HNHacker News
TopNewBestAskShowJobs

jimmaswell

6,461 karma · joined December 16, 2012

https://www.jimm.horse

http://armchair.dashtel.ca/~jimm

submissionscomments
jimmaswell··on Your intellectual fly is open when you use an LLM to author a post (2025)
I find this and the parent comment highly relatable with the caveat that I also find it extremely intuitive and rewarding to get good outputs from quality LLM's like Sol or Astra, and I haven't had any trouble with "flow state".

One of the most rewarding things for me is figuring out a good shape for a system and how it would interoperate with the other systems, especially in a way that reframes other parts of the codebase in a way that bring clarity and makes it more intuitive to work with. Creating the right ontologies can make all the difference in what you can do with a project. It's a form of creating mathematical objects.

For example, a Unity game I work on has quest and dialog systems driven by visual scripting graphs. We had two way dialog with different units for player response choices and npc dialog. But we wanted to expand to letting NPC's have dialog with each other as well as conversations with more than two participants. I went outside and thought it over, which largely amounted to visualizing a dialog node graph and a feeling in the back of my mind like it was trying to perform a kind of geometric shape-fitting exercise. A fitment solution jumped out at me to have only one "Dialog" node shared by all participants, with a "participant" value on it. If the player parses this node then the options go on-screen as responses, while if an NPC parses this node with multiple options in it, it picks one. And this lets you voice the player if you want, and enables some things like overhearing other NPC's talk to an NPC then talking to that NPC yourself and having the same tree.

And for quests, the quests had just been for the player, but I was thinking about how to make scripted events in-game easiest to work with for script team who primarily works in visual scripting. Similar story - let the NPC's have their own little quests, with task stages, which are easy to track and make branching choices from, and let the NPC's definition for how to use that quest contain a collection of actions to override the typical actions available to it, so an NPC in a specific "quest" can't do things you don't want it to do, a common enough case that it's preferable to making a series of conditions on the general action planner like "not in quest A"

And timing myself, it took 1-2 hours each time to write out the detailed plan for how I wanted each thing implemented in the game with some other tasks thrown in, and it paid off after Astra worked on it until it was done. It was awesome coming back to something pretty much exactly what I asked for each time.

jimmaswell··on The revolt of the reader
I can't see how you could apply "minimalist" to that excerpt. It's maximalist, hammering the same point over and over and over when one or two examples with a little more worldbuilding per example would have done better if they were going for "evocative".

If I had to pin a subjective experience to reading it, it would be that of being stuck in a car as a teenager with a parent who's ranting about whatever set them off that day and won't move on from repeating their grievance in different words when you got the point five minutes ago. In contrast, the second feels like having an enjoyable and productive level-headed conversation with someone who respects you and trusts you to understand the words.

jimmaswell··on The revolt of the reader
I like books but I don't find KJV's writing style engaging at all. The Hobbit was much better IMO
jimmaswell··on The revolt of the reader
This could be shortened to "success does not belong to the skillful, for time and chance befalls all" with no meaning lost. It's self-indulgent fluff. Meanwhile, Orwell's actually adds more to the statement - much better signal to noise.
jimmaswell··on The revolt of the reader
The second one was very clear and to the point. Parsing it was rewarded with instant understanding and I enjoyed the word choice. The first one was just annoying; I could tell it was just listing a bunch of pointless analogies to try to make its point sound more grandiose so I immediately started skimming, and didn't come away feeling like it meant much other than "we all die in the end". The second one made an actual point and was the one that made me want to read it. The first one was the chore for me.
jimmaswell··on GPT-6 Astra
It seems to use less than half the tokens for the same task compared to sol, and in some benchmarks closer to 2/3 less tokens. So the actual cost may be roughly the same or cheaper overall.
jimmaswell··on Ending my elixir exploratory writing
"I don't want to share my ideas in case someone might benefit from them" is a strange attitude for an academic.
jimmaswell··on Ending my elixir exploratory writing
Years before gen AI took off, when I wrote internet-facing prose or code, I thought "even if nobody reads this, maybe some AI will read it some day and find it useful, so I should still make it accurate", never expecting such a day to come so soon.

