silentbob

Karma: 2,088

silentbob 16 Sep 2026 5:56 UTC
3 points
1
in reply to: leogao’s comment on: leogao’s Shortform
I have some doubts that you can achieve “equivalent in performance” and “human-understandable” for most of these. At least in so far as the given size is anywhere close to the minimal practically achievable size enabling that performance.
Would the idea be that the python code performs in a similar way to the model and hence help understand it? Even if it works, I can’t really imagine it being useful, as I’d assume the code to end up as an often arbitrary mess of case distinctions, non-obvious weighted averages and so on. Perhaps for a CNN this could be halfway decent, but even there I would assume the majority of computations the network performs to either be so abstract you can’t find good variable names for them, or exploit random statistical correlations from the training data that look like pure noise but are still relevant for performance.

silentbob 8 Sep 2026 8:05 UTC
3 points
0
in reply to: Haiku’s comment on: Ian McKenzie’s Shortform
I kind of doubt that people would be nudged away from their “doom is just marketing” lense by snappy reports. But maybe the question is more what others who then read the exchange take away from it, as many of them may not have any strong opinions yet.
I would guess that from the overall population, some small percentage is following the “doom = hype” narrative, and some probably smaller percentage is following the “AI is a dangerous technology and any incidents that happen show that we’re getting closer towards existential risks” narrative, with the vast majority of people either having heard of neither, or not even being aware of AI being a thing that deserves any attention at all to begin with. And I’m not sure how best to address that latter group.
By the way, I don’t think “doom is marketing” is completely false. I do think the narrative is misleading and that many individual cases where people claim that something is marketing, like the Hugging Face incident, make no sense. But it’s true that AI capabilities are correlated with their dangers. I think of it like this:
- if you already believe that AI is (going to be) a transformative technology, then pointing out AI’s dangers makes you less excited about AI and more willing to regulate the technology strictly
- if you believe that AI has no value, the bubble will pop soon, and we’re all going back to our pre-AI normal life afterwards, then any emphasis on the dangers of AI violates your view of its capabilities, so it looks to you like talking about the dangers is just meant to misguide you into taking the capabilities more seriously
So one can make the point that talking of doom makes some people take AI more seriously. I don’t think it explains much (or any) of the labs’ behavior, and I don’t think they overall short-term-financially-benefit from pointing out the dangers of the technology. But if you’re convinced that “AI is glorified autocomplete”, then I can see why any talk of the dangers of AI seems like “doom trolling”.

silentbob 8 Sep 2026 7:47 UTC
4 points
2
in reply to: Ian McKenzie’s comment on: Ian McKenzie’s Shortform
My best guess for the nature of #2 is that some people like Cal Newport or Internet of Bugs (and to a degree Ed Zitron) have a strong tendency to moralize issues, seeing the world through a lense of human intent being the main driver of everything that happens, and hence you constantly need to focus on identifying culprits and holding them accountable, or at least convincing your audience to condemn them.
If human intent is what drives everything, then talking about the technology itself becomes problematic, as it distracts from the thing that actually matters, which is judging the bad people. In a response to Anthropic’s post on recursive self-improvement, Newport said “I would almost say shame on Anthropic for dropping this report. No suggestions, no reassurance, no culpability, no responsibility.”
Newport is also particularly drawn to explaining some aspects of how LLMs work (with varying degrees of correctness), to then conclude that clearly they cannot have goals or intentions. When your main goal is to find the bad people and make everyone agree with you that they are bad, then you can neither grant any agency to AI, nor would you accept the more systemic issues or coordination problems to play an important role.

silentbob 8 Sep 2026 5:19 UTC
3 points
1
in reply to: Ian McKenzie’s comment on: Ian McKenzie’s Shortform
Narrative #1 seems like a fit meme: it’s fun to have and share this opinion, because it makes you seem smart and non-gullible (well, to some people), and it’s the sort of self-reinforcing trapped prior that you never leave once you’re in it, because any evidence can be twisted to match this narrative.

