|
Welcome to the Control Plane newsletter, drafted by a real human being – hi, I’m the editor, Alex Perala – and covering some notable developments this week.
Aschenbrenner Again
I’ve been talking a lot in recent weeks about the predictions made by former OpenAI safety researcher Leopold Aschenbrenner, who basically warned in 2024 that the government would eventually, in some form, seize control of private sector AI firms once their technology reaches, or starts to reach, superintelligence.
Is that starting to happen? The White House is reportedly considering some kind of government vetting requirement for new AI models before they’re released, on the grounds that they could enable serious cybersecurity risks – which of course is what Anthropic is already claiming of its new Mythos model, which has only seen a very limited release that is meant to let a bunch of Anthropic’s partners stress-test their cybersecurity systems before it’s rolled out widely.
And of course there was that public dispute between the Pentagon and Anthropic over how the latter’s AI could be used by the Department of War and who really gets to make those kinds of decisions. That dispute ended up getting Anthropic blacklisted as a supply chain risk, a designation that Anthropic is now fighting in court.
Oddly, a bunch of other tech companies besides Anthropic, including Microsoft, Google DeepMind, and xAI, have already voluntarily agreed to let their frontier models be reviewed by the government prior to release.
None of this can really be described as the government seizing control of private AI systems. But doesn’t it kind of point in that direction?
AI Infrastructure Is Complicated
Control Plane has been in touch with some AI infrastructure people over the last few weeks, and together they painted an interesting picture of what’s going on behind the scenes as AI really starts to get adopted into live enterprise production settings.
When you’re using AI to vibecode some small piece of software that helps you do whatever, you don’t really mind that it maybe takes a few seconds to load or that it’s a bit buggy. It’s operating behind the scenes with respect to whatever your boss or your customer sees of your work, and you’re just using it for your own convenience. But if you’re trying to launch a live product – like a voice agent to help redirect customer calls – that needs to work, like, really well. It needs to *work* work.
As Anoop Dawar, Chief Strategy Officer at Deepgram, told us, “A conversation is a tight feedback loop. Add 300ms of delay and the caller pauses, unsure if they were heard. Add 500ms and they repeat themselves.” At that point the conversation “unravels.” So it’s really important to have AI voice software that can run seamlessly.
But then there’s the underlying and connected infrastructure. What if the customer query needs to point to information hidden in some database, and that database is in Kubernetes, a popular tool for managing applications that was not originally meant to handle live, stateful data retrieval that these kinds of AI systems rely on? That’s the issue that Don Boxley, CEO and co-founder of DH2i, highlighted for us. “If you treat data like stateless compute, you will break things,” he said.
And then there’s the bigger-picture issue of where in the cloud all this processing takes place – so many companies are getting kind of defaulted into big hyperscaler cloud services that have hidden costs and aren’t always reliable. So it’s making sense to a growing number of companies to look for dedicated or custom cloud infrastructure. “Over time, you’re not really designing infrastructure anymore,” explains Leaseweb USA CEO Richard Copeland. “You’re adapting to someone else’s menu.”
It’s a set of problems that the end customer doesn’t really think about. And ideally they won’t ever really think about this stuff because their AI services will just work.
Siri’s Stupidity Costs Apple $250M
Finally: Apple has to pay for Siri being so dumb. In a $250 million settlement of a class action lawsuit.
It really is kind of astonishing how bad Siri is sometimes. I say this as an iPhone user. I have been stuck inside Apple’s walled garden for many years. Locked in here with Siri, an idiot who can’t recognize when I say, “Hey Siri” but will leap into action when I use the word “serious” in conversation with another human. An absolute buffoon who can’t understand that when I say, “Siri, play,” I mean keep playing the podcast I was just listening to a minute ago.
Of course, Apple isn’t paying for Siri simply being stupid, but for marketing improvements that didn’t materialize when selling the iPhone 16. Claimants could get up to $25 to $95, depending on how the details shake out.
How do I actually file a claim? I can tell you who I *won’t* be asking.
|
