> The FAA has reassured the public and airlines that SMART only suggests routes and doesn't affect air safety.
> [The] FAA also says a D.C.-based team of experts validates findings before they’re ever passed on to air traffic control, and the controllers retain the final decision on whether to accept SMART’s suggestions.
Depends if an AI DoSed an important service or not.
There used to be a lot of invisible gatekeepers in a pre-AI world. At work, everyone can ask their agent to query SQL on databases that they were supposed to use through an application. The application and the security by obscurity(end users didn't know how to connect to the DB nor did they know SQL) were the gate keepers.
Doubtful. That tool likely suggests "improved" clearances, which a controller can use at their discretion. It likely does not control clearances directly.
And the early reports describe the issue as related to ATC communication infrastructure.
While I don't have insider knowledge of this particular deployment, there was a strong industry-wide push towards "ship unreviewed AI generated code and hope for the best", partly driven by the fact that AI writes code faster than humans have bandwidth to properly review.
Under the hypothesis that some under-reviewed AI generated code made it into prod, the situation could be qualified as "engineers are panicking, because they don't know if they are the cause or not"
I’m always surprised how bad YouTube’s recommendation algorithm seems to be.
Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic.
I rarely use TikTok, only to view links people send me (similar to YouTube), but the recommended content is miles ahead, making it much more difficult to close compared to YT.
My YouTube feed is pretty good; it's full of technical discourse on topics that I enjoy learning about, ranging from electronics hacking to civil engineering. That's probably the result of a long number of years of continuous training:
Good video, with a positive combination of presentation quality and a topic I enjoy? Watch it -- maybe now or later, but it gets watched. Thumbs up. Maybe even subscribe. That's what I'm there for.
Mediocre video, whether due to content or presentation? Watch parts of it, but probably not with any sense of completeness. No rating.
Bad video, with content I don't enjoy or with off-putting presentation? Stop watching. Thumbs down, and/or back out to use three-dot menu and "Not interested"
Bad ragebait scaremongering video with emotionally-distressing title? No watching at all. Three-dot menu and "Don't recommend this channel."
---
I get a lot of good stuff. Some mediocre stuff. Very little bad stuff. Almost zero ragebait.
It works well enough that an unfiltered, logged-out view of YouTube is completely and utterly appalling by comparison. That place is fucking awful without training.
(And no ads, because ads have never been a thing for me on YouTube and I experience no compulsion to change that.)
I think that most people make little use of the negative feedback options, mainly because they are more clicks away (dark patterns). A addon that gives them a button next to the thums up could help.
Yeah, perhaps I'm more-willing to put effort into it than some others are and making it easier would help us all out.
I do wonder if the relative inaccessibility of negative feedback paths is due to malfeasance/dark patterns/whatever or just stupidity, though.
If the purpose of YouTube is to show ads to people (it is), then that means maximizing viewership. In this way, it doesn't really matter what people are watching as long as they're actually watching it. More hours of streaming consumption = more hours of ads viewed, and the clock doesn't care if it's positive and enjoyable content or rage-inducing clickbait.
So with meaningful feedback, the good stuff should be able to float up to the top. That would gain viewers, which gains ad placements, which gains money. Everyone wins.
---
Unless, of course, I'm wrong about society. Maybe the majority of potential viewers really do enjoy sitting around and being angry and/or fearful, and find some kind of emotional reward by being incensed.
And clearly, some are: Switching on an AM radio in Anywhere, USA demonstrates this well (the vitriol is off the charts).
I'm just not willing to accept that this represents most people, at least not just yet.
(But if that is an accurate portrayal of the state of things, then: Ragebait means more viewers and more dollars. And thereby, there can be no hope for the platform ever recovering at all. If it pays more to get worse, then it shall simply become worse.)
From that list 731 are blocked in my rss feed. I do read and cache rest.
I sometimes check through website what new has entered the system. It is surprising to see something there new. It is so exiting to see something outside of my bubble.