silentbob 25 Aug 2026 20:28 UTC
2 points
0
in reply to: Philip Harker’s comment on: Philip Harker’s Shortform
Congrats on the grant! That does sound like an exciting opportunity. As for the rest of your situation, sorry to hear that your job is dragging you down. Being stuck in such a situation sucks, and I guess the way forward depends on many details that us commenter don’t have much context about. Like, would returning to your parents be worse than staying in this job for longer?
Quitting with only a few months of runway sounds pretty risky to me, and chances are that much of the time that you would want to focus on working on your game, you’d be stressed out about your financial situation and where you’ll be at two months later (at least I would).
Maybe some other questions one could raise:
- is there any way to improve your working conditions, maybe reduce your hours so your situation becomes more bearable?
- can you reduce your living expenses and thereby gather more runway while working for another few months, after which you’d have more flexibility?
- is there any way to reframe your work or change your focus that would have non-obvious upsides? Maybe there are some opportunities to learn something useful while you’re there and make the best of the situation.
From what you’ve described, I’m not overly optimistic about any of these, but who knows.
Overall, I think something along the lines of “hang in there as long as necessary while at the same time looking for alternatives (whether a more enjoyable and/or better paid job, or more funding for your project)” would seem wise to me.

silentbob 25 Aug 2026 5:50 UTC
1 point
−1
in reply to: Brendan Long’s comment on: How the Sausage is Made—Why we Hate Slop
Indeed, they mention in the video description that AI was involved, which seems fine to me. (I also like it, the lyrics are cool)
I was speaking in more general terms here, should probably have clarified that.

silentbob 23 Aug 2026 6:13 UTC
4 points
0
on: How the Sausage is Made—Why we Hate Slop
For me, it’s primarily two things:
1. I really dislike dishonesty, and publishing AI generated content as your own seems quite deceptive. I’m much less likely to have a bad opinion of an AI piece of any kind when the human behind it is upfront about how it was made. Of course, not every use of AI necessarily has to be disclosed and it’s a fair question where the boundary should be. But when, say, someone posts a post on lesswrong where the majority of words are LLM-written, or publishes a song made with Suno—however much effort the human has put in—that imho deserves disclosure.
2. The basic question about distributions, priors, and how I can avoid wasting my time. I’d assume that 99% of clearly AI-created media (whether Suno songs or texts or shorts of inspiring old people in armchairs) really is low effort slop and I’m well advised to skip them as soon as I realize that. Sure, this comes at the cost of also skipping the 1% that do include deliberate effort and intention. But if the alternative is that I drown in slop just because maybe that instance is different, then I’m happy with being very occasionally wrong in my slop assessment. Notably, I likely won’t skip such cases if the creator is open about how it was made and perhaps even points out while it might still be worth my time.

