Genuinely, for your sake, personal disagreement aside, it would be safer for your career to delete your account and move to Bluesky at this point.
I don’t know why you think screeching about this topic will help you when your boss is very publicly and aggressively taking the opposite side of this.
“also obviously correlated to artistic values”
“…I also will not likely care about something that the developer themselves didn’t care about”
The prominent reply that Casey quoted in the thread that you’re subtweeting was from Tim Soret. Tim has been working on The Last Nite for over 10 years at this point. The opposite of what you’re claiming here.
I assume one reason is that these people have earned a ton of money and want to start spending their time using these incredible tools to make cool things of their own.
Yeah it seems like the programmers who actually care about getting stuff done (and done well) have all embraced these tools: Torvalds, Carmack, yourself, many others.
On the other hand, I noticed the programmers who spend their time coding on twitch streams, who presumably care more about being perceived as smart, tend to be against the tools. I think it's an ego thing.
Perfect analogy. I'm now able to prototype crazy game ideas I've had bouncing around over the years in the course of a weekend or two. Any one of them would have been years worth of weekends in the past!
What I've been surprised by are the developers who seem to have wrapped up their identity in being manual hand saw guys, and are lashing out at those who choose to use power tools. It's very strange.
But we already know what software is made from: binary data.
To apply your analogy properly, you’d be asking for a chocolate bar to not just list the ingredients and country of origin, but which machines, which subcontractors, which brand of forklift, what type of cooling was used in the warehouse, etc.
“never accomplished anything” too strong a claim, they haven’t accomplished things in the world of software engineering but there’s plenty of newcomers who have accomplished heaps of impressive stuff outside of software engineering
“people better than them” better at writing code yeah, but these newcomers will be better than engineers at many things too
I think there’s actually a lot! Christopher Nolan speaks about this (from The Nolan Variations, which I’m currently reading, it’s very good!)
I think you’re right! Or at least some form of this seems right to me.
Some of the best progress I have made on a current project has come from an obscure second hand book I picked up on historical Chinese architecture; the photos in it gave rise to some scenes that I don’t think would have been possible by text descriptions alone.
? I’m aware it’s written in Odin.
I’m saying creating the Odin language is, to me, more impressive than the vfx stuff. You’re free to disagree, but I don’t see why that’s a controversial statement.
Thanks for the detailed response to that. My assumption was that models aren’t far along enough yet to be making valuable contributions to the sorts of problems that you work on, and that confirms it.
And yeah- one problem I have been thinking about is the one you allude to; how we train the next generation of engineers to achieve the baseline competence that allows these tools to be used effectively.
“Do it by hand until you’re good” is definitely one approach but I still wonder if there is a way to safely introduce these tools early on. Framing them as “bullshitting engines” might actually be helpful as a pedagogical technique for that.
People are asking this because it sounds like a huge departure from your recent remark that you would “most definitely never use agents” and it’s an interesting reversal. What changed?
xitter.cf/thegingerbill/status/2…
This comment was asking if I used LLMs/agents for anything, so I wanted to explain why I have been recently using them for.
I mostly use LLMs as a chatbot. I will most definitely never use "agents". Below is what I have been doing for my assembly project specifically:
As an initial way to bounce ideas for syntax so I could refine my concepts faster, even when the suggestions were bad. It was also weirdly passive-aggressive, especially about how my ideas were worse than its own, lol.
To produce a bunch of unit/regression tests for these `asm` templates once I give it the syntax (something LLMs are kind of built for, even if they make mistakes).
As a search engine (which I still verify through other sources).
To generate large tables from pre-existing data, and "clean" them up.
To suggest improvements to error messages, and general grammar and typo fixes in the messages too.
I give it the source code and ask it to find bugs, which it does sometimes (especially typo-related ones, which I'm well known for).
To give me new ideas on what to work on, even if a lot of them are pretty bad.
In general, I'm not using LLMs to program for me, because I very much doubt that they could even architect anything like what I'm trying to build. That takes deep subject knowledge and a real sense of code architecture. These tools are impressive, but damn are they also dumb in a lot of ways. I guess because they have largely been trained on the mean programmer's code, which has essentially no understanding of good architecture.
P.S. The whole idea of "LLM agents" seems insane to me, and even worse than the standard "vibe coding". Please just treat LLMs as a tool, and understand that tool's limitations.
I think that’s the more accurate framing, yes.
And my experience is that the frequency of confident mistakes is going down drastically with each new release. Perhaps that will plateau but I’ve yet to see a strong case for that.
Making a mistake and correcting it is not a strong definition of bullshitting.
They are trained to score high on evals that punish bullshitting.
Yeah I agree with you about design and taste.
I haven’t seen you make the case behind the “fundamental nature” as you describe it. I don’t see how next token prediction is “literally” bullshitting. If you can link me to where you sketch this reasoning out, I will take a look.
