Grok Is Three Years Old and Somehow That's Supposed to Mean Something

Elon Musk posted something this week about xAI being “only 3 years old,” apparently as a defence of Grok’s current standing in the AI market. The implication being: give it time, it’s just getting started.

This is a strange argument to make in a field where three years is basically a geological epoch. The models people are actually using, Claude, GPT, Gemini, have each had major releases in the last few months. If anything, three years of runway with Musk-level capital expenditure and not being at the frontier is a concerning sign, not a reassuring one.

The version numbering situation is genuinely funny, though. Someone in the thread pointed out that the chaos of Grok 4.3 versus Claude 4.7 versus GPT-5.4 nano/mini/whatever means nobody actually knows what generation anything is anymore. This is not an exaggeration. I follow this stuff reasonably closely and I cannot tell you with any confidence which model version is “current” for any given lab at any given moment. The versioning has become pure marketing theatre. At least Elden Ring’s DLC had a coherent naming scheme.

There’s a real split in the discussion about why Grok underperforms. Some people say the model itself isn’t competitive. Others say the pricing is the problem: roughly 50% more per month than the alternatives, for a product that isn’t clearly better. Both of these things can be true simultaneously, and apparently are.

But the third thing, the one that kept coming up, is harder to quantify. A meaningful chunk of potential users have simply decided they won’t touch anything with Musk’s name on it. That’s not irrational. When a product is entangled with a public figure whose behaviour has become genuinely alarming to a large portion of the population, some of that sticks. You can argue about whether that’s fair to the engineers building the model. It probably isn’t. But it’s also just how product reputation works, and pretending otherwise doesn’t help.

The thing I find genuinely interesting, underneath all the dunking, is the capital question. xAI is reportedly burning through something in the range of $30 billion annualised. Anthropic is doing more with roughly a fifth of that. Meta has spent enormous sums and the returns have been modest. This suggests the money alone isn’t the variable that matters most. Architecture, talent, and the absence of someone constantly interfering with the product apparently count for something too.

The Tesla data angle keeps getting raised as a potential trump card: all that real-world driving telemetry, the physical world training data that nobody else has at that scale. Maybe. But it’s been “maybe” for a while now, and the people who work on xAI’s actual models seem to have largely departed. You can have the best ingredients and still produce a bad meal if you keep rotating the kitchen staff.

The Cursor acquisition might change things for coding tasks specifically. Or it might not. I genuinely don’t know, and neither does anyone else pretending to have a confident take on this.

What I keep coming back to is how strange it is that we’re at a moment where the most visible AI products in the world are being evaluated partly on the political and personal behaviour of their owners. That wasn’t supposed to be how this went. The technology was supposed to speak for itself. Instead we’re debating whether to use a product because of what its founder did at a public event, and that’s not a trivial concern to wave away.

Three years is young. It’s also long enough to have figured some things out. The question is whether the current trajectory leads somewhere better, or whether the talent drain and the reputational weight become structural problems. I don’t have a clean answer. I’m not sure there is one yet.