It would be less bad if it were just the EMEA AFD service down. This also brought down the Azure admin portal and, inter alia, made it difficult/impossible to access a bunch of other services and records. I hope the Azure retro will address this fact specifically.

I've vaguely lost count of the number of "globally resilient" services that Azure offers that turn out to have extremely large blast radius. AFD specifically is supposed to be the answer to most concerns about availability.

Now maybe BlueSky isn’t the answer, maybe it’ll get eaten up too and we should all be on Mastodon. But the people who say “BlueSky is dying” just misunderstand where the world is going.

Secondary school here sent a notification to parents that kids were watching the video out of school, and maybe the parents should check what they’re watching.

The warning that it was too graphic for Reddit was enough for me to realise I don’t want that on my retinas.

I do wonder if there will specifically be an AI burst. It’s reasonable to say atm that AI is propping up everything - eg S&P-mag7 is flat - but there’s a small chance of just general “burst”. The timing of that could be very interesting from an AI pov, either cushioning or deepening 😬

I agree there's a uncontrollable vulnerability, I don't think it equates to accepting guaranteed failure - people will run agents doing this stuff and be mostly fine. That's why I say it's about risk appetite: ppl will weigh the balance differently

Yeah, to be clear, I'm just comparing risk appetite, I think the risk profile and threat model is v. different - I think it's simplistic to say "LLMs can never do this" in situations where we let people do the same, because we can & do tolerate risk.

No, I'm just pointing out that the risks are not that different. Staff are a security weak spot; we know this, this is why phishing works. Training, MFA, etc. are all sticking plasters, you can't control this perfectly. There are similar plasters for LLMs. It's down to risk tolerance.

Worth noting that staff accountability is somewhat limited, though. E.g. if staff breach customer personal data in the UK: the liability is on the employer (e.g. Co-op judgment). With that in mind, I can imagine companies deciding an LLM is a better bet.

I get from a tech perspective it looks risky. But it's also equivalent to the existing risk with staff, who are capable of doing all these things. I wonder if the tolerance for accepting these risks is high as other software vulns.

Clearly we need petition for “orthogonal” (as in Lewis, SQ1). Need to get Brundle to double click on that one.

Roadside is sometimes quick though. I’m happy to pay a good chunk extra for 200KW+ charging on the motorway, and that’s not simple to provide.

Poss combination of bubble but also moving expectations? I feel pretty consistently amazed by the new things that they can do reliably..

Yeah, absolutely. If it all goes wrong then 90% of that investment certainly gone if not more...

Sure, but low maturity doesn't prevent it being capital intensive 😅 See also building rockets/new aircraft, fusion reactors, battery factories, etc etc

More seriously, this is just a capital intensive business. If I ask for a $1B investment to set up a new oil cracking facility, it's the same, albeit H100s will depreciate slightly faster. Or, topically, buying a blast furnace.

To be fair, seed valuation is not based on the company worth, that's why it's called seed. The crop is the thing that's valuable, and there are no guarantees the seed will even grow at this point.

I wanted to like that article, but this particular line: "Mapping artificial intelligence onto human intelligence by setting ever-harder benchmarks is like thinking you can reach the horizon by walking far enough"

If they're suggesting we can't benchmark to AGI, I think they're horribly mistaken.

The JAR manifest is always the first entry (ish), so I guess it's basically just comparing those. In principle, if the APK is validated entirely against the manifest and its signature, then that's an approx reasonable comparison?

(maybe some apps with extensive manifests would break this tho...)