what I am curious about is - what is more expensive, the GPU compute of a traditional fine-tuning loop, or the LLM tokens spent trying to lockpick together a solution.
It's impresive - but it definitely feels like it takes a lot of luck and brute forcing with a ton of tokens. Nevertheless, if we are able to train tiny models to be really good at weight-space programming, then the calculus may change.
I think this would’ve been very handy to me a couple years ago when I was learning to use Blender and asking LLMs for help performing certain actions like “how do I display the normals of all the vertices in my mesh?” I spent a lot of time trying to figure out which button the model was talking about.
Interesting project! I scan 35mm and medium-format film with a mirrorless camera, macro lens, and Lomography film holder, then convert with Negative Lab Pro in Lightroom. I’m happy with the results, but the Lightroom subscription has me considering writing my own converter.
Could you share more about your NumPy pipeline, or publish the code? I’m particularly curious about how you use the blank film reference and roll-level measurements to handle the orange mask, color balance, and tone curves. Do you use camera/scanner calibration profiles or film-stock-specific adjustments, or estimate everything from the scans themselves?
Also, your screenshot appears to show an OpticFilm 7400. Would you recommend it, and roughly what did you pay? Have you tried camera scanning? I’ve been considering automating film advance and capture on my setup, but consistent color conversion is the part I’m least confident about.
You should check out the negpy project. It in my view has the best inversion process. It claims to emulate the whole analog print chain in the inversion process. They have a fold works cited.
It also automates most of the process. It works with SANE, some other legacy scanners, and camera scanning. It can even do trichrome scans automatically.
I've played with this lately. I'd pair it with something like VueScan as the scanner setup is a bit lacking still, but overall it is really impressive. The only difficulty I found is coming into the terminology it uses fresh - it's quite different to most other photo software in that regard.
How lately? They have been releasing quite frequently, and have worked a bit to clean up the UI and workflows after adding a ton of new features. I've been using it and it's working quite well for me after version 0.60-ish
About a month ago I think? It didn't have enough options to run the Coolscan 4000 ED I have lying about, at least not reliably, and I don't think it had any batch options (might be wrong about that). I don't really mind going through VueScan first but I'll take a look again soon and see if that area has been improved.
CoolScan support has improved a lot in the last month or so. I don't have one so I cant comment in practice, but its a very active project and I'd definitely encourage you to submit bugs if you have a coolscan.
Ooh great I'll check it out. I have a couple of them actually, although I couldn't seem to get one working through the complex series of adapters I had, so I'm gonna set up a new machine with an actual FireWire card and reassess it then. I do at least have one that works in VueScan so I can at least provide some feedback (it's a modded 4000 ED that does whole rolls).
A colorist I worked with did something similar by using reference film stock of the same film that was being scanned. This was anime 16mm film prints from the 80s that had been scanned multiple times and color set on arbitrary decisions by who ever was working with the film at the time. After using the reference stock to base decisions, the director said it was the closest he had seen to what he wanted.
The pipeline is configurable from the interface itself which helped me a lot on "guesstimating". I modeled my personal understanding after darktable's negadoctor feature. Takes the negative and determines a baseline, and with that come the configurations for color, light. The rolls I work with are sometimes funky and they're tinted differently, for example I have this Lomography Metropolis roll thats a bright neon vomit green. So I also needed a method of manually inputting that, so its captured as density and recorded in a library so numpy can guesstimate the right corrections that need to happen per stock.
Depending how you want to capture it can come from multiple areas, like a blank frame, the trailing* end of developed film, or the gaps in between frames. Interestingly enough along the way I was able to just skip cropping entirely so the script had some extra blank film to go off of for initial estimates.
I just uploaded the code in case you wanted to audit that: https://github.com/shannadige/35mm-studio. I know some others have suggested already-existing libraries that do this, likely better? I've not the time to test it out.
(P.S. for what its worth I found your reply very human, in contrast to most of the wording/phrasing that I let Opus do on my first pass of the blog)
(P.P.S. - OpticFilm 7400 is definitely a good starter scanner. Great output resolution for the price. I see some folks are more partial to camera scanners, but this is easy and quite portable. I got it for 200$ pretty much brand new off of Ebay. I have no intention of upgrading it so I would recommend it. It does need to be paired with good scanning software though, and I think plustek does not maintain any software for Mac. I ended up buying Vuescan a few years back, and it is pretty solid in terms of available controls.)
Of course, it's impossible to know for sure what was LLM processed or not, but some of your posts (like this one) have been getting classified that way.
It is a very sensible and reasonable comment- literally all questions I'd want to ask to replicate the setup, in a format that's easy to parse. If the author did use LLM it did not substantially alter the content of their question.
I thought it was a reasonable comment and, like you, I thought they were good questions. I haven't followed what kind of watermarking LLMs are putting in their output that might be a telltale for their use so maybe it's readily apparent in a way that isn't to me.
I'd be mortified if a comment I made was accused of being LLM-edited or "created". I don't use any of that junk but if I'm reading LLM output and not realizing it I just might be adopting some of its mannerisms and styles with realizing it. That's chilling.
I ended up discussing it with the moderators and they interpret "transcribe and formatting" as contributing enough that they want to "educate" the user to not do that.
In my mind, using an LLM to lightly format a legitimate questions falls below any threshold I'd apply to AI-written writing. I often dictate HN messages to my phone, and in my case, personally edit the message (tediously), but if there was a tool that produced text that looked lightly LLM, I wouldn't care about it at all.