Glad to hear you’re doing well then, happy to have misread that aspect.
You’re right too about the current state of that codebase being a mess. I didn’t see it so much as an announcement of a release as a rallying call for people to assist- but I could also be wrong about that too.
I was under the impression that you only used the web chat aspects of these models and that you refused you use the agentic aspects, apart from some annoying database connection code I recall you mentioning. That you consider them bullshitting machines. Perhaps I was mistaken about that also (if not- how do you square the recent model progress in maths with the conception of them as bullshitting machines? I don’t see maths as a realm that tolerates a modicum of bullshitting)
I agree with you that the jangafx stuff is superb, and far more impressive than that is Odin itself. Bill is awesome, he’s done incredible things!
But struggling emotionally, yes clearly he is, evidenced the recent spate of threads like this.
For more context, the ceo of jangafx was posting recently about how good AI has gotten and how it represents a huge threat to their business model now. So it is a very stressful and challenging time for all of them.
xitter.cf/nickseavert/status/209…
>be me
>Realize in 2014 that GPUs are insanely powerful as is
>Have the motto if it ain't realtime it ain't shit
>Have most people believe that realtime fluid simulations are not possible
>assemble a cracked team to build the best VFX tools based on my naivety
>Work as hard as possible to build the world's fastest pyro tools that are ready for production use
>Team codes almost everything from scratch
>Take the hardest paths possible for future stability and control of code base and ownership
>By 2021 we're in a large amount of game studios
>Try hard to be a force for good with a focus on artists experiences in tools
>See machine learning for upscaling volumes is popular and invest in R&D
> Learn that GPUs are the bottle neck and give up after no success and a lot of money spent
>Investments in our tools pay off by 2022, company is crushing it
>AI starts to make serious waves, and seemingly are going after our small niche, VFX/games/film
>Video game industry starts to collapse in 2023/2024 due to embracer fumbling the ball
>Orange cheeto is elected for a second time and our growth immediately plateus due to worldwide aftershocks (we literally have data to prove this)
>Get Crohn's disease and be sick for 2 years
>Congress enacts OBBB and forces amortization of foreign software expenditures over 15 years
>Now have to pay taxes on software dev costs even if we have no profit for a given tax year
>AI is now the global phenomenon but our industry is the epicenter of anti-AI sentiment (rightfully so imo)
>Get hospitalized due to Crohn's and stress
>Economy is crashing and yet stonks are still high
>Crime becomes legal
>We release a new tool called illugen that I think is going to be our next flagship
>Sidefx releases COPS within a few weeks before our launch which is basically the same thing as illugen, with a worse UI, but has houdini level power, so it's awesome as fuck
>Indie sales are taking a big hit
>More massive layoffs back to back
>Permanently instate our layoff assistance program
>Health is 100% again
>Need a hail Mary and illugen seems to be pulling it's weight, record q1 in 2026
>People start trying to One-Shot code our tools, with some moderate success
>We try Claude and see that it is unfortunately quite good
>Become disillusioned in what to do next
>All things appear to be against our survival
>Pay the government over $200k for new taxes we shouldn't need to owe as it's robbery
>Face potential company mortality realities
>Continue building and double down and hire more
>Economic pressure finally imploding and most Americans are affected, yet stonks are still high, crime is legal and rewarded
>Revenues going down, senators ignore my pleas for them to realize that the power to tax implies the power to destroy
>Taxes will be due in 2 quarters
>Not know what the future holds for us
>Somehow still muster up the strength to carry on and desire to keep innovating without resorting to generative AI
>Wonder if I'm stupid
>Unusually depressed and burnt out
Yeah you’re correct, but I believe this is more a case of him grieving at this point xitter.cf/faircall/status/210995…
It’s sad, but understandable, that it’s the younger guys in the community like Bill who are struggling the most with this. They haven’t gotten to enjoy their hard earned time in the spotlight. They had programming idols growing up, and they were on their way to self actualising, to becoming one of them. And then it’s all taken away. It’s deeply unfair.
Bill has spent the past decade building a name for himself as an elite systems programmer, and rightly gained a lot of status in various communities as a result. He’s no doubt sacrificed a lot over the past decade of creating his own programming language- it’s finally scheduled for official release this Christmas, an incredible achievement. And suddenly, seemingly out of nowhere, most people stopped giving a shit anymore about programming languages, or handmade, or the joy of programming.
It’s obvious to anyone who has been paying attention that fixing all the issues raised in the original thread is well within reach of today’s models, let alone what’s coming. And it would feel awful, seeing a project that’s been out for 1 week acrue as many GitHub stars as your 10 year project.
So have some grace, allow him to air some grief. It is hard to see clearly through tears.