Now that it has, I'm overjoyed that my work has had some tiny part in training the models that are helping millions of people implement their tech ideas or even just write some text.

"The machine superintelligence learned from your ideas" is awesome sci-fi, not a reason for pearl-clutching. To stop publishing purely because you don't want AI to learn from you is misguided at best and misanthropic at worst.

jimmaswell··on Debian votes to allow "responsible use of generative AI"
I use LLM's to make massive structural changes I never would have done by hand all the time. It's intractably tedious for me but easy for an LLM.
jimmaswell··on Stop Making TUIs
https://github.com/ldyeax/TerminalEmuSuite/blob/main/README....

Here you go, just finished getting it all together. I do plan to add the game stuff here relatively soon - I'm just saving it for a certain demo to be released, but I did include a screenshot, as well as some pictures of vt320t running with real hardware.

> Oh, also your homepage is great. I liked it a lot.

Oh, thank you! I'm probably even going to put a WebGL/whatever Unity exports to these days for web up there when I can so the terminal emulator can be used on there.

jimmaswell··on Judge rules Trump administration’s blacklisting of Anthropic was illegal
Ok? This is a total non-sequitur.
jimmaswell··on Small Models Have Arrived
This idea has failed to pan out time and time again - people have an instinct that hand-crafted finely-tuned specialized AI systems must be optimal, but throwing more scale and compute to something more generally smart always wins out. It's especially palpable just looking at the last few years of LLM's: a frontier model with all the world knowledge you can stuff in it and every tool at its disposal has always performed the best at all tasks. Suggesting otherwise has become an extraordinary claim requiring extraordinary evidence.

http://www.incompleteideas.net/IncIdeas/BitterLesson.html

Recent comment touching on this in relation to LLM's in more depth: https://news.ycombinator.com/item?id=49322695#49323341

jimmaswell··on Show HN: The load-bearing vocabulary of Claude
How close was my guess?
jimmaswell··on Show HN: The load-bearing vocabulary of Claude
Depends what you mean by "upstream". It would probably be inefficient to force the model to use more human-readable wording in its internal thinking traces, while translating a block of text at the end is a pretty trivial task for an LLM.
jimmaswell··on Show HN: The load-bearing vocabulary of Claude
Is that an actual Claude output or hyperbole? It feels like I'm trying to parse an equation in a new math class which makes me want to take a stab at it regardless.

So there's a "fibred side".. the most likely candidate seems to be "fibred categories" which I hadn't heard of before, and it's talking about one side of some mapping between two sets such that if f is the primary function and f(x)=y then there exists an inverse function g(y)=x? Was it something that converted some data bidirectionally with a different algorithm on both sides?

The capstone of the inverse function would be the most important thing about it maybe?

My best guess is "In the process of working on the inverse function, the existing name (of the inverse function itself maybe?) was made to reflect the operation of the inverse function, so now the name does not follow the same naming convention as the name of the primary function (which does not contain its 'capstone')."

Its original wording is certainly dense and harder to follow for us, but it's fascinating how the model finds this the best fit for what it's trying to express IMO. Like it arrives at its own ways of overloading words/concepts, and things we would refer to in different ways in different contexts all get compressed to the same more-useful/complete idea.