It's not a big deal because in five years, the idea of chiding a user for using an LLM to enhance their text is going to be considered ridiculous. I've lived through enough computer revolutions to know this; once, I didn't get a job (in ~1989) because I could type 80wpm but made a bunch of errors... on a typewriter! The agency told me they needed people who could type 80wpm on a typewriter.
For whatever it's worth: It's definitely phrased a bit strangely but it passes my own sniff test. If they really are using a dictation-oriented method for getting text into a box (as they wrote in a followup reply to themself) then that may adequately explain the strangeness.
There was a time when scanning and [photographic] printing of 35mm film negatives was what I did for a living; it's likely that I've touched more film than most people ever will. I think they're asking questions that are a good combination of solid, thoughtful, and practical.
I can see why this comment would trigger people's AI alarm, especially if it fits a pattern for this user. The questions seem legit at first glance, but completely uninteresting. Too general and specific at the same time, if you see what I mean.
It wasn't a generated comment at all! Read the comment, and then what the author wrote elsewhere saying they spoke it into their phone.
https://news.ycombinator.com/item?id=49948614
"To reiterate, those were legit questions. I used voice dictation to my phone, with an app that transcribed and formatted my message."
harassing a user for doing that is just unncessary and I dont't think it violates the site's guidelines. Further, the comment itself was high quality, and contributed to the discussion.
It's not a hill I want to die on either, ultimately the moderators run the site and make the rules and we comply with them at their pleasure.
dang believes that the "transcription and formatting" adds a signature. I entered into Gemini Pro (at the highest level of reasoning) and it calls it human written text and provides a far more detailed description of why:
"If an LLM was used here, it was likely just used as a grammar checker or "polisher" rather than a generator. If we had to flag anything, it would be this section:... .his reads like a standard, high-quality comment from a technical forum like Hacker News or Reddit's r/analog. If the user utilized an LLM, they likely wrote the entire comment themselves and merely passed it through a tool like Grammarly or ChatGPT with a prompt like "Fix my typos" without letting it change their voice or vocabulary."
I tried ChatGPT and Claude (all the most recent models I have access to) and they all agree: human text, slightly modified to be more readable.
Looking at Pangram's output, it only mentioned the last sentence as AI generated:
"Would you recommend it, and roughly what did you pay? Have you tried camera scanning? I’ve been considering automating film advance and capture on my setup, but consistent color conversion is the part I’m least confident about."
That's absolutely a sentence a human expert might write. Even if it wasn't, the content contributes to the discussion. Seriously, why would grammar checkers for voice dictation even be remotely an issue?
I used voice dictation to transcribe my rambling thoughts on the article, which was thereafter re-arranged into a still verbose but somewhat more coherent reply. I should have taken the time to reword it properly - my laziness is on me.
I was not aware the rule against AI-assited text at the time, and will be more careful next time. Nonetheless, I find it unfortunate that we've spend more time talking about AI usage in a comment, than the blog which prompted it.
In case it helps at all: I post that type of reply (https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...) precisely when the commenter is a good contributor who we don't want to get dinged by HN's software. Usually it turns out that people weren't aware of the rule (as was the case here).
(I've marked the subthread offtopic and collapsed it now)
I don't have a public writeup, but the general approach for orange mask removal is: scan a blank unexposed frame from the same roll, take its per-channel mean as your reference white, then divide each pixel by that reference per channel before doing anything else. That normalizes the mask out instead of trying to subtract a fixed color. After that, a per-roll linear stretch (black point from the darkest scanned frame, white point from the lightest) gets you a rough base. Tone curve is the hard part: NLP uses a nonlinear S-curve tuned per stock, and matching that from scratch means a lot of trial and error with reference images you know the expected look of. If you want a quick win before rewriting things, Lightroom lets you write develop settings to XMP so you're not locked into the catalog even if you cancel later; the actual RAW/TIFF processing engine is what needs a subscription, not the metadata.
Yeah, I get that form follows function, but I’m hoping for more native features of quality control, some sort of dashboard to catch slop. Also looking forward to more domain-specific harnesses.
Did you look into the workflow tools? The agent can write a TypeScript program that encodes a subagent graph, so you can codify a review workflow and ensure it's actually followed. Kind of neat, I think other providers also have these now, but it was new to me, and interesting to see how it worked.
Adding dashboards etc is probably where the "everything is a plugin" starts to pay off. They ship an agent preset for working on the harness itself.
Exactly, and the classical lock and key analogy ignores electromagnetic and hydrophobic/hydrophilic interactions, which we can fully model in classical molecular dynamics systems
That's not correct. Lock and key absolutely does take into account em and hydrophobic interactions. Also, molecular dynamics is nearly useless at modeling these things (which is why heuristic models such as alphafold, Rosetta, dominate)
Fable's ridiculous. It's flagging basic biology research questions as a security risk. I'm talking basic fundamental genetics topics that make working on any genetics-adjacent codebase unusable.
As a workarond, I have noticed that a better estimate of time is number of sessions required to fully implement, review, and test a given feature. You can ask the LLM to help you estimate how complex a feature will be during the planning phase. The number of tokens can easilly be translated to sessions, asumming that you normally aim to keep a session below 60% of the context limit. As a rule of thumb, a simple feature can be fully implemented, tested and reviewed in a single session. Whereas a complex feature might take multiple sessions. Then depending on your workload and attention span, you can have a decent idea of how many full context sessions you can actually plan and review in a single day. Of course, some features can happen in parallel but in general the bottleneck is your ability to understand the context that goes into each session.
As someone with ADHD, it’s a lot more nuanced than that. Coding agents can remove task paralysis, but they also introduce many other distractions. Being one prompt away from zero to one is a double edged sword, because it means any random thought, idea and side project is also a prompt away.
reply