fxwin

that's fair, i haven't thought about it that way.

> My economically valuable skillset – programming – suddenly feels less valuable. What shall I teach my son?

Maybe that's just me, but this strikes me as a bit odd. When i think about my parents and what i appreciate about them and how they raised me, hard skills are not very high up the list. Things like encouraging curiosity, agency and freedom to pursue things that I showed interest in even if it meant spending many weekends and many eurodollars on me that they would no doubt have found ways to spend otherwise come to mind first, along with general values about society and how (not) to treat people.

Edit: there are "hard skills" they taught me, but they are more along the lines of "we want you to be a self sufficient human being capable of dealing with life" (cooking, cleaning, how to budget and think about money,...) and less about increasing my value as a participant in the labor market

> When you know the writer personally, it becomes harder to see the text from the author’s perspective if you know it’s not her style.

I think this is why I absolutely hate reading AI-generated PR descriptions, Readmes or wiki entries at work. I work with these people every day, we talk to each other in meetings, but then I read perfectly formatted Slop in the freshly opened PR that has your name and picture attached to it, yet it's almost impossible for me to care about it because (subconsciously or not) i can't bring myself to see it as yours. Same or even worse for reviews.

absolutely unhinged, i love it

I have no idea what I'm looking at here

whether people want to use the service or not is irrelevant, the only thing that matters is whether or not (and how much) they actually do use it. Also there are subscription/flat rate based cinema services that have been in business for almost two decades now, so it doesn't seem like that wouldn't work.

> I don't actually believe that you can stack the dishwasher wrong as such

There are definitely more and less correct ways, but personally i think it's "wrong" to load it in a way that leaves the arms blocked or (not) tilt pieces that get excessive water pooling on top

Because it's kind of true, but also mostly wrong. People conflate two things: "Subsidized" subscriptions as a monetization model, and (unsustainable) subsidization through outside investors.

When you max out your Claude Max plan every billing period, are you consuming more tokens than you could realistically pay for with the same money spent on API inference? Definitely! This works because many Claude Max subscribers don't use it to the max (and you probably aren't using it to the max every single time either), so overall subscriptions might still be profitable. Technically, the subscribers subsidize each other, but this is entirely sustainable (see any other subscription model that is based on its consumers not maximizing their usage, e.g. gyms).

What people mean (or want to say) when they say inference is subsidized is that it is subsidized in an unsustainable way (e.g. with investments that will not last forever - if this is true at scale it should worry us!). What happens now is that they take the fact that an individual subscription used to the max by itself is a financial drain for the company offering it, attach the "subsidized" term to it, and infer from that a bunch of stuff that is usually associated with the term (such as unsustainability) but doesn't really apply here.

Why would LLMs need injected randomness to work effectively? Their nondeterminism isn't by design, it's largely a byproduct of how (batched) inference actually happens on hardware [0]. In case you are referring to the actual token sampling (be it temperature, Top-K or whatever), i wouldn't really see this as part of the LLM itself (And I've frequently seen benchmark sites run them at t=0 anyways)

[0] https://thinkingmachines.ai/blog/defeating-nondeterminism-in...

I agree with basically everything said here, BUT i wish people would stop (mis)using the word "deterministic" in this way:

> Compilers are, for the most part, deterministic in a way that current AI tools are not. Given a high-level programming language construct such as a for loop or if statement, you can, with reasonable certainty, say what the generated assembly will look like for a given computer architecture (at least pre-optimization).

> The same cannot be said for an LLM-based solution to a particular prompt.

This is the correct and meaningful difference to point out here, but it has nothing to do with determinism. LLMs could be perfectly deterministic and still suffer from the same problem. The problem isn't that LLMs themselves are nondeterministic, it's that language is imprecise, and language models themselves are (for the most part) black box text processors. An imagined piece of functionality ("feature") has to pass through both of these somewhat opaque steps before it ends up as code, unlike code that is processed by a compiler, where both the constraints and the structured /formal understanding of the input are much stronger which allows us to reason about and trace the relationship between inputs and outputs in ways that we can't for LLMs and natural language.

I've seen these posted and amplified in a few places over the past few days now, has anyone who is familiar with the adobe suite actually tried these in depth and can comment on their fidelity/quality?

> While the author didn't make proper use of it, this is actually the best example of the articles point.

sure, but they don't actually illustrate that, they just point to a different "starting point", committing the same fallacy

The irony of this article being fully LLM-written

While this applies to many conflicts, it is very pronounced in I/P specifically. I find it a bit ironic that the author picked this example where, for almost any individual event or date in its history that seems particularly horrible or evil, one needs only to go back a few years to get a very different impression of the conflict.