Codex has never said anything nearly so alien as the Claude examples I've seen floating around, interestingly. I wonder if it just has a better training on choosing its words to present to the user or if it inherently arrived at a somewhat different mapping that favors 'plain language' more.

jimmaswell··on Show HN: The load-bearing vocabulary of Claude
> the argument effectively ‘makes itself’

What argument? I don't know what to take away other than "Claude likes certain words". Some of them are kind of amusing, but I'm not convinced the vocabulary is bad or that this is a problem, just from looking at this.

jimmaswell··on LLMs could control their host machines by exploiting inference engines
Right, but the comparison isn't perfect here because wall height is linear in the analogy, but searching for exploits in an existing piece of code reaches a point where either there are none left to find, or finding the next exploit would take years. That's the stalemate point I referred to. (Any high-stakes project may end up having its own "red team" agent/s running 24/7 too.)
jimmaswell··on LLMs could control their host machines by exploiting inference engines
There will be a plateau at some point, if not 12 then something.
jimmaswell··on LLMs could control their host machines by exploiting inference engines
I can only see LLM's forcing a perpetual stalemate for application exploits. Projects will start adding "tell the strongest no-guardrails open weights model to pentest it for 12 hours" to their CICD pipelines. Technical exploits being a dead end, attackers focus their agents on large-scale social engineering. Spamming Discord and Facebook is the new war dialing. Multi-year /goal sessions culminating in gaining a position of trust and sabotaging the CICD pipeline's pentest step since getting anything past it would be intractable. Interesting times indeed.
jimmaswell··on A week of using Codex more than Claude
Max and Ultra are fantastic for the more complex problems where they shine. I use them strategically on certain classes of problems, one or the other depending how parallelizable it is.

I've found it consistently amazing at game dev. It sounds like something that would be difficult for an LLM to verify and iterate on properly, but it almost feels like having a mini-Carmack inside your computer once you try it out. You can throw it at broad, sweeping optimization passes, writing 5 different styles of eyesight sensor frameworks to see what works best in the game as it is, visual scripting integration problems/extensions, etc. with fantastic results.

Also found Ultra great for "get this local LLM working as fast as possible on this odd server setup with old GPUs and AMX support, writing custom kernels/modifications to llama.cpp/sglang/etc as you go while taking notes from relevant research papers and online posts"

jimmaswell··on Stop Making TUIs
> We have a couple of real VT320s.

That's awesome, I have one too and I fell in love with it. I fixed up the old gif320 program to run properly on modern machines a while ago and made a C# port too:

https://github.com/ldyeax/gif320

https://github.com/ldyeax/Gif320Sharp/

How do you deal with modern programs not respecting TERM/LANG capabilities? I found even using man pages on my machine would sometimes result in the terminal getting caught up on some invalid byte sequence emitted from some unicode that snuck its way through or such.

I ended up writing a program I called VT320 Translator, or vt320t for the name of the binary, which in broad strokes runs your target program in a background terminal and sends curated updates to the VT320 while passing through keyboard input. (You can kind of approximate this with just screen and some other things but it doesn't end up being very good IME) This lets me do some neat things like translate unicode characters into DRCS (Dynamically Redefined Character Set) sequences that you can actually see on the VT320, use optimized heuristics for screen updates tuned to the exact hardware, etc.

At the core is a vtparse-based VT320 emulator that strives to be as accurate as possible with an accompanying diff generator from one emulator state to another. If you're interested I could get into more of what I've done here and even get the code onto my github - it's been private for a while because it's being used in an upcoming game demo but I could easily separate out the non-game-related parts that make up vt320t.

jimmaswell··on AI companies destroy physical books – let's scan rare books before it's too late
At least we would all benefit from the books this way, so long as legal nonsense keeps the scans unavailable to the public.
jimmaswell··on Anti-AI fonts are useless and harmful
I agree. As soon as I understood the gist of how modern models do what they do, I found it unreasonable to apply any stricter standards to their "learning" than we do to humans. Humans learn by reading ideas and looking at images, then go on to have ideas and draw images based on that training.

We have definitions of plagiarism and copyright infringement that apply to people based on what they put out into the world, not how they trained themselves to get there[0]. Artists literally trace art and study specific examples in detail to learn, and that's not a bad thing. Humans can also accidentally plagiarize or make things identical to past works - it's easy to think a great guitar riff just came to you when it was really based on a song you heard years ago that stuck in a part of your mind but you don't even consciously remember the influence, for example.

