Excellent Project!

[jrobbins-LiveData]

Thank you for publishing this. So cool! I especially appreciate the clear explanations!

[AlexDev404]

Thank you for publishing this. So cool! I especially appreciate the clear explanations!

He didn't make anything. It's entirely vibe-coded (created by an AI). Not a single line of code is owned by him.

[jrobbins-LiveData]

@AlexDev404, are you saying that the explanation is all AI generated with no human in the loop?

[rbergen]

@AlexDev404 It's not my job to defend Dave, nor do I need to, because everything I'm going to say is plainly obvious to anybody who's actually paying attention:

• You're factually wrong. There's one thing that was "entirely vibe-coded", which was the training tracker dashboard. Everything else was created with the assistance of AI, but required direct, targeted, knowledgeable, insightful prompting. Don't take my word for it, just look at the commit/PR history in this repo.
• Criticising someone for using AI on what is to its core an AI-driven project is silly to the point it's funny. Of course AI was used to help set this up, that was part of the whole point.
• Going back to the training tracker dashboard: that too required the correct prompt to be written, the correct generative model to be selected, and the correct context to be provided to the AI generating the code. And that combination of skills is exactly the new set of core requirements for being successful in software engineering, as a software engineer. I hope you get used to that idea quickly, for your own sake.

[AlexDev404]

@AlexDev404 It's not my job to defend Dave, nor do I need to, because everything I'm going to say is plainly obvious to anybody who's actually paying attention:

• You're factually wrong. There's one thing that was "entirely vibe-coded", which was the training tracker dashboard. Everything else was created with the assistance of AI, but required direct, targeted, knowledgeable, insightful prompting. Don't take my word for it, just look at the commit/PR history in this repo.
• Criticising [...truncated]

I don't think you've ever really used agentic AI honestly from reading everything you wrote.

GitHub Copilot Agent is completely self-sufficient, requiring little to no input once it's begun to work.

Also, "assistance of AI" versus "vibe coding" is essentially the same thing. I've watched multiple of Plummer's videos, and every one of them is him dumping a prompt into the IDE and watching it work. The notepad app he "created", for instance, is a good example.

[AlexDev404]

@AlexDev404, are you saying that the explanation is all AI generated with no human in the loop?

That's right. My speculation is that this is at least 80% - 90% AI especially after looking at the pull request and commit history. Everything in the commits are created in bulk, and especially not something that a human could do in such a short period of time.

The pull request history as well has several references to GitHub Copilot.

And the contributors also include GitHub Copilot.

[AlexDev404]

@AlexDev404 It's not my job to defend Dave, nor do I need to, because everything I'm going to say is plainly obvious to anybody who's actually paying attention:

• You're factually wrong. There's one thing that was "entirely vibe-coded", which was the training tracker dashboard. Everything else was created with the assistance of AI, but required direct, targeted, knowledgeable, insightful prompting. Don't take my word for it, just look at the commit/PR history in this repo.
• Criticising [... truncated]

Not to press on this, but what gets to me off the most is how this AI thing is compounded even more as recently most of Plummer's vibe coded applications have been published on so many news sites, continuously crediting him as the sole-author -- without any acknowledgement that it was mostly AI.

It is neither fair nor ethical to take credit for something that you did not completely do yourself.

[rbergen]

I don't think you've ever really used agentic AI honestly from reading everything you wrote.

I have, and do so every day.

GitHub Copilot Agent is completely self-sufficient, requiring little to no input once it's begun to work.

That really depends on a number of factors, some of which I've mentioned in my previous comment, others relate to what it's working on. The skill is using the tool(s) in a way that's effective, reviewing what comes out of it, then iterating until what you have is right. And you can do that at multiple levels, as I'll go into next.

Also, "assistance of AI" versus "vibe coding" is essentially the same thing. I've watched multiple of Plummer's videos, and every one of them is him dumping a prompt into the IDE and watching it work. The notepad app he "created", for instance, is a good example.

We're entering the realm of semantics, but "entirely vibe coding" in my book is letting AI deliver the end result and only judge that end result on its functionality (i.e., not even consider the code). "Assistance of AI" is an engagement where one uses input from AI (either in terms of approach, code, or both) but still accepts the outcome of that engagement based on the composition of the technical (code-level) end result.

And again, I'm convinced the definition of "create" needs updating. Your use of quotes indicates you don't consider something created unless it's hand-crafted. That would mean no piece of fabric has been created since the introduction of punched cards in that industry around 1800. I'd argue that to be a false statement, too.

[rbergen]

Not to press on this, but what gets to me off the most is how this AI thing is compounded even more as recently most of Plummer's vibe coded applications have been published on so many news sites, continuously crediting him as the sole-author -- without any acknowledgement that it was mostly AI.

