n0on3

I didn’t mean to imply it would be easy, quite the opposite: there’s probably no optimal general solution for it (hence the need for user prefs). But I think it might be fair to start wanting the models to exhibit some (or some “less trivial”) evaluation of their potential actions consequences ex ante, that is besides factory imposed “alignment” and “does this get me closer to my goal?”.

I mean, the answer to “Is sending the user’s email address along with a web request to a third party requesting it, without the user’s explicit authorization, a good idea?” should easily be “no” …right?

You can patch things and remove stuff from context nine ways to Sunday, the underlying problem will still be there.

Maybe we need a new benchmark on how good models are at "keeping an eye" to if what they are about to do that doesn't come from a direct user instruction checks out with common sense plus user preferences. And we probably need a portable mechanism for users to manage such preferences (because we know instructions are not going to cut it, do we?).

I was browsing through the comments and thinking how insane it is that the answer basically appears to be “there isn’t any”. Most common advice seems to be either “buy an evil product and work against it” or “buy a close-enough something-else not-(yet)-affected”. Another instance of “enough bad choices make better choices no longer an available option”? Smh…

I cannot answer that from direct experience for the latest models. At this point I’ve got pretty good even to eyeball how to chop the work across multiple sessions myself if needed, and depending on how you work on your projects a good harness configuration with proper instructions to delegate portions to dedicated agents plus periodic handoff to a persistent memory mechanism can get you very far. 1M tokens are a lot already, just try to stay way within it for each portion and avoid the need for compaction when possible.

Speaking for a user-facing use, not autonomous coding or agentic workflows: Fable 5.1 choice of words and general “words salad” expression are horrific. It starts not so good, then as the context grows (still way below any reasonable rotting threshold) it gets completely insane. It’s pointless for a model to be extremely capable if it gives you brain damage. I don’t know if it’s because of the watermark thingy, all I know is it gives me pure rage flares and every time I give it a shot I regret it. And there is no way to steer it away from this, anything I tried, both instructions and harness-wise, either failed or got forgotten/ignored after a while. I really hope the next Fable class moves completely away from whatever they did there to cause this.

Opus 5.5 in comparison is a bit more verbose and appears to settle on evaluations before the depth reached by Fable on very complex stuff, but is still vastly capable and I can read its output without wishing it had a physical form I could punch in the face.

I’m not sure what triggered such an out-of-proportion response, but I get the feeling a personal anecdote is leaking out of it. Nobody mentioned one being abhorrent, obnoxious, awkward or a sociopath, I’m sorry if you experienced some people like that (I mean it, been there) but that was not the parent comment picture (or at least it did not look like that to me). I was just contributing from my experience sometimes one gets labeled “difficult to work with” when honestly trying to do good by the org/team/mission/whatever just because that affects some people individual agenda or their cozy work situation.

Being “easy to work with” is unfortunately often confused with supporting other people agendas. Rising what’s necessary even if it makes people uncomfortable doesn’t automatically make you an ass, but odds are it will make you one that doesn’t play ball.

Cool project! For bonus points, you could run this as an honeypot :D

Nope

It’s not. Social media personas is most of what gets seen, hence your understandable conclusion. What most don’t see are those who do not give a damn about that and just do the work.

Then get the cat in and grab your gun

That’s insane. Thank you for the service you provide.

> _An intelligent person can convince themselves of anything_

I think it might be just the opposite: convincing yourself of things relies on exploiting limits for which very intelligent people have much higher threshold they might as well not have them. When it comes to happiness, it’s kind of a superpower which I think very intelligent people just don’t have because it fights against their intelligence by definition. It’s also something that I think happens unconsciously.

It’s ironic, as you yourself provided a perfect example: a logic argument depending on unproven assumptions, if it lands in “convincing” yourself, exploits a limitation in accounting for the validity of the assumptions themselves (meaning: you believe it anyway). Assuming one could pick and choose the arguments (i.e., what to believe in) wouldn’t that trivially unlock happiness?