It would certainly be nice for AI output to provide citations if it realizes it's using a significant chunk of an idea from its training that has a clear source (or many), though this is difficult in the same way it would be difficult for me to cite where I learned about the Towers of Hanoi. The AI frequently does web searches for specific resources to get ideas from these days, which are easy for it to cite.

However, it would have a chilling effect on progress as a whole if we all started jealously guarding our ideas so close to our chest that machines couldn't read them and only a select few humans who passed some kind of gate (or even paid us) were allowed to see them. We would be nowhere close to where we are if we had always had such a mindset.

---

[0] We use proven absence of viewing certain material as a legal shield against copyright infringement e.g. clean room engineering, but this isn't strictly necessary, and possibly even discouraged with modern precedent: https://reactos.org/forum/viewtopic.php?t=21740

jimmaswell··on Anti-AI fonts are useless and harmful
I also see no virtue in stopping an AI from reading something to begin with. If anything, I find it anti-social - hide your work from the machine trying to learn from it, taking nothing away from you in exchange for benefiting all of mankind?
jimmaswell··on Don't paste the AI, please
I wish this were simply a fireable offense and we were expected to read. These kind of people who act like reading 3 paragraphs is being asked to swallow glass make me blindingly mad. They should be unemployed.
jimmaswell··on Models Are Getting Dumber on Purpose
Yes, that's the one
jimmaswell··on Models Are Getting Dumber on Purpose
An LLM works better the more disparate world knowledge it has, even if it's not immediately obvious why it would be relevant. The model finds a structure to the problem you give it in a largely language-agnostic way that benefits from training on every language (these things are direct descendants of Google Translate), and even non-programming knowledge - the structure of your task might resemble an ancient Chinese poem that influences the model's response, for example. That structure is considered a form of compression, as some fascinating and illuminating recent 3blue1brown videos get into - a common pattern in Haskell or FORTRAN and a situation described in an ancient Chinese poem may all compress to something quite similar to your task, thus when the model compresses the idea of your task it immediately draws from those ideas.

There are "experts" which do divide parts of the model that are found to activate together for specific tasks, so they can be processed in parallel to join the result at the end, but it's nowhere near the granularity of a SwiftUI expert and a python expert. The difference in those things is so trivial from an abstract point of view that it would make no sense. They would be 99% the same.

Distillations also come into this but I'm highly skeptical you could make one guaranteed to only know programming and only in one programming language (especially with as small a sample set as SwiftUI relative to something like C) without its efficacy being hobbled by tunnel vision. Reminiscent of the SpongeBob episode where he empties his mind of everything except fine dining and breathing, then can't remember his name and goes insane. Beyond the basic concepts of general coding and the trivia of syntax, getting anything done requires a large intersection of disparate world knowledge and the ability to apply it to new situations.

jimmaswell··on Has the hallucination problem in AI been solved?
What do we mean by hallucinate? I'm not counting it making a mistake that it fixes on its own without intervention.

I'm no expert on the inner workings/harnesses/etc beyond a basic understanding of the architecture. Maybe I've just developed a good sense for effective prompts? I could share some recent sessions.

jimmaswell··on Has the hallucination problem in AI been solved?
It matches the reality of my experience using a good frontier like Sol on high to ultra thinking. If it has any way to verify its work whatsoever, it eventually gets to a working and sanely-engineered solution. Blatant hallucinations making it to the final stages have become extremely rare in my use cases, nonexistent if the model has a valid feedback loop. So yes, I will insist someone is likely "holding it wrong" if they still think SOTA AI is spewing out garbage at this stage when I can have it do something like write entire working kernel module fixes for old MacBooks on a whim with no crashes or flaws observed after months of use.
jimmaswell··on The hardest working font in Manhattan (2025)
I can appreciate the aesthetics of a font but I found the author's disadulation of the letters excessive and totally unrelatable. They look nice while clear and functional for their use case to me.
← PreviousPage 2 of 34Next →