Clojurians Log v2

Is it from the same guy who recently reslopped Photoshop and other Adobe bloatware?

dunno

I like how they dare to claim that it is a clean room implementation

I'd definitely wait and see. This one has had 13 pages of GH issues created in just 3 days. :) The Photoshop "clone" also suffers from it. Superficially it works, but it's actually a not-working software underneath. (Not claiming Photoshop "just works", but at least seems more reliable.)

For proper reverse engineering you would need access to a binary? Maybe a closed source SaaS is a bit harder to reverse engineer? More like decompilation?

true... like... reverse engineer claude?

i’ve heard the rust photoshop things are terrible. But that future is probably very close. Perhaps licensing is what prevents this. But i wonder if in 14 months you can say “i need a photo editor” and one is created on the fly for you with sufficient quality that a pre-built one holds not sway?

or you just tell claude to edit your photo? I mean, why even bother editing it yourself?

The real question for me is why any person in their right mind would want to have or build a clone of Word? Especially if there are open source alternatives like LibreOffice?

My wife uses Excel instead of LibreOffice because she likes it better 🤷

it’s the ultimate in the promise of open source. “fork it and adapt it” is so trivial now. If a user wants an editor that has different UI stylings, minimal UX, focus just on the words. built in collaboration, etc. There are a ton of reasons you might want to customize some notion of a text editor

Word hardly feels like the golden standard, but that is probably personal. I still look back with some fondness to my LatTex days 🙈

> why even bother editing it yourself? Depends on the operations that you need to perform. It would take roughly the same amount of words in a prompt to (1) create a small app with very well known functionality or (2) add a short transducer chain at a particular location in your code. So if it takes me 30 seconds to convert mentally the code I need to write into a prompt, write that prompt, wait for the agent to finish, and then check its work, while it takes me 5 seconds to do it manually, there's little reason to still prefer the former.

But also, there's always a way out of this for them. "Word documents certification" - bam, here you go. Anything can create DOCX, but only Word from MS can create the DOCX. One that's guaranteed to be safe, because nobody wants to be unsafe. Large institutions love that shit. "No, we do not accept ODT, only DOCX. Your DOCX indicates that your copy of Windows has not been activated. Goodbye."

I’ve tried Artcraft quite a bit, and there are a lot of issues that Photoshop doesn't have. Tough for someone who doesn't use Photoshop professionally, those might not be dealbreakers (but such a person doesn't necessarily need Photoshop anyway). The bigger comedy is watching the PRs: multiple different approaches to solving the same thing, a lot of conflicting code / ideas, and it's completely beyond the maintainers' capacity to review it all... It’s a perfect showcase of what happens when you motivate anyone with a coding agent to contribute, without anyone actually understanding the product or having a cohesive vision behind it.

I wouldn’t be surprised if Nintendo and some other major IP holders are biding their time, watching which legal arguments against AI providers hold up in court. If a clear opening emerges, things could get messy fast.

I edit photos myself instead of asking for Claude because technology is (maybe by accident, probably) biased to people with lighter skin tones. Every "automatic edition" (including the one in Lightroom that fixes lightning) makes my wife either a shadow, or make her seem like a painting. Same for word - my mother works for the government as a translator. There's a lot of requirements on headers and footers, and Libreoffice simply doesn't support (I ended up installing a Microsoft Word over Wine in a Linux machine because I was tired of giving support for Windows. Surprisingly, 5 years and a single problem that I had to intervene)

i used claude to build an editor to invert scans of negatives for me. but i want the final control.

BTW - I've been testing this tool they are using to reverse engineer the binaries. In my case, I had an old game and I asked Claude to patch it with some features... ... surprisingly, it worked

(After a lot of convincing, it was insanely slow, took a long time to finish and took 5 tries, but it worked)

my sentence "why don't you ask claude" was a reaction to "what if in a couple of months agents are so fast that they can hammer out a photoshop for you in 5 minutes" (don't remember the exact wording). in that case editing a photo probably happens in under even a second if you can come up with the right description. not that I would do that now, it's a terrible experience imo

> just tell claude to edit your photo? I suspect a probabilistic machine is not reliable enough to edit binary images to professional needs. A small model embedded* into photoshop that could invoke APIs might do this better. "Photoshop - put the the background into a layer, lighten it, and save the layer stack as xyz." (* "agentic", although I'm growing to dislike that word)

yeah sure, the harnass or whatever that is could be smart about figuring out how to do it

“to professional needs” “make this image pop for our blog” is probably gonna happen, or actually, probably already does happen

> true... like... reverse engineer claude? Do you mean DeepSeek or Kimi K3? 🙂

https://clojurians.slack.com/archives/C03S1KBA2/p1791068287204419?thread_ts=1791066447.946919&cid=C03S1KBA2

Yeah I think that's part of an intuition you start to develop on when you know whether it did its homework properly and when to push back. Awhile back I witnessed another dev use claude and constantly jump to conclusions that were unfounded, but the dev did not know that.

