Interview with Google Cloud CEO Thomas Kurian on Google's Enterprise AI Strategy | Hacker News Reader

Interview with Google Cloud CEO Thomas Kurian on Google's Enterprise AI Strategy

stratechery.com

46 pointskjhughes28 comments

Spending time trying out Google's new AI products you can see the astonishing amounts of enormous bloat around every single service it offers

Sergey shouldn't be pair programming right now he should be doing the hard work with Sundar of bringing some semblance of coherence to their offerings but it seems like the organizational dynamics are too challenging for them at this scale.

> you can see the astonishing amounts of enormous bloat

In its category, hyperscaler public cloud, GC appears consistently leaner compared to AWS Azure

I understand your frustration, but you need to realize that if you are paying nothing and you are complaining, then you are simply criticizing someone who are not serving you. Like criticizing Biden being a bad president from a Chinese citizen.

This type of criticism carries negligible amount of potency and relevancy.

This might be an unpopular take but Google is fast regaining LLM platform dominance that OpenAI / MSFT took away for a year.

GPT-5 release needs to blow people's mind for MSFT to rejuvenate the Azure AI conversations because momentum is shifting again to GCP.

P.S. Many customers aren't very happy with AWS Bedrock but they should be ok as long as Anthropic & Mistral doesn't get annihilated by Gemini / GPT-5 in the next year.

Only recent news is keynote demoware yesterday and a model release that is 5 _requests per minute_ and was released 2 days ago. Gemini Ultra still doesn't exist, and needs 32-shot CoT to beat GPT-4 from March 2023 5-shot.

Impossible for it to be fast regaining LLM platform dominance, for those reasons, and Google had no LLM platform before OpenAI/MSFT.

Annihiliated, sure. They can coexist. This isn't an olympic swimming competition.

I don't see any discussion of their deprecation strategy. What's killing GCP is not that they don't have the latest toys, it's that *NO ONE TRUSTS THEM ANYMORE* [1].

Why would you build anything on GCP when there's no guarantee the GCP thing you're using won't be the next Stadia in 18 months? Otherwise, you'll just use them like you'd use DigitalOcean, in which case they'd be commodified.

For the CEO not to be laser focused on this issue shows they're entirely missing the point.

1. https://steve-yegge.medium.com/dear-google-cloud-your-deprec...

I actually heard somebody say gcp was going to be shuttered, didn’t verify, and just assumed it wasn’t going to be around for long. Now I assume what I read was nonsense, but it’s worth noting how non-suspicious of it I was when I read it

The @killedbygoogle account is funny and also a complete menace sometimes. I wish they'd drop the sarcastic tweets, they've won thoroughly, no need to be an infohazard

Not to mention they were out in their Paris availability zone for a month last year.

Google Cloud DOES have a deprecation "strategy". They publish it in Google Cloud Terms of Service.[1] In short, any service that is listed as "Generally Available" (i.e., not alpha/beta) is guaranteed to have at least 12 months of notice prior to deprecation. And services that are in alpha/beta are clearly marked as such in the documentation and UI, and require explicit steps to use via the command-line or API.

Additionally, if you literally Google "Google Cloud Deprecation strategy", a statement on Google Cloud's deprecation policy and links to the Terms of Service and Google Cloud's service lifecycle is the very first result.[2]

I've been a user of Google Cloud for a long time, and the company I'm with has a yearly GCP bill in the millions. While Google Cloud does have its frustrations (e.g., their various price hikes, and the architecture of their network infrastructure that makes them more vulnerable to cross-region outages), Google Cloud being "deprecated" is not a thing that I worry about.

Can this tired meme please die already?

[1] https://cloud.google.com/terms#:~:text=(d)%C2%A0Discontinuat...

[2] https://www.google.com/search?q=Google+Cloud+Deprecation+str...

I didn't say they don't have a deprecation strategy. My problem is that they do have one, and they're very happy to use it.

12 months of notice is laughable for enterprise customers. From planning to getting something rolled out at scale, an enterprise will have eaten 2 quarters of time.

More than full deprecation, we often have API deprecation from google services, increasing our maintenance burden.

Say what you will about AWS or Microsoft products, if you build something on those platforms you know it'll be supported for a long time.

Google fundamentally values their own code above the code of people using their platform. But good platforms do the opposite - they keep the technical debt on their side of the fence to make sure you don't have to think about anything if your code runs on their stack.

> Say what you will about AWS or Microsoft products, if you build something on those platforms you know it'll be supported for a long time.

AWS literally has the same terms.[1]

Azure claims that "certain" Azure services have up to three years of notice; otherwise, it similarly offers a one-year notice period like GCP and AWS.[2]

[1] https://aws.amazon.com/agreement/#:~:text=Notice%20of%20Chan...

[2] https://learn.microsoft.com/en-us/lifecycle/policies/modern#...

Again, you're not getting the point. It's about behaviour. How often have you seen AWS deprecate stuff?

The BOTO API for S3 is an ugly mess, but you can still run BOTO2 code from a long time ago. With Google I guarantee you they'd have killed off the old API in a similar situation.

Similarly, how often did you see Windows deprecate core Win32 API calls? There's no SLA that says Windows 11 should run XP era programs. It just does, because they don't want to piss off people who depend on them.

As a user, I'm happy with deprecation happy products (hi Apple!). As an enterprise customer? Nope. We're paying you to ease our problems, not the other way around.