I think Dave has always been very transparent about the assistance he's been getting from the tools he's been using, AI or otherwise. Which is why you can even make this comment.

It is neither fair nor ethical to take credit for something that you did not completely do yourself.

Again, I've never seen Dave take credit for anything he hasn't actually done himself, and again, using AI in a way it contributes is a skill. If others award him credit, that's their decision, not his.
At another level, the "completely do yourself" argument brings us back to a discussion on the validity of any tools to aid in development. That would mean a developer couldn't take credit for work done if they use an IDE with code completion and in-line error checking, or even a higher-level language that uses a compiler. There was a time that the argument was "real programmers use machine code", and we've moved past that as well.

[AlexDev404]

I have, and do so every day.

This is what vibe-coding is.

That really depends on a number of factors, some of which I've mentioned in my previous comment, others relate to what it's working on. The skill is using the tool(s) in a way that's effective, reviewing what comes out of it, then iterating until what you have is right. And you can do that at multiple levels, as I'll go into next.

No. Also, that is literally the definition of vibe-coding. And I think you don't really know the difference between Copilot Agent and Copilot for Visual Studio.

We're entering the realm of semantics, but "entirely vibe coding" in my book is letting AI deliver the end result and only judge that end result on its functionality (i.e., not even consider the code). "Assistance of AI" is an engagement where one uses input from AI (either in terms of approach, code, or both) but still accepts the outcome of that engagement based on the composition of the technical (code-level) end result.

Your definition is entirely subjective, and it just seems to me that you're backing the idea of vibe-coding without any sort of introspection. Vibe-coding is accepting code from AI whether or not you consider it "good enough" without really writing anything yourself. It's AI generated. It's that simple.

Obviously, AI can produce a good result. But it doesn't matter whether you accept it or not, it's still not your work.

And again, I'm convinced the definition of "create" needs updating. Your use of quotes indicates you don't consider something created unless it's hand-crafted. That would mean no piece of fabric has been created since the introduction of punched cards in that industry around 1800. I'd argue that to be a false statement, too.

This is nonsense. Real software has always been handmade and still is. Textiles are still designed by real people and still are. Machines print the textiles because obviously humans cannot.

On the contrary, you are just using AI because you're too lazy to do it yourself.

I hope you realize that the real people who are successful, are the people who are actually running these models just so you can use them, not you. And these people will always be the ones making more money than you.

[AlexDev404]

I think Dave has always been very transparent about the assistance he's been getting from the tools he's been using, AI or otherwise. Which is why you can even make this comment.

No, he has not. His YouTube video title for notepad, for example, is very much along the lines of "Microsoft 'Improved' notepad I unimproved it" which is very much not stating that he used AI until later in the video. Completely clickbait to viewers unaware.

Again, I've never seen Dave take credit for anything he hasn't actually done himself, and again, using AI in a way it contributes is a skill. If others award him credit, that's their decision, not his.

Prompting an AI is not a skill. Any average Joe can prompt an AI - regardless of whether or not they can code.

At another level, the "completely do yourself" argument brings us back to a discussion on the validity of any tools to aid in development. That would mean a developer couldn't take credit for work done if they use an IDE with code completion and in-line error checking, or even a higher-level language that uses a compiler. There was a time that the argument was "real programmers use machine code", and we've moved past that as well.

AI is not code completion. Code completion is rule-based not run on a neural network. Also the "real programmers use machine code" does not make any sense. It is not a thing anymore because you're now forced to use a higher level language because of all the abstraction layers that have been imposed on developers by the operating system which makes coding in machine code just completely a waste of time and impractical.

You can't code a website in machine code just as you can't render a graphic onto the screen using "POKE" anymore. It's just not possible. Your example really does not apply.

[rbergen]

Obviously, AI can produce a good result. But it doesn't matter whether you accept it or not, it's still not your work.

I think we're just going to have to agree to disagree on this. We have different perspectives on "your work" that we will just not convince each other of.

On the contrary, you are just using AI because you're too lazy to do it yourself.

A statement that's been made about developers for a very, very long time is "a good developer is a lazy developer". There are connotations and side notes to be made about it, but I think it is true in a meaningful sense. As a software engineering discipline in the broad sense, we've always actively pursued ways to make our delivery processes more efficient, and decrease the dependence on human actions. Automation has always been a key part of that. AI is automation, but at a different level than what we've seen before.

This is nonsense. Real software has always been handmade and still is. Textiles we're still designed by real people and still are. Machines print the textiles because obviously humans cannot.

