Brussels orders Google to share Android's AI sandbox with the other kids

Those pencil pushers at the European Commission are drawing up measures to ensure Google opens up its Android smartphone platform to something few users asked for – competing AI services. Landing on Google's desk are proposed measures the Chocolate Factory may have to implement, aimed at ensuring third parties get effective …

[Sproggit]

Portabililty Between Models

This article opens the door to a dimension of AI adoption that I haven't really seen discussed yet: portability.

Back in the day, organisations like ANSI (the American National Standards Institute) developed portability standards for technology that covered programming languages, including COBOL, FORTRAN and even SQL. [ And yes, I appreciate that these are older languages, but COBOL and SQL, in particular, remain two of the most widely adopted languages in computing today.

The thing is, we have no way of knowing which of the current AI companies [if, indeed, any of them] will establish the most capable or effective solution. We don't know if the companies will continue in business, or be bought out, or sued out of existence if something goes wrong.

What we do know is that companies that use technology in general - everything from banks to healthcare to engineering - is rushing to experiment with and adopt AI-powered solutions, even if this largely seems driven by FOMO (the Fear Of Missing Out).

But the problems will start to manifest in maybe a year or two, when the AI race starts to solidify around leading players and we start to see major rounds of consolidation, with smaller and less effective players being consumed or simply fading to insignificance. But what of companies that invest in those doomed providers?

The thing about a programming language like COBOL is that it is relatively simply to port it between systems [as long as it sticks to an external standard, like ANSI]. We have no such collaboration happening in the AI space right now. What happens if you're a downstream consumer of i.e. Anthropic as a major part of your business process and then Anthropic is bought or our withers and dies.

If you've experimented with AI at all in a business setting you quickly learn that it is a brutal, Garbage-In/Garbage-Out paradigm and that crafting really good prompts is essential to getting good, efficient results. Once you have effectively hard-coded that in to other software by means of an API, things get more complex still.

I dare say it will come... but it would be interesting to ask companies who are rushing to adopt AI and just ask them what their contingency plans are if the AI provider they are partnering with is toast within the next 12-18 months.

I have a hunch that not many of them will have an answer for that.

[doublelayer]

Re: Portabililty Between Models

It would be very difficult to get anything like that because there's no defined set of functionality from model to model. That's the whole point; you can ask it to do literally anything and it will parse your request, run something, and give you a result that might have done what you wanted or might have destroyed anything it had access to. You can't standardize the inputs to that when the inputs are theoretically unlimited, and you can't restrict the outputs when nobody's found ways to keep them from going wrong on simple prompts every time, just most of the time.

[Sproggit]

Re: Portabililty Between Models

I don't disagree with your observation that "there's no defined set of functionality from model to model".

Perhaps AI is generic enough that at some point process can be switched between models from competing vendors and still return acceptable results.

We're not there yet. I'm seeing differences between two iterations of the same model that have force a complete rethink in terms of basic process design.

I made my original comment because of the way I've seen models "drift" in terms of their performance, even over a couple of dozen iterations of the same basic process.... To be sure, there are going to be a good number of business cases where an iterative, learning approach that renders different results from the same data inputs [over time] is not an issue... but I suspect that for every business process where such variability is OK, there are going to be 2 processes where the business owner is going to want a better degree of consistency.

I think portability is tied to that consistency challenge - because of course you're going to want to be consistent in your prompts and to get consistent results. There's no point in developing a model to produce insurance premiums given a set of vehicle, drive and location data that starts to deviate too much, because it wouldn't be reliable enough to be profitable for the business trying to run it...

And I'm not sure I'd fly in an aircraft where the autopilot was engaged for the landings and simply decided to iterate or experiment with how to get the aircraft on the ground...

Then there's that piece over on Axios today where OpenAI and Anthropic are both saying that their latest models are rewriting themselves... and I'm not so sure this is heading in the right direction...

[Anonymous Coward]

But the UK will have to suck up whatever the US wants ?

Not being in the EU and all that.

[Snowy]

Apple

Will Apple have to give the same access?

[doublelayer]

Re: Apple

I'm not sure Apple has that level of access for anything, so probably not. Under the law, if they also had a control interface for their AI model, they'd probably be asked to open it as well.

[DS999]

Re: Apple

If Apple gets to the point Google is at pushing AI as hard as they are while only allowing their in house model, probably. But so far Apple is actually more "open" AI wise since they already support multiple models (at least OpenAI and Gemini, and I believe Anthropic as well)

[druck]

FOAD option

Ideal opportunity to insist that there is an option of NONE for any AI integration.

[BasicReality]

Just shut down all Google and Apple services to the EU. Leave them in the dark ages again.

[Dinanziame]

Yeah sure it's nice to be principled, but you know, Google and Apple make a shitton of money in the EU. That's why the EU gets to make rules, and if they want to keep making so much money they have to follow the rules.

And that's just "basic reality"