As long as GCP doesn't take this concern seriously, they'll be relegated to being a commodity provider, along with digitalocean and friends. We'll use them when they're the cheapest, and only for stuff we can migrate in 24h.

Guess what? If I'm looking for the cheapest cloud for VMs and storage, there's much better than GCP out there.

And as long as their freaking CEO doesn't get this they won't come out of the hole they dug themselves.

I'm still using GCP services that were deprecated years ago (e.g., legacy VPC, classic non-HA VPNs, decommissioned images, etc.). In some cases I can't create new instances of these services, but those that already existed continue to work.

Similarly, we have infrastructure created with very old versions of Terraform whose GCP providers continue to work despite being years old. Indeed, I can't recall a time when Google Cloud changed a public non-beta API in a compatibility-breaking way.

I'm sure we can continue to go back-and-forth with our own personal anecdotes, but they are just that. If a more generous deprecation policy was actually something these cloud providers were willing to stand behind, they'd add that to their terms of service. Azure has some extended lifecycle services (to their credit), but otherwise, GCP's lifecycle policies are in line with the rest of the industry.

Yeah, no. I've ridden the Google train enough times to know how often products get killed which I rely to keep my business running.

Google would need to be, say, half the cost of AWS to get my business.

Toss in hiring head of product from Oracle? No one in their right mind would use Google Cloud

I am honestly not sure how any business can take Google’s AI products seriously given the long history of products being shut down and the recent issues with Gemini.

I know someone is going to claim that Google Cloud is different from the rest of Google and doesn’t deprecate things in a hurry, but that’s not really reassuring. Enthusiasts adopt useful products at work because they’ve had a great experience personally using those products. The reverse is also true. If they’ve had a bad experience with a brand personally, they’re less likely to trust that same brand professionally.

With AI specifically, there are other bigger issues. Most businesses that could be a customer of Google’s enterprise AI products serve customers from all parts of the political spectrum and from all backgrounds. If you’re a large company you probably also have customers in other countries that have very different values from someone living in the Bay Area. Google’s culture is known to be heavily influenced by the politics of its employees and that came out in Gemini. Which company wants their own products or their own customers to be subject to the potentially offensive political whims of Bay Area Google employees by depending on Google’s enterprise AI? There are better choices out there that are safer bets, that don’t come with Google’s biases. It’s not just that Gemini isn’t sufficiently better than any other platform to stand out, but that it is demonstrably worse due to this issue.

If Google wants to change this perception, its leaders need to show backbone, shut down the overly political culture and make it neutral with respect to the political spectrum and operate with a totally different mindset. They need to do this across all products, both for consumers and enterprises. They will never earn the trust of the world unless they do this in an obvious and visible way.

PS: The “open-washing” in this interview is laughable in how flagrant it is. The only actually open approach is to be open source. For AI, that means having open source training code, open training data sets, total transparency on how training data is pre-processed, etc. If a third party cannot reproduce the weights, it isn’t open.

Let's see the products. I'm so tired of hearing Google talk a big game about AI without delivering.

a) I think they have in a lot of ways, that are maybe not super flashy. Vertex AI offers more API/functionality/product than anyone else, by a mile. I think in a year or two this might be more important than the quality of model (not that the model won't have to be good, but all of them will be good enough to solve most peoples business problems)

b) Gemini 1.5 Pro public preview with 1 mio tokens is out now (including API access). Has anyone benchmarked it?

> not that the model won't have to be good, but all of them will be good enough to solve most peoples business problems

In a year or two? Mate I think you need to step back from the AI hype for a moment. We're not even close.

When I write that tooling will be more important in the future, I am not making any crazy claims about the future of AI. My claim is: It does not require much more AI than we currently have to solve most business problems. But it will take better tooling.

I don't think AGI is even remotely required to solve most peoples business problems. Better tooling is.

Maybe we also have different ideas about how mundane most peoples business problems are.

Having tried all 3 (Gemini, chatgpt and Claude) I'll agree with GP that 1-2 years seems conservative if anything.

I don't think AI will be won on the model but the tooling. Google is arguably ahead.

The models I'd say it's a toss. But it doesn't matter as before the year is up all of them will release new versions one-uping each other. It's a race to the bottom. They either collude on price or most will go bankrupt.

Edit:typo

#1) That's absolutely, harrowingly, shockingly, untrue. They're just rolling out system messages! I actively integrate with Mistral, OpenAI, Gemini AI API, Vertex AI API, and Anthropic's API and Google's are the weakest.

#2) It's Gemini 1.0 with a 1M context. It's not very good, it works like OpenAI long context for me (pulls out salient text, but it doesn't seem to be 'reasoning' over all of it)

It’s almost like they hire exclusively for your ability to leetcode rather than trying to hire people who can produce software

Maybe the ai teams were too busy keeping up to date with their tree reversals to read those textbooks on their nights off

Gemini is just as good if not better than GPT-4?

Extraordinary claims require extraordinary evidence... :-))

You can use the product yourself or check one of the many benchmarking papers

Benchmarking papers are borderline useless. Goodhart’s law is a thing when a measure becomes a target it ceases to be a good measure. That’s why I keep coming back to GPT4. It has the most “magic” and “mind reading ability”. Try to quantify that.

1. Train an ai on everyone’s data

2. Try to replace everyone, and their companies too, but do it quietly

3. Try to sell ai tools to make those companies dependent on you, until your super ai can replace them completely

Reply on news.ycombinator.com