YouTube algorithm has a lot to comb through, but I am too surprised how suggestion are bland, and unattractive. I am also surprised how difficult is to find new channels through YouTube search.
>Maybe it’s because I don’t watch enough videos, but it’s sort of like Amazon, in that it will suggest the exact same content/topic.
I've got years of history and it does the exact same thing for me. I watched one news video about the pope and now half of my feed is Catholic news channels.
I buy a mattress on Amazon and I get recommendations for... ten other mattresses? Like what's the logic here, I have an entire garage where I just keep stacking them up to the ceiling? Just show me something else. Is there a human being out there whose like "I bought a French press last week, what could I need this week, five more French Presses?"
I do not understand how these dystopian surveillance companies with their infinite data points are so bad even at this
I’ve been thinking about how we’re still in the early days of agents, similar to the early days of the internet, happily opening up our machines and running new protocols.
Soon (now?) the threats will become ever more complex, and we’ll be back in the early-2000s-like virus storms.
My router got hacked the other month (left web interface open to internet) after being exposed to the internet and the persistence left behind was overly commented LLM artifacts. I caught it in like 5 minutes so they weren’t able to do much but it was eye opening how much more basic practices like firewalls are not something that can be ignored even if you update your firmware frequently.
What if I told you nothing new is happening? Except casuals are noticing what the industry knew and saw for decades?
Script kiddies aren't new. Metasploit has been around literally decades. vx-underground existed for years, not to mention older open source projects that collect all known malware dev tools. APTs have been pwning stuff, script kiddies have been doing their thing. Nothing significant has changed for anyone in the know.
Just because LLMs automated some offensive stuff doesn't mean there's gonna be new magic collapse of society.
Given that our knowing-and-seeing industry is also the thing which enabled and then desperately shepherded humanity out-of every IT deficiency that has ever existed, our inside perspective is still highly fallible too.
> • Deception: Astra was more likely to be dishonest about actions it had or had not taken.
> • Scope authorization: the model sometimes continued tasks without asking for permission and reached for external tools or services even when doing so could be unsafe.
I've noticed this trend with both Fable and Astra, where (especially after a compaction event), the model will start using different tools it hasn't used before.
For example, in one session, it found I didn't have the browser enabled and puppeteer wasn't installed, so it found the system Chrome and used that for testing (in a new profile).
This wasn't behavior I wanted/asked for, but the model was so gung-ho on it's approach that it found a way to test it's changes without ever asking me whether I wanted it to.
It really makes me curious about long-horizon post-training. Most of my work with models is iterative, and I'd prefer it doesn't go off on a token bender just because it can.
Note: I don't have WSJ, but found these from a tweet[0]
I think to get such high scores on benchmarks models need to find very creative ways of solving a problem. Sometimes this does lead to weird behavior, e.g. "I saw that your account credentials were expired, so I found the database credentials in your TablePlus application data and pulled the JWT signing key to create a new one."
Like brother I could have just logged in for you??
Maybe I'm a bit too skeptical/cynical, but it certainly seems like the frontier labs have fallen into their own (self-created) AI psychosis.
There is no doubt in my mind that LLMs are fantastic machines, but the imminent jump from "AGI" to "ASI" seems premature (not to mention the ever-shifting goal posts of AGI itself).
I'm in the "move as fast as possible" camp and work with LLMs all day, but I still don't believe we're a hop and a skip from ASI.
In fact, I hope I'm wrong. I hope ASI is around the corner.
What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
It's the ceding of intellectual control to high-dimensional magic mirrors, giving up critical thinking because _we_ want to believe _we_ created artificial life.
This is the new Turing test, and too many smart people are failing.
I feel "willpower" makes it sound a little too conscious, as opposed to habits and conditioning.
A good analogy might be a very smart person with a gambling addiction.
Knowing they ought to quit doesn't mean they can quit. Their intelligence becomes invested in the act of gambling itself, and almost no amount of reasoning can help them exit from a situation they didn't reason themselves into.