Another item I would add to the list is that ai's ability to cheaply launder IP theft is likely to cause more folks to avoid publishing or sharing their ideas.

@smith.adriane Absolutely. I had planned to publish a blog detailing my experiences learning Clojure, but now I figure - what's the point? Few engineers read blogs anymore, so a blog may well only be scanned by the AI. It's almost all downside for me. See also: The Pragmatic Bookshelf no longer publishing.

Another thing to add to the list: deskilling. Rich got at this in his tool vs automation gist. It's been shown that people leaning too heavily on LLMs' "thinking" start to lose skills and think less critically about problems. I'll also add that without discussions around problems, community is going to need to be much more intentional.

I've already seen it with my coworkers. I don't use AI but they do. And I can see their judgment slowly decreasing in quality.

I'm skeptical that we have the legal and economic frameworks to handle AI. From a tech perspective there's some ways to go, but nothing insurmountable.

Also, I'm quite bullish on renewables, so I don't believe compute will end up being an issue.

When you don't feel the pain, many problems won’t be known as a problem. I just worry about they might still be a problem.

@richardc.2074, @shiyi.gu - yes. On the subject of renewables - I want to believe, but at least for now, in the US datacenters are using relatively inefficient turbines (compared with the grid turbines)

I'm still skeptical that overall AI will result in my life being better, my kids' life being better, my job being more fun and enjoyable, my social interaction being fulfilling, and so on.

I'm also skeptical that if you take actual impact into account from a business standpoint, if using AI to develop software is having a dramatic impact? And is it actually making projects go faster and product be materially better?

And I'm a little skeptical if AI is enabling us to do things we couldn't do before and expanding the space of what is possible, or if it's just giving us the same thing but more of it and maybe faster or cheaper.

having been inoculated against Effective Altruism relatively early on in its history, I've been skeptical since OpenAI was founded that the main AI labs can be trusted at all to act in anyone's best interests, much less the future generations they claim to speak on behalf of. Nothing since then has changed my mind. This story is a helpful and thorough timeline. https://www.iankduncan.com/personal/2026-09-16-sex-ai-and-the-apocalypse

@didibus said it very well, so I'll just summarise what I'm sceptical about: I'm sceptical that AI is good for our industry, our society, and our ecosystems.

On the renewable energy front: datacentre power demands are growing way faster than renewables are being added to the grid. I'd be optimistic on renewables too, if we just stopped sucking up every new MwH into our GPUs.

I admit that LLM coding tools are way more capable than I would have expected a year ago. This is largely due to the harnesses getting better. Sure, models are improving as well, but it was the explosion of MCPs and connectors and all of this stuff that really unlocked the effective application of models to the various tools we use (both coding tools and also the general SaaS landscape).

However, this is a double-edged sword. Making things easy to do tempts us to just do them without thinking about whether they're the right things to do in the first place. Easy not simple is not what we're going for in the Clojure community, right?

I, for me, can confidently say that AI is helping me a lot to fix the issues on the completely unfamiliar codebase that I am working right now on my job... ... but at the same time, I am 80% sure that these issues are happening because the AI itself put them there

I have seen LLM agents being very helpful to senior engineers in the short term. What I haven't seen is the long term health of codebases developed this way, and more importantly, the long term health of the people working in codebases developed this way, whether they are using the tools themselves or simply co-existing in a codebase where others are using these tools.

After a quarter of a century in this "programming business" my main rule has always been, whether building a product or consulting, that "YOU DO NOT NEED SOFTWARE". And this is a call to dialogue about the actual problem at hand. I want the other party to clearly articulate their problem and argue why software is the best solution for this particular problem. If they succeed, great, let's build great software together! With AI everybody just vibes it. There is not enough of scrutiny about the issues and problems when you can very cheaply just generate code. And the major problem is in the details, you know, maintaining software costs money, requires fixing, may disrupt your own business, creates issues in the non-functional requirements -zone and so forth. Someone has to think about software in the context of other software, like integrations, data consumption, storage etc. And don't get me started on the resource hogging, utilitys rates, environment and power shifts.

I feel like professionalism has taken a major blow with the advent of AI.

Just in case people in this thread aren't familiar with https://politechs.dev/index.html, it is a long form discussion of reasons to be sceptical of AI.

What can it do today which it couldn't do a year ago?

Plenty, actually. One of the examples that's trivial to see in comparison is video generation. If properly done, it's incredibly difficult, if not impossible, to tell whether something has been AI generated.

Ah, indeed, the evil usages of sophisticated tech..

Pretty much every aspect has gotten better, sometimes incredibly so. But it's not always easy to tell. And there are still uncountable areas where the current AI is just a bad tool for the job.

And that ended your skepticism? @p-himik

Not sure how to unwrap the question. "Skepticism" is not a binary value. There are areas where I know I can use an LLM and have a result that is more worth it than a human (me) doing the same job. Many more areas where I wouldn't even approach an LLM. Plenty of areas in between.

A year ago you were extremely skeptical that LLM will be able to generate better videos? I'm just curious what @naomarik768 is mentioning.