silentbob 13 Aug 2026 19:12 UTC
59 points
2
on: silentbob’s Shortform
I was surprised to hear from three different people that they had some serious misconceptions about the moon and eclipses, sooo, just a few facts about our favorite celestial bodies for those in need. :P
1. the reason the moon usually has a crescent shape has nothing to do with Earth’s shadow and is usually entirely explained by the fact that half of the moon is illuminated by the sun, and depending on what the current angle between Earth, moon and sun is, we end up with different crescent shapes. Most of the time, no shadow casting is involved.
  1. A side effect of this is that the crescent moon is a bit like an arrow that points towards the sun.
  2. Did you know that Venus, Mars, Saturn and Jupiter are quite often visible in the sky? They (sometimes, when they’re close to Earth) are notably brighter than all stars, so when you see a particularly bright star-like thing in the sky, then it’s probably a planet. Specifically if it’s close to the extended line through moon and sun (or the line implied by the “crescent arrow”)
  3. Did you also know that you can actually recognize Mars by having a red tint? This sounds like BS, but it’s really true. If you ever look up at night and see an unusually bright star that appears to have a red tint, it’s probably Mars (but Mars is only brightly visible at the night sky for a few months every ~2 years).
  4. Another fun fact: “the dark side of the moon” is a phrase that does not actually refer to whatever side of the moon is currently dark, but it describes the “back” side of the moon. From Earth, we always see the same side of the moon, as it’s tidally locked. So what we call the dark side of the moon is actually illuminated by the sun half of the time.
2. In fact, Earth very rarely (~1.5x per year on average, and then only for a few hours) casts a visible shadow onto the moon. The reason for this is that space is huge and the moon is very far away from Earth, meaning that Earth’s shadow cone is relatively small compared to the space through which the moon moves.
  1. One could naively assume that every full moon (i.e. ~12x per year) should come with a lunar eclipse, because full moons happen when Earth is between the moon and the sun, and hence there should be a shadow. However, orbits are not perfectly “flat” but tilted by a few degrees, meaning the moon, during a full moon, is often “above” or “below” Earth’s shadow and hence evades it.
  2. Even when a full moon does lead to a lunar eclipse, the eclipse is only visible for a few hours, so lunar eclipses are easy to miss, even though full(ish) moons happen about monthly and many of us notice them all the time. Taking together that there are only ~1.5 lunar eclipses per year, ~half of which happen while you’re on the wrong side of the planet, many happening while you’re asleep, others happening while it’s cloudy, plus the fact that they’re rarely widely advertised, many people may have never actually seen a lunar eclipse.
  3. A total lunar eclipse (a bit less than half of the 1.5 per year) has an interesting look where over the course of a few hours you first see Earth’s shadow covering the moon from one side, then for an hour or so you see a very dim moon (which is usually tinted red, aka “blood moon”), after which Earth’s shadow disappears to the other side.
3. Solar eclipses tend to be a much bigger deal that people get excited about, but they actually happen about equally often as lunar eclipses. The reason they’re a bigger deal is that per location they’re extremely rare. So when they happen close to where you live, that’s special and potentially a once-in-a-lifetime event. Whereas a lunar eclipse can be observed from the entire half of the globe that’s facing away from the sun.
  1. Often, solar eclipses and lunar eclipses come in pairs, two weeks apart, as is the case right now: yesterday there was a total solar eclipse above parts of Europe, followed by a lunar eclipse two weeks from now. This is because the moon is currently in the phase of its orbital alignment where it intersects with the line between Earth and the sun; which makes it likely that two weeks form now, when it moved around Earth by 180°, it’s still aligned well enough to intersect with Earth’s shadow.
  2. Speaking of solar eclipses, when you get a total solar eclipse, the duration varies depending on several factors such as the moon’s distance to Earth and the Earth’s distance to the sun at that point in time (and from where on Earth you observe it). Spain yesterday had ~2 minutes of totality. Under ideal conditions, up to 7 ¹⁄₂ minutes of totality are possible. On the opposite site, you can also end up with an “annular solar eclipse”, which is when the moon appears smaller than the sun and hence cannot fully cover it, even when moving perfectly in front of it.
  3. Observing the solar eclipse yesterday from a location with ~90% coverage of the sun, I was surprised to realize that this was very hard to even notice without equipment! It did not seem to get notably darker despite only 10% of sunlight coming through. It most likely would have been easy to tell in a direct comparison, but I think had I not heard of the solar eclipse, I just would have not noticed at all that anything was different. The only equipment-less way in which it made itself noticeable was that all sorts of shadows and reflections took crescent shape, which was cool.
Some upcoming dates, see also here:
- Two weeks from now, on August ²⁷⁄₂₈, there will be a (near-total) partial lunar eclipse, possibly visible in e.g. Europe and America
- August 2 2027: total solar eclipse above parts of Spain and Northern Africa, with Egypt getting over 6 minutes of totality, close to the 7.5 minute maximum, and thereby the longest totality of the century
- December 31 2028: next total lunar eclipse, visible from about everywhere except South America and the Eastern half of the US

silentbob 13 Aug 2026 14:45 UTC
12 points
7
in reply to: gabrielrecc’s comment on: gabrielrecc’s Shortform
Thanks for the summary!
Interestingly, my impression often is that these views:
It’s not an “agent”: AI doesn’t understand or want anything; humans configured this, and anthropomorphic doom-talk misleads
The alarm itself is a product: doom discourse is fear-selling and marketing hype (IPO timing, bailout bids); company accounts can’t be trusted
are the most common—at least I encounter them constantly, on all kinds of places (Youtube videos + comments, reddit discussions, hackernews, work colleagues). But I can’t rule out that some filtering beyond my control or awareness is skewing what I encounter.
In your sample, they seem less popular, with only ~12% of comments and ~8% of votes, which seems like a good thing.

silentbob 13 Aug 2026 5:59 UTC
9 points
7
in reply to: MichaelDickens’s comment on: LLMs Are Starting To Noticeably Accelerate Our Work
I think neither one year ago nor now nor anytime soon will “LLMs improve productivity” have a clear binary answer. It’s a jagged landscape that depends on things like
- The domain
- The concrete task
- The human’s experience
- How they’re using the LLM
- How conscientious they are in the process
- ...and 50 other things
And with every year, a larger part of this landscape comes up positive for LLMs.
Ironically, I can imagine that part of what leads to the relative improvement of LLM performance is that (some) humans, due to extended LLM use, have seen their skills atrophy, and hence the bar was lowered. I think it’s very difficult to use such technology without “moving backward”, as John has put it.

silentbob 11 Aug 2026 15:05 UTC
3 points
0
in reply to: J Bostock’s comment on: Jemist’s Shortform
What LLMs are doing doesn’t look like “interpolation” of the training data to me, in the most commonly-used sense of the word.
Interpolation usually means taking the convex hull over a set of points
I’d say that even if it was interpolation, this would still mean we’re interpolating within an absolutely huge space of tens of thousands of dimensions, the overwhelming majority of which is entirely unexplored. Interpolation can seem a bit basic when you look at low-dimensional cases, but I just don’t think that applies the same way to higher-dimensional spaces.