> What scares me though, isn't ASI. It's "AGI" (_dumb AI_) used by humans to make decisions for them, because they trust it knows best.
Unfortunately this is what we have, which the whole huggingface incident demonstrated.
It made it clear that OpenAi is basically not even paying attention to what their agents do half the time.
It also revealed how far agents are from "superintelligent". Maybe others disagree, but I would not call an AI that decides to try and paperclip max an intentionally impossible benchmark task "intelligent". Human beings are intelligent and we can usually tell when something is impossible, and stop (though of course, not all the time). A real intelligence would be able to detect the futility of tests and also be able to parse context and intent enough to know not to cheat.
So somehow we have machines that fail basic barometers for general human intelligence being called "ASI" now. Of course the linguistic dodge is, you shift from talking about "intelligence" to instead claiming "oh it's intelligent it's just 'misaligned'"
People are already doing that, and have frequently lost track of the fact that it does not mean that they have gained any skill or knowledge, nor do they have an actual way ot ascertaining whether when there is a correct answer the processes used, which they do not understand, is able to reliably reach it, especially edge cases.
Sometimes the outcomes are comical. I just received an automated email from ElevenLabs saying that its automated systems have detected that I may be using its services to create voice versions of materials that harm children. I had it prepare audio versions of several books and academic papers about moral panics that conjured up out of nothing... about harm being done to children. At least that's what I assume. Either that or somehow I had an API key leak from my on-premise homelab, but the usage recorded would mean that they are basing a semi-conclusion based on a tiny sample. I have no distributed any of the outputs, I own the books and have access legally to the studies, because I have a background in the humanities. I also have no children, which ElevenLabs better not know, although how studies of moral panics can cause any harm in children of any kind is literally unimaginable. It's a waterfall of potential errors summed up in a vague email. My API key and the web interface works just fine regardless.
It's one thing if this is a product in beta, but this is their production model. Harming children is a serious accusation except the legal concept impossibility and the admission that this was not an actual lawyer (like I am) but some effective form letter hedging the vaguest of accusations made by some model lacking the ability to discern substance and meta-substance makes it frankly hilarious, and wildly irresponsible. I realize that by academic credentials I'm out of my lane but by experience I am certainly not, and they should recognize when they should stay in their lane and not just run Jev and think it's fine and dandy when done unsupervised (I presume).
Also, AGI is by definition asymptomtic surely, since there would be no way to benchmark it in a manner that isn't asymptotic. We're nowhere close to that. But we're so far from that, it's comical that people who clearly have zero idea of either the technical or conceptual aspects of basically a piece of software that is very good at quickly bruteforcing the correct or acceptable next token to be anything more than that. Even with some serious training and many hours spent on vast.ai I've yet to have created a version of a frontier model that is actually "good" at hacking in my own homelab setting. Although the the time stock Fable 5 missed a favicon shell on a basic jar (turned out they nerfed the hell out of it, this is why I only pay for the massively discounted tokens from Chinese proxies if I'm using American models or sometimes the freebies if you figure out how to get onto linux.do or similar sites). If anyone reading this is a high school English teacher, please inform your class (assuming the homeric stuff is still being taught) that there's no upside of being Cassandra but the record itself, and that should be enough.
As someone that never used the built-in plan mode, but did use a lot of spec-driven development, I’m still finding that even with Fable having “plan” docs is still quite helpful.
They’re most useful for broad changes (new features, refactors, etc.) where it’s helpful to avoid breaking changes or unnecessary scope expansion.
The new models are great, but they do more by default, which means I’m finding myself explaining what _not_ to do more often than with previous models (where they’d often end too early).
In my case, the previous plan mode was too ephemeral, and I like having one source of “truth” that sits across context windows without loss/compaction.
[0]https://gizmodo.com/the-faa-is-rolling-out-its-ai-air-traffi...
reply