I think the commenter was saying google squandered its lead ("goodwill" is how I would refer to it) in providing open and interoperable services, not the more recent lead it squandered in AI. I agree with your point that they've made up a lot of that ground with gemini 2.5.

Yeah you're right, I should have read their comment more closely.

Google's API's have a way steeper learning curve than is necessary. So many of their APIs depend on complex client libraries or technologies like GRPC that aren't used much outside of Google.

Their permission model is diabolically complex to figure out too - same vibes as AWS, Google even used the same IAM acronym.

> So many of their APIs depend on complex client libraries or technologies like GRPC that aren't used much outside of Google.

I don't see that dependency. With ANY of the APIs. They're all documented. I invoke them directly from within emacs . OR you can curl them. I almost never use the wrapper libraries.

I agree with your point that the client libraries are large and complicated, for my tastes. But there's no inherent dependency of the API on the library. The dependency arrow points the other direction. The libraries are optional; and in my experience, you can find 3p libraries that are thinner and more targeted if you like.

I feel like the AWS model isn’t all that hard for most of their API’s. It’s just something you don’t really want to think about.

I sometimes feel the complexity is present by design to increase the switching cost. Once you understand it and set it up on a project, you are locked in, as the perceived cost of moving is too high.

This is bizarre to read. gRPC is used _widely_ outside Google. I'm not aware of any API that requires you to use gRPC. I've never found their permission model to be complicated at all, at least compared to AWS.

Gemini 2.5 Pro is so good. I’ve found that using it as the architect and orchestrator, then farming subtasks and computer use to sonnet, is the best ROI

You can also farm out subtasks to the Gemini Flash models. For example using Aider, use Pro for the "strong" model and Flash for the weak model.

I’ve tried flash. Sonnet with computer use has been a blast. Limited anecdata led me to 2.5 Pro + 3.7 Sonnet, but this all moves so fast that it good to reevaluate regularly.

OOI what's your preferred framework for that managing agent/child agents setup?

I use Roo Code. It’s very good.

The models are great but the quotas are a real pain in the ass. You will be fighting other customers for capacity if you end up needing to scale. If you have serious Gemini usage in mind, you almost have to have a Google Cloud TAM to advocate for your usage and quotas.

We have moved our quota system to Dynamic Shared Quota (https://cloud.google.com/vertex-ai/generative-ai/docs/quotas) for 2.0+ models. There are no quotas in DSQ. If you need a guaranteed throughput there is an option to purchase Provisioned Throughput (https://cloud.google.com/vertex-ai/generative-ai/docs/provis...).

The thing is that the entry level of provisioned throughput is so high! I just want a reliable model experience for my small Dev team using models through Vertex but I don't think there's anything I can buy there to ensure it.

While we are talking about quotas, can you maybe add an easy way of checking how much you've used/got left?

Apparently now you need to use google-cloud-quotas to get the limit and google-cloud-monitoring to get the usage.

VS Code copilot managed to implement the first part, getting the limit using gemini-2.5-pro, but when I asked gemini to implement the second part it said that integrating cloud-monitoring is too complex and it can't do it !!!!