silentbob 11 Aug 2026 14:53 UTC
3 points
0
in reply to: Ameya Panchal’s comment on: Ameya Panchal’s Shortform
Not to be mean, but the above reads very LLM-like to me. Even in the case that my impression is wrong, I find it very difficult to understand what point you’re really trying to get across. E.g. this part:
I did not understand the math, but I knew what to look for: how long the sort can run before motion stops reading as pixels, what edge detection has to preserve for a face to survive. When a render came back wrong, I could say how. It worked.
I have no idea what this means. This is the first time you mention “sorting”. What was sorted? And why? What does it mean that “motion stops reading as pixels”? What edge detection? What does it mean for a render to “come back wrong”? What “worked”? So many questions. The same is true for most other sentences, fwiw. Like “What varied was how specific the target was”. What target? Why did it vary? But I could really quote almost any sentence here.
I’d appreciate if you could write a human paragraph on what you want the reader to take away from all this. My best guess is that you vibe-coded a thing that is cool and you learned something in the process and want to share your insights. But I don’t even understand if you considered this project a success or not. Early you say “It worked.”, later you regret that you didn’t end up with more satisfying animations. How does “When a render came back wrong, I could say how.” relate to “What perplexes me is that I could never localize a failure.”? These seem contradictory. Oh well. :D

silentbob 4 Aug 2026 8:38 UTC
7 points
5
on: silentbob’s Shortform
Useful hint for vibe-coding: if you’re building something mostly for yourself but that should work cross-device, you don’t need to deploy it to the internet, but can have a server running locally that’s exposed to your wifi, which allows multiple devices to easily talk to each other, allows easy file persistence and so on. Comes with much less friction than having to set up some account/backend/database somewhere and likely has lower security risk (but certainly not 0, as people may be able to connect to your wifi without you knowing). (Edit: one more advantage is that a coding agent can end-to-end test what it builds more easily this way)
Some examples of what I’ve used this for:
- using my phone as a remote control for my PC, letting my phone act as a touchpad + keyboard so that I can control my desktop PC (which is connected to my TV screen) from the couch without having to buy extra hardware like a Bluetooth keyboard. There’s a tiny app on the PC that opens a terminal where it displays a QR code which exposes a locally hosted web app for the PC controls. So I have to scan that once initially, and then my phone is connected. Any phone (that’s in my flat and in my wifi) can scan this QR code and then control the PC through its mobile browser^[1].
- for a recent party, I vibe-coded a side quests app. Every guest who was in the wifi could follow a simple link to open the web app in their browser. They insert their name, which is used as a seed for a personalized side quest selection. (at previous parties I just used Claude’s integrated web app hosting for that, which worked just as well and doesn’t require people connecting to your wifi; but then this app gets public, which you may see as a disadvantage, e.g. in case some side quests reveal personal information about people)
- for the same party, I vibe-coded a “photo station” app. I had my laptop set up in a well-lit corner of the party, with a web app that accesses the webcam and has ~20 different modes of doing fun things to the webcam contents, e.g. replacing the background with nice pictures / fractals / memes, or surrounding people with animated shoggoth-esque tentacles, or just blurring the background for portrait shots, or many other things, including some fun interactive modes where people can play around and create cool images featuring themselves. This app also hosts a gallery of all photos taken during the evening (which are stored on the laptop’s hard drive). The gallery, when displayed on the laptop, contains a QR code which people can scan to see the entire gallery on their phone (as long as they’re connected to my wifi) where they can then download these images.
Just knowing that this is possible unlocked a lot of cool options that, if I had to rely on hosting things on the internet, I likely would have dismissed due to friction/risk/implications being beyond what I’d find acceptable for building any given thing.
But of course, be careful with this. Ensure your wifi is well-secured and that you trust the people you give access to it.
1. ^
  Out of caution, I ensured to only let one device connect at a time (meaing when I connect my phone, I can be sure nobody else is able to connect even if they find the QR code), and the server shuts itself down after some time of receiving no input, so I can’t accidentally leave it running and forget about it. There are some more security features baked in, but probably not worth getting into these here. Plus, if any individual was sufficiently motivated to attack me, they’d likely find easier ways than through this app anyway. I basically ruled out accidental as well as spontaneous misuse of it, which seems sufficient.

