Gadget Guide: IFA Berlin
Sep 15, 2026 · 42 min · 12 segments
Rob Chipperfield and Lawrence Micallef check out news and announcements from IFA Berlin.
- Speakers
Lawrence MicallefHost
Rob ChipperfieldHost
Sep 15, 2026 · 42 min · 12 segments
Rob Chipperfield and Lawrence Micallef check out news and announcements from IFA Berlin.
Lawrence MicallefHostRob ChipperfieldHostAnd This is an interesting turn of events, isn't it, because the US are pushing for possible legislation after anthropics researchers raised safety concerns.
Yes, we've heard about AIs going rogue already and doing things like targeting other AI systems and setting up their own chat boards and all sorts of interesting things.
It's almost a little bit of closing the barn door after the horse has departed, but let's think about what they might actually achieve with some better regulations.
It does feel like a lot of the incidents that we've seen have been effectively inadequate not-AI controls.
So, in other words, everyone sort of acknowledges that an AI is a non-deterministic thing.
So, if you need it to not do a thing, you need to physically prevent it from being able to do that thing.
And I think a lot of the stuff where we're seeing sort of longer running what's called agentic tasks, you're sort of going, go off and solve this problem and come back when you're done type thing, rather than just a conversation that you might have with it.
We're giving it a lot of power and we're not building those sort of really solid guardrails around it.
A lot of the controls at the moment seem to be, oh, well, I'll ask it nicely not to be bad.
Book me a gym class that's oversubscribed, but please don't go and hack the system that actually runs that.
Yeah, then it won't hack it, but it'll do something else and bribe the manager or something.
I'm sort of... ingest here somewhat, but it does feel like we need to get better as an industry at actually building those solid constraints.
If you want to build something that's general purpose, it can do lots of different things that you didn't necessarily think of first.
But it's going to be so important because, you know, we can scale these AI systems out much faster than we can scale humans.
And This is an interesting turn of events, isn't it, because the US are pushing for possible legislation after anthropics researchers raised safety concerns.
Yes, we've heard about AIs going rogue already and doing things like targeting other AI systems and setting up their own chat boards and all sorts of interesting things.
It's almost a little bit of closing the barn door after the horse has departed, but let's think about what they might actually achieve with some better regulations.
It does feel like a lot of the incidents that we've seen have been effectively inadequate not-AI controls.
So, in other words, everyone sort of acknowledges that an AI is a non-deterministic thing.
So, if you need it to not do a thing, you need to physically prevent it from being able to do that thing.
And I think a lot of the stuff where we're seeing sort of longer running what's called agentic tasks, you're sort of going, go off and solve this problem and come back when you're done type thing, rather than just a conversation that you might have with it.
We're giving it a lot of power and we're not building those sort of really solid guardrails around it.
A lot of the controls at the moment seem to be, oh, well, I'll ask it nicely not to be bad.
Book me a gym class that's oversubscribed, but please don't go and hack the system that actually runs that.
Yeah, then it won't hack it, but it'll do something else and bribe the manager or something.
I'm sort of... ingest here somewhat, but it does feel like we need to get better as an industry at actually building those solid constraints.
If you want to build something that's general purpose, it can do lots of different things that you didn't necessarily think of first.
But it's going to be so important because, you know, we can scale these AI systems out much faster than we can scale humans.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
