Google launches an end-to-end AI platform
techcrunch.com
techcrunch.com
12.1 The following terms apply only to current and future Google Cloud Platform Machine Learning Services specifically listed in the "Google Cloud Platform Machine Learning Services Group" category on the Google Cloud Platform Services Summary page:
Customer will not, and will not allow third parties to: (i) use these Services to create, train, or improve (directly or indirectly) a similar or competing product or service or (ii) integrate these Services with any applications for any embedded devices such as cars, TVs, appliances, or speakers without Google's prior written permission. These Services can only be integrated with applications for the following personal computing devices: smartphones, tablets, laptops, and desktops
[1] https://cloud.google.com/terms/service-terms#12-google-cloud...
For instance, as GigaMart, if your ML system finds that you are going to need widget-x in region y two weeks ahead of time, you can plan for that including the logistics and inventory.
Hospitals? I'm sure there are lots of things in the day-to-day operations that correlate with patient outcomes that currently go unnoticed.
Etc etc. It's not about creating a huge new invention, mostly it's about creating improvements to operations of current incumbents.
I would say if you want to be a provider, be a provider, don’t treat your customers like potential enemies.
This platform focuses not on the this-AI-is-magic-and-can-solve-everything like many AI SaaS startups announced on Hacker News, but focuses on how to actually integrate this AI into production workflows, which is something I wish was discussed more often in AI.
The announcements here, including AutoML Tables (which is coincidentally similar to my own Python package: https://news.ycombinator.com/item?id=19492406), the new BigQuery BI tools, and the new Google Sheets integrations, make me a very happy data scientist.
I'm taking the rest of the week to figure out how to integrate everything announced into my team.
We’re a startup that helps with AI deployment management (our only focus is production AI - embedded everywhere) and I wonder how many companies are who’d be interested in their end-to-end offering.
From Amazon there is no automated ML solution (autoML that you can train models with few clicks)
In brief, my experience was quite frustrating. First, getting my dataset to the cloud required quite a bit of manual labor. Uploading my 25GB of images on my 40mbit/s wasn't really ideal so I ended up spawning an virtual machine on GCP and downloading directly from Kaggle. Unzipping the files and writing it to Google Cloud storage with some terminal commands.
Furthermore, to get the label data into AutoML I had to write some generation script that generated the exact format CSV that AutoML requires - which was hidden in their documentation somewhere with no mention in the AutoML environment itself.
Nothing too cumbersome but generally not a very user friendly experience, or something I wish to repeat many times if I get a new/different dataset.
Ultimately, when the data with labels was in the model started training. Then I found, that I didn't really have the tools and information to assess the model performance. They did a decent job of characterizing model performance through precision recall graphs and displaying incorrect predictions but that didn't really satisfy me. I was interested in getting more details about where it was misclassifying images, specifically how classification performance was distributed across the 28 classes the model was predicting (in a multi label context).
This is the point where I think the downside of working with a platform such as AutoML starts showing. I tried reaching out to someone about gaining more insight into model performance by opening a ticket, since there was no phone number. After a couple of days I finally received an email from a product representative that told me that for any assistance I should contact one of their local cloud partners.
These are third party vendors that typically assist companies in deploying cloud based applications in the GCP. However, after calling two of these companies that were highly recommended by Google's vendor page I was told that they don't have any experience with AutoML and that I was on my own. The other company didn't reply at all.
In my view, choosing a product such as AutoML - for a company that is serious about adopting AI to improve their business - is currently not a good path (yet). And I see this space as being wide open for competition with current solutions not cutting it from my point of view.
This is classic commoditization of your complement. On one hand, Google is pushing to centralize the integration, data management and computing platforms for machine learning, so that these things become as much of a commodity as possible.
On the other side, they are offering massive compensation packages or acqui-hiring as much AI talent as they can, not really because they have useful work for these folks to do, but to artificially reduce the supply of statistical algorithm talent, making their consulting and pre-packaged AI solutions go up in value in a manner that is pretty much the same as De Beers pushing silos full of diamonds to keep diamond prices artificially high.
This is very much the opposite of democratization, and my advice to anyone considering services like this or like Amazon’s out of the box models, don’t do it!
If you think it’s going to save you money over paying competitively to get your own in-house machine learning staff, you’re wrong, and you’re going to waste probably hundreds of thousands of dollars before you learn you’re wrong.
>> It’s so disingenuous for Google to brand
>> these efforts as “democratizing AI.”
Cloud vendor lock-in and proprietary hardware, software, _and_ datasets is not in any way "democratizing" anything.