silentbob 4 Aug 2026 6:04 UTC
9 points
5
in reply to: Philip Harker’s comment on: Philip Harker’s Shortform
I don’t get the downvotes, I think it’s an important question to ask when I arises.
To me, finding undisclosed LLM text just is a very strong indicator the person publishing it wants to get away with sloppy work and doesn’t respect (the time and epistemic environment of) their readers very much. Clearly this doesn’t apply to every case, it’s just by far the most common occurrence, so by posting LLM text, you’re moving into a different reference class. The most likely scenario behind someone posting LLM text, purely on very general priors, is that they wrote a short prompt and got a slop response. If the person invested more effort into the interaction and has reason to believe their LLM text is of high quality and worth reading, then they should point out why (which involves being transparent about LLM usage in the first place).
I think I’d still have this preference if there was only a single LLM in existence which was both aligned and superintelligent. Text is supposed to have effects on readers, so being transparent about the nature of a text just seems like the right move.

silentbob 4 Aug 2026 5:55 UTC
2 points
0
in reply to: Gavin Runeblade’s comment on: Philip Harker’s Shortform
I think the subcontractor/robot situation really depends. There are cases where I hire someone purely so that something gets done, without caring about who does it. If they send another person, or a robot, I may be entirely fine with that. But there are cases where this doesn’t apply, say when I hire a person because they specifically have a great reputation or I otherwise think highly of their work and I care a lot about details and quality.

silentbob 2 Aug 2026 20:18 UTC
10 points
0
in reply to: Richard_Kennaway’s comment on: Richard_Kennaway’s Shortform
Here’s where it gets interesting.
This is the point most people get wrong.
Then there’s Y.
What I keep coming back to is X.
The things that surprised me most is Y.
Something something “on Tuesday”.
The X <verb describing a change or increase>ed. The Y barely moved.
Things “land” and “sit” all the time and everything happens “quietly”.

silentbob 21 Jul 2026 20:35 UTC
2 points
0
in reply to: Togame’s comment on: leogao’s Shortform
touch your pupil without flinching
Am I the only one cringing at this so heavily that I’m basically 100% certain I’ll never wear contact lenses in my life? I have 0 intention to ever learn such a thing. :O

silentbob 21 Jul 2026 7:12 UTC
3 points
0
in reply to: silentbob’s comment on: NameRank: The Model Knows Your Project, Not You.
Another thought: every individual may be involved in many projects. So if you look specifically at their most well-known project, then that’s somewhat of an unfair advantage for the project. Whereas if you were looking at a randomly selected project that a person published, then it may well be that that one’s less well-known than the person themselves.
E.g. the study you (OP) linked says:
Pair a well-known artifact with the person who built it and, in 7 of 10 pairs, the artifact wins. The reinforcement-learning library Tianshou scores 0.78 while its sole author Jiayi Weng sits at 0.22 — the work propagated; the name behind it did not.
I’d say this is not surprising, as when you condition on “well-known artifact”, you should expect something like regression to the mean in well-known-ness when you move on to another entity.

silentbob 21 Jul 2026 7:03 UTC
3 points
0
on: NameRank: The Model Knows Your Project, Not You.
I’d say it depends on the person and project—sometimes a project gets really well known and gets a lot of coverage (in pretraining data) while the creator is only mentioned occasionally. Similarly, I know hundreds (maybe thousands?) of pieces of software by name, but only for a fraction of them could I name the creator(s). But then you have other cases, where people get better known than many of their projects. E.g. if you think of any actor, every LLM out there will likely know who they are. But they are less likely to know about every single movie they’ve made.
As a funny example mixing both spheres, actress Milla Jovovich was involved in creating https://www.mempalace.tech/ which will most certainly remain less widely known than herself.
(edit: while I linked to mempalace, it looks like pure AI slop. I haven’t looked into it more deeply and hence me linking to it is not a recommendation. fwiw, the NameRank website + paper also appear to be fully AI-written)

silentbob 18 Jul 2026 15:55 UTC
9 points
2
in reply to: dynomight’s comment on: Kaj’s shortform feed
I’ve seen some software engineers in tech orgs behave in similar ways, where I’m really unsure whether they actually managed to 10x their productivity using the super elaborate systems they constructed, with hierarchical teams of AI agents, each having different roles, and the humans just being their orchestrator, or whether they just get extremely excited about these possibilities and set up all these layers of “automation” without this actually leading to more useful output. I can certainly imagine it leads to more output, but I’m somewhat skeptical if it’s high quality, sustainable output, or just huge amounts of ~slop.