ignirtoq

First, don’t need to respond to arguments made in bad faith. There’s no net positive outcome possible for the other person who is coming to the conversation in good faith.

Second, not having their own solution does not invalidate anything critical they say about the one under discussion. There doesn’t have to be a “better” solution to justify not implementing something that has no positive impact on the targeted problem and severe negative impacts elsewhere.

Determinism means performing the same way every time it is run with the same inputs. It doesn’t mean it follows your mental model of how it should run. The article you cite talks about aggressive compiler optimizing causing unexpected crashes. Unexpected, not unpredictable. The author found the root cause and addressed it. Nothing there was nondeterministic. It was just not what the developer expected, or personally thought was an appropriate implementation, but it performed the same way every time. I think you keyed on the word “randomly” and missed “seemed to,” which completely changes the meaning of the sentence.

LLMs often act truly nondeterministically. You can create a fresh session and feed it exactly the same prompt and it will produce a different output. This unpredictability is poison for producing a quality product that is maintainable with dynamic LLM code generation in the pipeline.

Bluesky is one, single platform. It stores the complete data for any given user post in its databases and provides that through its data stream and APIs. This means every different client someone writes has access to all the same data as every other client, because they’re all going through Bluesky. This also means if Bluesky doesn’t support some feature, no clients can either.

The architecture of the Fediverse is different. Forgetting ActivityPub for a moment, Mastodon is one platform and Pixelfed is another. This means each one has its own data model, internal storage architecture, and streams/APIs. Because they were built for different purposes, they support different features. I don’t use either, but I expect there are image-related features in Pixelfed that are just not possible in a Mastodon client, not because someone hasn’t written a client capable of it, but because Mastodon doesn’t have the internal data storage nor API to support it in any client.

Where ActivityPub comes in is a unified stream language. When a post pops up on a platform, that platform has the complete data and translates as much as it can into an ActivityPub message to send to other platforms. Some platforms haven’t figured out yet how to pack all of their relevant data into an ActivityPub message, so some data may be lost in the sending. And different platforms may not support storing all the data in a given ActivityPub message they receive, especially if it’s from a feature they don’t provide, so some data may be lost in the receiving.

Ultimately this means even with ActivityPub linking things together, the data flow isn’t perfect/complete. So different data is available to any even theoretical Mastodon client compared to a Pixelfed client because the backend platforms are different. Their APIs expose different data in different, often incompatible ways, so even if someone wrote an image-focused client for Mastodon, it wouldn’t be possible to do everything an image-focused client for Pixelfed could do, because the backend platforms focus on different things.

The walled garden (micro services in an isolated network) is the first line of defense. In case a malicious actor finds a way into that network, the second line of defense would be to authenticate the service-service traffic, so the micro services reject direct requests from clients they aren’t expecting.

Nadella is adamant that these kinds of boosts that AI provides will justify AI and carry the industry, stressing less spectacular and more practical applications of the tech.

This is a huge about-face on the earlier proclamation. I really wonder what changed his mind from “AI will radically transform every industry” to “it doesn’t need to be used to discover the ‘magical molecule,’ but provide some other tangible, less extraordinary benefit to developing the product.”

Sure, everyone here has seen the writing on the wall for years, but until now his paycheck has depended on him not seeing it. I wonder if he’s getting internal pressure from some on the board of directors.

Really wish I didn’t have a health condition that made it too risky to see in theaters. I hope there’s a digital version I can buy at some point.

There’s no substantive evidence in this article. They present 2 kinds of evidence: giving the text to LLMs and asking if it’s written by AI, and asking representatives at major food delivery app companies if it’s about them. How are either of those better sources of “truth”?

The article then also cites second hand stories from other journalists. Apparently the original author of the post acted suspiciously when the journalists tried to get more information. That would be great to corroborate solid evidence, but in the absence of good evidence it’s just gossip.

I’m not saying I believe the original post, but I definitely don’t believe the claim in this headline.

Are we there yet? Is the bubble popping? Do we need the Ron Paul gif (or am I showing my age)?

My hypothesis on that is people responding to others’ body language to get the same snap-out-of-dissociation effect. The people closest to Batman would see him and then look around at others more to gauge their responses. Others further away wouldn’t see Batman, but would notice the more-attentive-than-usual other passengers and be similarly more attentive to try to find out what’s going on. They then would notice seemingly unrelated things, like the pregnant woman, and respond more than usual. The paper also says Batman entered from a different door, so a ripple effect of attentiveness could explain this effect without needing responders to directly see Batman.

The problem is that some small but non-zero fraction of these bugs may be exploitable security flaws with the software, and these bug reports are on the open internet. So if they just ignore them all, they risk overlooking a genuine vulnerability that a bad actor can then more easily find and use. Then the FOSS project gets the blame, because the bug report was there, they should have fixed it!

I was really hoping that the AI they would be talking about would be the specialized algorithms known to have real, beneficial effects, like the one that basically solved protein folding. But no, they are talking about LLMs making policy, which sounds genuinely horrifying at this stage.