

So Perplexity is betting that 1) Google fears losing and having to sell Chrome 2) at this early stage, there may not be other buyers and 3) getting access to the user base of Chrome would massively improve their market position overnight.

---

---

There is a lot that matters in the above news that I have already addressed. So here’s a summary:

• LLMs will continue to improve but not as fast as we’ve seen the past few years.
• As a result, the AI companies are shifting their focus to productization to drive adoption by making LLMs easier to use and taking a product approach
• LLMs aren’t the only game in town – we’ll see entirely new applications of generative AI technology in other ways, like AI agents and Google’s world models (Genie)
• We’re seeing a repeat of the dotcom days, where companies are spending massive amounts of capital to obtain market share
• In the hopes that said market will be worth billions when all this settles out
• It’s not clear what that market is yet. Is it AI companions? AI entertainment? AI agents? Control of the access to it all (the browser)?
• But it is clear that the market they’re going after is you, the consumer. They want your eyeballs, your browsing habits, your contacts, and most of all your wallet.

But, most of the dipping into your wallet comes later.

You’ve been warned.