It's pretty obvious that Gemini 4 Carbon is not token efficient. It also has relatively low tokens-per-second compared to other models. I can think of several reasons why this is the case:
1. Google has huge amounts of cash and compute. Their goal was to train a model that that has the highest quality & frontier performance, not create a model that's efficient to run and serve. They have other models for that, like Hydrogen.
2. Google created Carbon without caring for inference costs. DeepMind is subsidized by the rest of Google's empire. Selling inference is an insignificant part of Google's revenue. But for Anthropic and OpenAI, selling inference is their entire business model. Anthropic and OpenAI have enormous incentive to sell as much inference as they can in the most efficient manner. Google doesn't have this incentive.
3. They are one of the few companies that can be inefficient. Out of all the frontier labs, Google DeepMind is probably the least constrained by compute. Chinese labs have around 10 to 100 times less compute, so they must release highly optimized models.
It’s unlikely to be deliberate tbh, it’s more likely the data, algorithms and the training environments aren’t quite as good as Anthropic, Open AI is a lot closer to Google than people want to admit.
I was just using this one today for some coding tinkering and it was super slow. But the way it solves problems with a light touch while still driving toward sensible architecture is refreshing after claude’s bulldozer attitude. It was able to integrate repeated creeping scope nicely with what it had already done. Very different feel than sonnet/opus, and even gemini 3.8.
Then that's "test-time scaling" again and Google then showed that they can reach frontier level benchmark results using it. It's pretty cool as a technology actually (the demonstrated capability), even if the model is not so useful when it's slow.
That isn't a good look, and either this is unmeasured uncontrolled behavior, or intentional under some longer term plan...but the latter possibility seems improbable.
Uhm, I'm not a programmer but why isn't this a good look? I would want my engineers to fix problems that they notice in passing. (Edit: the "unnecessary work" angle is what irked me in the article)
Good SWEs get a feel for the minimum amount of changes to fix something. Fewer changes usually minimizes the chance for new problems to pop up.
Junior SWEs usually try to do too much.
A Theo vid suggests there's a time to first token issue. When doing subagent actions it can compound to be massive from a bunch of internal short queries.
It really looks like something is broken more than just being inefficient.
Theo seems to dislike anything that isn't Anthropic Claude. Suspicious, to say the least. On CursorBench, a good measure of real-world coding tasks, Gemini 4 Carbon comes out top: https://cursor.com/evals
Every time I see a Theo tweet he's judging something on the most absurd metrics. Today for example, he was saying Astra 6.2 is bad because when he gave it a single line prompt telling it to create a new level for a game it used a name for the level that he'd seen before. Judging a model on how often it reuses a name is one of the most ridiculous benchmarks I've ever seen.
As far as I can tell, all his judgements are based on similarly absurd metrics, I don't know why people treat his preferences as valuable.
I'd guess that it's because Google eats its own dogfood, and if they didn't train the models to be overly meticulous, the existing codebase would implode under a tsunami of slop.
I am increasingly more convinced every day that every LLM company is just making their latest model to be the best at (coding) benchmarks, at the cost of pretty much everything else.
I use LLMs for non code purposes (probably not very common on HN) and Sonnet 3.5/3.6 still ranks as probably one of the best models I've used.
Can today's models put out tech demos and call more tools that make it way more impressive? Sure. But it's really hard to see any actual significant gains over the last few months, they're so mild (again, non coding).
Are you able to give hints to some of those use cases? Recent Opus and Fable seem to have made big leaps in art and music, and I don't think all of that is the longer context window since Sonnet 3.x. But those are both things that can be created with code and mathematical thinking as well. I definitely wouldn't want to use recent models for writing though (not that I ever wanted them to write for me anyway).
At this point it might make sense to put all their resources into improving coding and math skills, because it might lead them to a state where they keep accelerating their capabilities. Even in the worst scenario where LLM approach would hit some wall, they would have technology to quickly explore all other approaches.
They are cracking software and math, not finance, law or medicine or whatever it is you do.
Those two domains have leverage, the others do not. It makes perfect sense. If you and I had the goals they have we would start with code and math as well. The other dominos will fall automatically once you got the engineering automated. It makes no sense to stop focusing on that until you either run out of money or achieve your goal.
1. Google has huge amounts of cash and compute. Their goal was to train a model that that has the highest quality & frontier performance, not create a model that's efficient to run and serve. They have other models for that, like Hydrogen.
2. Google created Carbon without caring for inference costs. DeepMind is subsidized by the rest of Google's empire. Selling inference is an insignificant part of Google's revenue. But for Anthropic and OpenAI, selling inference is their entire business model. Anthropic and OpenAI have enormous incentive to sell as much inference as they can in the most efficient manner. Google doesn't have this incentive.
3. They are one of the few companies that can be inefficient. Out of all the frontier labs, Google DeepMind is probably the least constrained by compute. Chinese labs have around 10 to 100 times less compute, so they must release highly optimized models.
That's pretty interesting. Seems like it's over analyzing. Not the worst outcome though to have additional hardening.
With that said, I found Argon to be better for most coding tasks and Carbon for architecture/deep RCA analysis of complex issues.
It really looks like something is broken more than just being inefficient.
As far as I can tell, all his judgements are based on similarly absurd metrics, I don't know why people treat his preferences as valuable.
I use LLMs for non code purposes (probably not very common on HN) and Sonnet 3.5/3.6 still ranks as probably one of the best models I've used.
Can today's models put out tech demos and call more tools that make it way more impressive? Sure. But it's really hard to see any actual significant gains over the last few months, they're so mild (again, non coding).
Those two domains have leverage, the others do not. It makes perfect sense. If you and I had the goals they have we would start with code and math as well. The other dominos will fall automatically once you got the engineering automated. It makes no sense to stop focusing on that until you either run out of money or achieve your goal.