Let me say upfront: I'm not denying that Google has had an enormous influence on AI -- starting with the Transformer -- and has made tremendous contributions over the years.
But watching their moves over the past year, I've come to feel that Google simply isn't interested in being number one in AI. I want to lay that out in some detail.
I see three main ways to advance AI:
- Serving AI products to the public on the front lines, and continuously improving them
- Contributing to academia by publishing novel research and open-source models -- not necessarily leading the front lines, but pushing the field forward
- Contributing at the foundation level through massive datasets and benchmarks
Up until last year, Google was leading in all three. But from where I stand now, while they may still lead in category 3, they've not only fallen behind in categories 1 and 2 -- they seem to have lost interest entirely.
1. Serving AI to the public on the front lines
What's the most important thing if you want to serve AI to the public on the front lines? It's service reliability. You either provide a stable API like ChatGPT, or you deliver fast speeds like Anthropic.
But the Gemini API is both unstable and slow. It's far too unreliable for production use.
On top of that, the hallucination problem isn't just some minor issue -- it's severe enough to call the service's fundamental trustworthiness into question. For example, even when code generation tools are available, the model refuses to generate 100 random numbers via code and then defends its own output as "plausible averages." This isn't something you can just laugh off.
And it's not just the LLM API performance. AI-powered services built on top of it -- search tools, IDE tools -- are at a level where practical daily use is simply not feasible.
The Antigravity incident
Google launched Antigravity on November 18th, and it was broken by elementary prompt injection within just one week. (See Simon Willison's analysis.)
Prompt injection is a powerful attack vector, yes. But given that Claude Chrome launched on August 25th and OpenAI Atlas launched on October 21st, shouldn't Google have delayed their release to harden their IDE/web tool's security?
OpenAI had even published official guidelines on prompt injection on November 7th. Why, by November 18th, had Google still not implemented even basic defenses?
"Other services get broken too"? Go ahead and try it yourself. See if the blog-level prompt injection that broke Antigravity works on Claude Chrome or OpenAI Atlas. I'm genuinely curious whether products that launched one and three months earlier get broken at that level. Claude Opus defends against MCP tool injection -- if anyone has an example of the same caliber of prompt injection getting through, I'd love to see it shared.
Security might seem like a minor concern, but when it's this sloppy, it looks to me like Google isn't serious about serving AI to the public.
Credit where it's due
That said, I think image generation models like Imagen are genuinely impressive. It's things like this that make me still consider Google "viable as an AI company." But I don't think they're at a level where they can compete with OpenAI or Anthropic.
2. Contributing to academia through research and open-source models
I won't belabor the point that Gemma has been essentially abandoned. Put it up against DeepSeek and there's nothing to say.
You might think that OpenAI and Anthropic don't contribute to academia because they don't publish many papers, but OpenAI contributes through GPT-family open-source models and the Alignment Research Blog. Anthropic goes without saying -- they're a solid pillar of academic contribution.
And Google? I think the research that Google Brain puts out at venues like NeurIPS is excellent. Proposing new paradigms like nested learning is no small feat. Among big tech, really only Meta attempts that kind of novel exploration.
But there's an important distinction to make here. OpenAI and Anthropic raising new problems with frontier models and creating entire new fields is a fundamentally different kind of contribution from trying out new models and conducting research in the traditional academic mode.
When OpenAI released the o1 preview -- the so-called reasoning model -- last September, the absence of a paper didn't mean there was no academic contribution. Quite the opposite: they pioneered the field. No matter how talented the big tech researchers or famous professors are, this kind of contribution isn't easy to replicate.
Ultimately, what top-tier AI companies contribute and what ordinary big tech or academia contributes are entirely different things. In this regard, I see Google as being on a different track from "what today's top-tier AI companies are contributing."
I'll continue with DeepMind below.
3. Contributing through large-scale datasets and benchmarks
I fully acknowledge that Google has made enormous contributions in this area for a long time, and that continues today. Models and datasets like AlphaEarth are things that no one else could produce.
But I personally view DeepMind not as being on the "AI frontier," but rather as doing AI research grounded in "large-scale datasets across diverse domains."
Take AlphaFold, for example. The life sciences dataset is incredibly valuable, and creating a bio-AI paradigm from it was remarkable. But that's closer to "applying domain-specific data to AI" than to "inventing new AI algorithms."
The same applies to other cases. With AlphaEarth, anyone who's done remote sensing will recognize quite a few familiar dataset names when reading through the dataset specs.
I can speak with confidence about 3D modeling and video models too: Google is not number one in the video generation model space. They've released a lot of open video datasets, but personally, I find the video model performance disappointing given how vast that data is. I'm genuinely puzzled why something like PlayerOne -- a video dataset paired with VR information -- didn't come out of Google first.