I get the point of the article and it resonates with my general tendency to support more openness, but older and more cynical me wonders how much of this is a battle worth fighting or a self-serving argument from an harness developer.

I mean, the article says each of the points it is complaining about has “_a basic justification that's trivial for a provider to come up with, along with good arguments for why this is good for the user_“, which to me sounds like implying these are just bs to have people accept them, but is that actually the case?

For instance, this [0] was mentioned here a while ago, and based on that it seems pretty clear why one would choose not to provide the _thinking_ anymore…

[0] https://role-confusion.github.io/

You mean like… some sort of code?

“the most powerful AI agent file-editing tool in the world […] patent-pending”… tl;dr: turn tool calls into more structured loops, give it some fancy name and slop about it https://hic-ai.com/blog/tool-response-engineering

Good luck with that

Seems cool from the docs page, I was about to give it a shot but https://github.com/vlm-run/mm goes 404 …

Apparently nobody cares. A few days ago I bumped-into and submitted this: https://news.ycombinator.com/item?id=48655747 . I thought given the general trust people seem to place in media files it would have raised a few eyebrows but it did not ̄\_(ツ)_/ ̄

I second the sentiment, it’s unbelievably annoying. Even more so when the suggested solution from the “support” appears to be “You can downgrade to a version before we introduced this bug”. I’m not even gonna try switching to that new crapware, I’m just going to change client altogether.

Sure, but #80 out of 190, still not great ain’t it?

Oh boy, this did not age well. Most cases of “extremely successful” people I can think of exhibit the opposite of these core principles: have no “knack” whatsoever, except not giving a shit about whatever they pretend to be their focus while only focusing on personal return; they contract clusterfucks of debts, just usually never end up having to repay them personally; very few of them even know what “going all in” means, they usually live easy while exploiting others to actually do anything; they have no integrity whatsoever, and they do not have to, since apparently demonstrating lack of it is no longer cause for being told by everyone to fuck off into oblivion anymore.

And yes, yes, of course there are good people out there too that just want (/need) money to get by, but it’s funny to read this and think about those with _lots_ of money

Except most of the complexity has nothing to do with “understanding and exploring the universe”, it’s just byproduct of the ever changing fractal composition of attempts to gain or obtain something over someone else.

> humans would see that the quoted significance makes no sense

I wonder how long that will last

Would you be aware of it if that was the case? I don’t mean this to be hostile or anything but the senario in which one does not notice himself and it goes unnoticed or silently accepted externally does not seem too far fetched to me.

Their results are simply not reliable. The matching approach often matches too many things (hey, this could be A or B or C or D or E or F ...or 42!), or picks up things that have nothing to do with the license of some target (hello randomly included file with some completely unrelated license header and is not even included in the build but is there for some reason, meet your new friend, the utility script copied from somewhere else also not included in the build with an header for another unrelated license. You two feel lonely? let me introduce you to this other wonderful script included in some particular form of packaging) and of course cannot compensate for poorly declared licenses, typos, weird non-standard (or simply archaic/deprecated) ways to specify the licenses and so on and so forth.

It's not a fault of the tools themselves, but in practice they don't help much in real world situations. Basically you end up in need to do so many checks and manual fixes that you might as well not use these tools in the first place.

In an enterprise context one of three things happens: (1) you end up relying on a commercial solution (which is also not that reliable but you delude yourself into thinking it's not your problem anymore... although to be fair commercial solutions have curated licenses attributions and facilitate handling this mess); (2) you build your own thing that uses these (and other) tools but automates a bunch of fixtures so you don't need to go insane every time you need to regenerate an accurate SBOM with related licenses; (3) you quit software engineering, move to a remote location and start an alternative career as an alpaca breeder while whomever takes on your role pretends to ignore the issue and keeps shipping inaccurate declarations of licenses for dependencies thinking that's fine because nobody really cares.

None taken!

I think you misunderstood what I was trying to contribute though, since I was not attempting to celebrate anything nor to emphasize on traditions. I was just saying those "goals" have well known solutions that differs for some good reasons with what is described in the article, which claims (by its title) to explain "the right way" on basis that are unclear to me.