Appreciate the transparent process! Is there a reason why you chose to process the entire list of options as one input to the LLM torso instead of splitting them and processing options separately? From what i can tell from my (limited) experiments, there is a lot of variation caused by simply reordering options, plus a heavy bias toward the first option listed [0], and my guess is that this is caused by hidden state of each option "leaking" into subsequent options

[0] https://news.ycombinator.com/item?id=49987076

... so training? Even "regular" pretraining is "inference plus training at the end" if you want to frame it that way

as far as gpt-5.6 could help me understand the code in their repository (and as far as their architecture is accurate [0]), the "hidden state at each option" refers to the hidden state at the last token in each option, which is then scored by the (learned) pointer head against the hidden state at the <answer>-position.

To vaguely confirm that this is plausible, i tested the sample query from the post with differently arranged options (This should produce slightly different outputs since each option influences all subsequent hidden states, including those of the other options):

"billing,sales,retail" produces billing -> 0.843, retail -> 0.092, sales -> 0.065

"retail,billing,sales" produces billing -> 0.470, retail -> 0.468, sales -> 0.062

"retail,sales,billing" produces billing -> 0.517, retail -> 0.415, sales -> 0.068

"billing,retail,sales" produces billing -> 0.803, retail -> 0.146, sales -> 0.051

"sales,retail,billing" produces billing -> 0.647, retail -> 0.127, sales -> 0.225

So each of these arrangements favor billing, but with pretty different confidence scores, and it looks like there is a pretty big bias toward the first option listed - which makes me wonder how viable it would be to have the torso process these options independently instead

Their docs say:

> An option's score depends on what it says, not where it sits. There is no per-option parameter anywhere in the model, so nothing can learn that "the first option is usually the right one"

This isn't quite correct. There might not be an explicit "per-option" parameter, but the hidden states themselves "contain" all prior options. My gut feeling tells me they messed up data augmentation by incorrectly permuting options during training.

Edit: the above numbers were obtained with the v19 version/checkpoint, v21 is much less sensitive to ordering, but still shows a first position bias.

[0] https://github.com/strands-labs/strands-decider/blob/main/do...

> AI Companies as Bad Actors > If you believe that the big AI companies is [sic] run by people who don’t have humanity’s best interests at heart

I believe that AI companies (much like any other company) are primarily driven by financial motivations (note that this might not extend to the individual people working there). Now often this manifests in ways that serve humans or even humanity, but other times it doesn't. This isn't unique to AI companies, and if you tried to only support companies that put humanity over finances, you'd be hard pressed to find many.

> Motivated by Bertrand Meyer’s AI for Smarties, if you believe that software engineering is the careful application of the mysterious knowledge of software engineers, encapsulated in rules like the SOLID principles, “prefer composition over inheritance”, “use you’re type’s good“, the Law of Demeter and so on, then AI can’t possibly write software because it doesn’t apply those rules, it just generates text that happens to compile.

... is this the "LLMs don't truly understand what they are saying" of software engineering? I'd also say there is a case for re-evaluating some of the established engineering principles, considering that they were built on assumptions that are no longer true.

> I guess I could try to cope by pointing the AI back at it but for the first time I’m forced to color outside the lines just so I can deliver.

You could, but then you shouldn't be surprised when nothing changes. If someone had (repeatedly) filed unworkable tickets/specs pre-AI, what would you have done?

It occupies the same space in my brain as the clickbait-type "This is the key" and "what you're missing" LLM-isms

Universities teach one of the most influential and ubiquitous programming languages, crazy

Would even recommend peaceful mode, biters to me always felt like the least fun part of the game, especially while still learning how things worked. It's like you have a coworker come over and interrupt whatever you were doing, except instead of doing small talk they close half your tabs, clear your browser data so you are logged out of every site, and steal your coffee mug

> I suppose that puts me in that group of programmers that are focused on the end-product, agnostic as to how the thing was made. I have truly never enjoyed the plumbing part of the job—just want clean tap water.

FWIW I think most people don't belong to either group exclusively. I often care mostly about the outcome (e.g. at work or when i make small tools for myself), but occasionally i care mostly about the process (e.g. when learning a new programming language or solving problems akin to advent of code)

I think "tools for personal use" implies both that their utility is more important than craftsmanship, and that they aren't public.

is that how jev works too? how do these differences make those approaches more "instinctive and emotional"? I know they're different on a technical level, I'm asking which characteristic difference necessitates the use of this loaded term from pop psych

so whats the difference between this and a non-reasoning LLM, or just any generic classifier method that necessitates a new term? there is nothing more intuitive or automatic about jev or this than any of the other currently used AI models

You're describing external aspects but to me, "instinct" says much more about internal processes than the properties you mentioned, and I haven't seen anything that tells me how these models draw on anything similar to these internal processes to generate their outputs (at least not more than generic LLMs do)

i'm well aware of the origin and meaning of the term

> "System 1" is fast, instinctive and emotional

this implies more than just "fast", which is precisely why i don't like its present usage

I really hope that "System One" won't stick around as a new buzzword simply meaning "fast".

if anything, using oh-my-pi opened that can more than using pi (or better: both!) would have