>> not really because they have useful work for these folks to do
When he raised the issue to managers that he wasn’t working on anything related to his specialization (deep learning for NLP) and this made him unhappy, the response was essentially, “Get in line.” I’ve heard similar stories about Google from a former boss who had been a long time manager in Google.
You essentially get paid super well to be put out to pasture so that your skill isn’t being used by other companies (leading to more demand for Google’s managed AI solutions).
In order to get career-developing work, you have to play political games or get hired in a non-standard way, like acqui-hire or poached, where you can negotiate your projects as part of your hiring conditions. Eventhen it will probably only be respected for a short time while it’s convenient for Google, and they’ll find a way to manage you out of that situation when they want to.
There are some tremendously talented AI engineers in places like Google. Some of them create awesome products and tools. A bunch of others sit around and atrophy working on dumb shit locked in golden handcuffs just to ensure they’re not on the market and able to help a company build things in-house more cheaply / more optimally than if they needed to buy it through some managed services through Google.
Discussing stuff with your manager is utterly pointless because your interests aren't really aligned. You want to do something else. Your manager wants you to do whatever you're doing now because finding a replacement for you is a bit of a pain in the ass. She gets no brownie points if you leave.
It is true that Google has a ton of PhDs who just copy one protobuffer into another and browse memegen all day while earning half a million dollars a year. But they also have a ton of PhDs who do meaningful work, too. It's not really Google's problem that someone can't be bothered to look around and find something meaningful for themselves to do. Or to be more exact, it is a problem _for_ Google, because there are a lot of people who can be deployed in higher leverage occupations, but not one that Google itself can solve, because one of the main tenets of how they operate is _nobody tells you what to do_. You're supposed to figure it out on your own. A lot of people can't deal with that.
Just move your shit to another desk and join whatever team you want with out talking to your manager or hr or updating your goals?
Yeah... nah.
Don't know about Apple or Amazon but Google/FB are like that. If you're very senior, a few months (no more than 6 no matter how senior) delay might be imposed so that you hand off your stuff, if you're less senior, a few weeks is usually enough. It is also expected (at Google, don't know about FB) that you'll stay on each team for at least a year, and that you'll wind down your obligations in an orderly fashion, which I think you'll agree is not unreasonable.
But _nobody_ will force you to do work you really don't like to do. In contrast, at most other companies it's easier to get a job _at another company_ than to move to another team.
I did something like this myself. I found the team I liked, talked to the team a bit about the work I'd be doing, and at my next one on one with my manager I told him I'd be leaving the team in two weeks.
You can certainly find yourself a better team fit and organise to move to it proactively, that’s true of any organisation, with greater or lesser degrees of red tape.
...but the parent assertion was basically, if you don’t like your job, just get up and walk off and drop your stuff on some other interesting teams table. Job done!
Don’t do that.
Then I also started working with one of the other teams I find interesting, and I now spend (more than) 20% of my time working with that team on various things.
Sitting around and stagnating your career for a nice salary is a profoundly short-sighted thing to do.
Any ideas what can I do with such a situation with my solution? Can I compete with Google?
I'm waiting right now for their model performance score. But I got feeling that they are only tuning Neural Networks. (however I cant find info about algorithms they are using).
Maybe this is crazy, but I feel that I can compete with them on model accuracy and UI. For sure, I cant compete with them on marketing.
Depends on how much you're paying them and the SOW you've signed.
> non transparant pricings and they disrespect privacy.
For pricing, talk with them. Or use a cloud broker.
You’re joking, right? Every other day there’s a top 10 article on HN about Google locking out a whole business customer with no humans to speak with.
Below is an excerpt their T&C, use those to your advantage. I can't imagine any serious enterprise customer wanting to tie them down to Google. Just tell the market that you do exactly the same as this platform without tieing them down, without privacy violation and without letting them down by abandoning the product (yours is open source)
God speed!
12.1 The following terms apply only to current and future Google Cloud Platform Machine Learning Services specifically listed in the "Google Cloud Platform Machine Learning Services Group" category on the Google Cloud Platform Services Summary page:
Customer will not, and will not allow third parties to: (i) use these Services to create, train, or improve (directly or indirectly) a similar or competing product or service or (ii) integrate these Services with any applications for any embedded devices such as cars, TVs, appliances, or speakers without Google's prior written permission. These Services can only be integrated with applications for the following personal computing devices: smartphones, tablets, laptops, and desktops
[1] If the tool is your primary value add, I think you're making a mistake in open sourcing it.
I do not think that you can compete with google on alg accuracy, since most of the underlying ML alg are open source (scikit learn or tensor flow). This is not a secret sauce.