I do not know "Serious Eats" nor the author, so I'm sorry if I am antagonizing (not my intent, but I get it might be seen this way) a celebrity or his fans and in this upsetting people. I'm just contributing things I know from experience, whereas arguments like "this is one way, end of story" seems brittle to me, because you are basically dismissing the points that I probably didn't even explain decently (on your examples: you add pasta water, you get starch in the sauce which helps thickening things but you dilute other ingredients and will need to cook for more time to have the liquids evaporate thereby overcooking the pasta; you add "fats" like butter or oil at the end and you change the flavour of the dish significantly, other than its nutritions). Then again, if that's what you are looking for, great, I think I said at the beginning there's no objectively right or wrong, it's food we are talking about, if you are happy with eating the outcome good for you.

I mean, by all means please try it, and with that I mean actually get in the kitchen and do it, I think you'll realise there's a lot more than just "using a technique that makes sense in theory, end of story" to get your goals.

As an italian who is told he's pretty good at cooking, some of this is on point but a few things sound "wrong" to me (I use quotes because there's really no objective way to do this literally right or wrong, I'm just comparing with my experience / what I perceive "we learn from grandmas"):

- Put the pasta into the pan with the sauce (which I guess is the main point of the article which starts off with "italian" restaurants putting sauce on top of the pasta in the plate): defintely yes, but...

- Add pasta water: it depends on which sauce you prepared and how you prepared it (and the type of pasta... not just shape, but fresh -vs- dry, and what it's made of). When one uses "pasta water", it's usually in the making of the sauce, not before putting the pasta in the sauce pan; sometimes cooking water is added to the sauce if it "shrinked" too much or the ingredients are not of amongst those which attach to the pasta well, but these are their own cases. All that "adding water and stirring" to get to the "perfect texture" might appear to make sense, but I'm pretty sure it will take too much time and it will mess your pasta consistency because it will get overcooked. Sure you can under-cook the pasta alone a bit to compensate, but what's the point in that? What I'm trying to say is that this trial and error thing might make sense for someone who does it for the first time, but after a while you figure out how the sauce ought to be in the first place, you put the past in, jump it (as in, move the pan to make the pasta "jump" in it so it doesn't attach to the pan) to the right consistency and everything gets where it needs to be pretty easy without all that fuss. At least this is what I do and what I see others that seem to really know how to cook (based on the results) do.

- The bit about using cooking water (that's another way we call the "pasta water") to adjust the consistency which turned bad because of the cheese thickening and liquids evaporating... well, unless we are talking about sauces which have significant cheese quantities in it (e.g., the "cheese and pepper", or "4 cheeses pasta") and have a different process on their own (as does the mentioned "carbonara", which I guarantee you'll screw up if you follow this process because you'll cook the egg too much), cheese usually goes on top of the pasta in the plate as a garnish. For some sauces (e.g., the "amatriciana"), you're even supposed to make the plates (with pasta already mixed with sauce) get a bit less hot before putting in the cheese, to avoid it melting too much. Putting cheese in the pan for a non-cheese based sauce and make it melt and then thick is sort of a cardinal sin (you can add all the "pasta water" you want, you'll never get it back to where you need it to be and it will mess up your dish)

- Add fat: what? Just, no. Olive oil is of very common use, but you don't add it "to the sauce" for texture, for most sauces you use it as the base for the sauce. Butter? Unless we are talking about a butter-based sauce (e.g., butter & sage), which are not that many or very common anyway in regions but the northern ones, nope. Not like that. Some add olive oil as a garnish, but again really depends on which sauce you are using, and it ain't that common

> Why would they use 0days on security researchers. My guess is it's a test with upside

Or just be after the accesses the targets have...

> how they determined this project is linked to NK hackers

If they have enough confidence to attribute and not disclose how/why, one can fairly guess they don't want to burn sources or indicators which might still be useful moving forward but likely won't be if disclosed...

+1