I'm skeptical of AI as an industry due to its externalities (resource hogging, utilitys rates, environment and power shifts, etc) that are not factored in to its costs, not regulated properly and so forth. We have not even touched the issue of using stolen training material and the effect AI has on publishing new stuff which is crucial for AI development as it is built on human effort and output. I use a local AI daily, mainly as a rubber duck, reviewer and generator of new approaches when I'm tired or feel like being blind to my own code. There are valid use cases for AI no doubt.

> A year ago you were extremely skeptical that LLM will be able to generate better videos? I read the OP question differently - the "extremely skeptical a year ago" sounds to me like a context for the question, not the required framing for those who answer. In any case, I doubt I had any useful thoughts on the topic a year ago. :) I was barely using LLMs, barely reading anything about them.

A year ago I did not trust LLMs to write much of the code I'd write myself. I did most the work myself manually and if there was something obviously easy/mechanical I'd pass it to them. That changed exactly when Opus 4.5 came out, and it has been getting better since. What's also getting a lot better is my skill using them to ensure I get the right outcome each time I let it do something. So my own skepticism was with regards to using LLMs for coding. If I touch on the topic of being skeptical about AI in other areas, I'm enormously skeptical about things like Hermes, Dots, etc where you hook up your apps to something that you do not control, do not understand, you do not control the context, you did not write the system prompts for, has access to the internet and whatever else, and expect it to behave full time and not cause security issues for people.

I feel like there are massive security issues waiting to happen (and a lot have already occurred) and I'm hoping no one at the bank I use that has access to important stuff is using any of these AI tools that're out there.

What prompted my question about this now in this channel was seeing what has recently come out that're AI driven projects: • Diablo 2 Resurrected in VR • Diablo 1 as a 3d game • Literally steamOS as an android app (DroidDeck). • Adobe suite being cloned as opensource

Did it really clone Adobe suite, did you try it? I saw many bombastic claims, most ended up as pure marketing. Like Mythos dicovering 26k security holes, OpenAI killing math etc.

@jussi.mononen professionalism in software development was never that high in my opinion - lots of "people trying random things until it compiles" in the companies I worked with, and unfortunately because of my experience with Rails, lots of people that understand nothing about databases, etc. But I think that in the past, we could have these discussions - the lack of professionalism had real consequences (bad code that slowed down new features, bugs exploding in production, hard to keep talents in the company, etc). Now, it's evidently worse - it takes less time to prompt and """"""fix"""""" things than actually solve the underlying problems (lots of quotes because in my experience, the "fix" is more like a band-aid and a kiss over a gunshot). As a real example, on my current job we're having insane technical issues regarding data consistency and the "fix" so far is prompting Claude to ask what's wrong (and it does produce convincing documents showing the issues, but they are only 80% correct)

The models have gotten better for sure, but as someone who used LLMs heavily since 2023, you could always use them for some things, even back then. What's changed is long-horizon tasks, and the size of task they can autonomously work on at a time. The appearance of "agents" definitely improved the UX, no more manually copy/pasting edits. It seems for coding, the big improvement came from reinforcement learning from verifiable rewards. Basically, you tweak the weights towards what generates working code. Prior it was mostly just RLHF, reinforcement learning from human feedback. The downside of RLVR is that the prose has become worse, when it tells you what it is thinking and doing, the way it talks is less aligned with human feedback as it gets tuned towards verifiable rewards. Since I have been a heavy user from a while ago, I was never skeptical of the capabilities, I think of lot of the disappearance of the skepticism is not due to the newer models being better, but to more and more people actually trying it out, getting experience with it, and realizing that ya you can find utility in it.

> it gets tuned towards verifiable rewards And showing itself to be correct too. It's insane how Opus 5.5 hugely tries to be "right" - it insists on things like "the error is here" and fail to show trust (it constantly says when I ask for help on a ticket "The conversation is clear that you need to do X, but please confirm it beforehand just to be sure"), and last Wednesday it insisted on a specific code that was a problem, and when I asked to make a script that called a background job in foreground (supposedly it would generate a CSV then send an e-mail - I asked for the script to run in foreground and not send the e-mail) Opus "tailored the script" in such a way that the thing it said it was a problem was indeed the problem (essentially, it bypassed the CSV generation that was problematic)

For "Claude's point of view" the reward existed, and was verifiable. But if I trusted Claude only and put the thing in production, it would not solve the issue we were having...

I once accidentally wiped our shared family desktop computer just after the turn of the millenium because I wanted to try a new Linux distro out. I lost all my earliest digital work (Flash animations and the source for some small games I made). That's how I learned to make backups.

Unlike you, the AI learned nothing.

"Dog ate my homework", version 2026.0

@paul AI doesn't dissolve people of their responsibility to back their important shit up.

No worries, they can Seadance the season anew in a few billion tokens.

I thought this was supposed to be the anti AI channel

@ben.sless coming out as radically anti-swedish television

Hey, I don't discriminate, it's not just the Swedes.