That said, to get better models, you will probably need to find better hyper parameters tuning method, which depends on the number of models that you are willing to run per model tuning session. So if you have a unique model search method, you can save 10X-100X search time, which translate to real saving.
In addition, the solution lacks in model management. In general, most business people would like to understand why a specific model make a specific prediction. Most ops people want to track the training data version, model version, alg version etc.
Moreover, The product itself has "best practice" page :
https://cloud.google.com/automl-tables/docs/data-best-practi...
which include a list of features that are not in the product.
And one of the biggest differentiation should be on-prem vs cloud. Are customers willing to put their data in google (or any other cloud, for that matter) ? Can they legally do that?
I think that this product actually benefit the ecosystem since it helps to create a category of auto ml for tabular data, backed by google marketing budget.
Generally, if a model is not performing or generating unexpected results, its almost always the data or how the question is being structured.
We’re also a startup in the ML space but solely focused on production deployment.
You’re right in that model accuracy is not the most important thing. There are many other considerations as to why a model should be used including training time, processing costs, value.
Furthermore, as someone mentioned earlier, Google’s service is horrendous and provides another angle to address. Their reputation as a company to be trusted with data is also somewhat shaky given all the privacy concerns.
Add on top of that better customer support, like other replies suggested, and your product won't be dying due to competition from Google any time soon.
I am not sure what you mean by methodology?
The one thing that I, as a tool developer, do not understand is how humans would be able to handle hundreds of models, including retraining/monitoring and deployment without automl.
If you build your business on this platform your business must be willing to evolve with it.
If I'm understanding your questions correctly, the main problems I see with this are:
- Using raw data instead of feature engineering (less of a problem given feature synthesis libraries like https://www.featuretools.com/ and other heuristic methods). I'd expect Google to do a good job of basic things like normalization of raw input features before training.
- Using features that it really shouldn't (if you just throw ML at your database for say, loan applications, then sensitive / personally identifying information can/will be used as features)
- Lack of insight / understanding as to what is driving the model. This can be partially overcome with post-training methods like LIME, Shapley values, etc.
I wouldn't expect predictions to be from a set of discrete values - if (say) predicting housing values and training a NN, the output should be continuous and based on the input features.
Can someone explain what this does in engineering terms? How does this differ from something like AWS Sagemaker?
With sagemaker you still need to provide the python code for training.
The big question is: If all your data sits in AWS, because your app that generates the data is there, do you reach across and try to use the Google AI tools, or are their tools compelling enough to get you to move your app and all your data to GCS?
Google is the one to blame for this being tiresome.
Its worthwhile reminding consumers to not jump on the next Google bandwagon blindly. Its Google's job to convince us wrong, looking at this should help get perspective: https://killedbygoogle.com/
This is HN. I'm not sure the readership here are the kind of consumers in need of this perpetual reminder.
No, you don't. Because any rational person would know that nobody is going to have or tell you the answer. So, still glib and disingenuous to boot.
Remediation? That's a cost center that google has automated away
https://www.programmableweb.com/news/google-clouds-threat-to...
I've had good results with it, but you do have to do things their way and its not always well documented. If you want more control you should create your own model and host it on google app engine, otherwise AutoML is what it is, no way to customize or tune it other than changing the training data you give it.
Google shuts: 700 HN entries: https://hn.algolia.com/?query=google%20shuts&sort=byPopulari...
(thought it would be funny to compare, it's not very scientific of course...)
Disc: Googler.
I know Azure ML as has been out for 3+ years - so I assume they have many features and enterprise learnings baked in over the years.
Does anyone have good comparison?
From my experience doing this for a handful of companies, it's almost always better long term to use ML libraries that fit into the organizational architecture that exists, rather than outsource the whole ML pipeline. It's a serious amount of lock in to do that.
Maybe it's a an easier sell if your entire pipeline is already built into GCP and you're just tacking this on as a parallel path.
I did speak with someone in financial services who’s been all GCP and they are quite impressed by the way everything is integrated and how they do not have to shift data from storage to train.
I should be able to use the documents and their tags to build more derivative tag data so that when I'm given a new document I can compute a set of automated tags for the user.
This way when you save a document to your repository it's given some suggested tags.
From Google: "At Google Cloud, we do not access customer data for any reason other than those necessary to fulfill our contractual obligations to you. Technical controls require valid business justifications for any access by support or engineering personnel to your content. Google also performs regular audits of accesses by administrators as a check on the effectiveness of our controls."
Justification reason codes for data access https://cloud.google.com/logging/docs/audit/reading-access-t...
That said, I agree with what you've said here: "Laws enshrining natural rights of privacy of personal data can't come soon enough." Nicely put.