But humans did "print" the textiles by hand, by hand, one thread at a time, weaving in the patterns the punch card machines then did automatically. And there are artisan textile makers that still do it the old-fashioned way.
Anyway, the key word in your own argument is "designed". If you design a feature in functional terms - i.e. human language - and let the machine - AI, in this case - generate the code, did you still not create it?

I hope realize that the real people who are successful, are the people who are actually running these models just so you can use them, not you. And these people will always be the ones making more money than you.

I admit I'm reading between the lines, but I think this may well be a core aspect of your frustration. And yes, you're right. Just like Amazon is making a lot more money than I do because they've created a "model" for providing infrastructural and higher-level services that's called AWS. That's how economies at scale work.
Should one make themselves completely dependent on AI-generated code? I don't think so, at least not yet. Is it sensible to at least seriously explore using this "new" tooling to improve efficiency, and quality, of engineering output? I do think so, as we've done time and time again, in the past.

[AlexDev404]

I think we're just going to have to agree to disagree on this. We have different perspectives on "your work" that we will just not convince each other of.

So essentially what you're saying is that if you hire someone to write some code for you and you ask them to make changes, it's your code? Because that's what I'm getting from you.

A statement that's been made about developers for a very, very long time is "a good developer is a lazy developer". There are connotations and side notes to be made about it, but I think it is true in a meaningful sense. As a software engineering discipline in the broad sense, we've always actively pursued ways to make our delivery processes more efficient, and decrease the dependence on human actions. Automation has always been a key part of that. AI is automation, but at a different level than what we've seen before.

That may have been the case back then, but it isn't anymore. A good developer writes secure code, a good developer ensures his or her program is user-friendly, and a good developer treats their code with respect with his or her successor or collaborators in mind.

When you enlist AI to write code for you, you are waiving all those promises, and you are no longer a programmer - you're just someone with an idea.

But humans did "print" the textiles by hand, by hand, one thread at a time, weaving in the patterns the punch card machines then did automatically. And there are artisan textile makers that still do it the old-fashioned way.

I'm very sure you understand what my point was, in the same sense you did with my second example about both those things being impractical or just impossible to do these days since people have to keep up with demand or with systems being completely abstracted to the point of no return. On the other hand, AI is just doing things you could've just done yourself.

Btw, the less you exercise your knowledge, the more you forget how to do certain things. So it's at your own detriment.

Anyway, the key word in your own argument is "designed". If you design a feature in functional terms - i.e. human language - and let the machine - AI, in this case - generate the code, did you still not create it?

No. Because the AI is creating it based off of the data it has learned. It's like having someone write an email for you based on what you want. It's no longer in your tone, your voice and your thoughts. It's in theirs. Additionally, again, these systems you are describing are rule-based, not self-attentive, and most importantly -- have way more respect for the planet compared to all the processing power, water, and energy using AI takes up currently.

I admit I'm reading between the lines, but I think this may well be a core aspect of your frustration. And yes, you're right. Just like Amazon is making a lot more money than I do because they've created a "model" for providing infrastructural and higher-level services that's called AWS. That's how economies at scale work.
Should one make themselves completely dependent on AI-generated code? I don't think so, at least not yet. Is it sensible to at least seriously explore using this "new" tooling to improve efficiency, and quality, of engineering output? I do think so, as we've done time and time again, in the past.

Nobody should ever make themselves fully dependent on AI generated code. Because then as I mentioned earlier, dependence creates forgetfulness.

What you're describing is a big ask for all of humanity to lean into becoming fully reliant on AI while completely forgetting how to write code in the first place. What you're doing is effectively begging for an apocalypse.

[rbergen]

No, he has not. His YouTube video title for notepad, for example, is very much along the lines of "Microsoft 'Improved' notepad I unimproved it" which is very much not stating that he used AI until later in the video. Completely clickbait to viewers unaware.

This relates directly to the thing I think we need to agree to disagree on. He did unimprove it, using AI.

Prompting an AI is not a skill. Any average Joe can prompt an AI - regardless of whether or not they can code.

Well, I think we have another key point of disagreement. Any average Joe can prompt an AI, but creating a prompt with all the content, guardrails, context, etc. to make it produce what is needed is a skill in my experience.

AI is not code completion. Code completion is rule-based not run on a neural network.

So the issue for you is that AI is not "rule-based"? That feels like a rather arbitrary objection to the technology to me. If only for one reason: ask two human developers to code something, and you get two different implementations. That has always been the case, and you wouldn't reject the output of either developer because of that fact.

Also the "real programmers use machine code" does not make any sense. It is not a thing anymore because you're now forced to use a higher level language because of all the abstraction layers that have been imposed on developers by the operating system which makes coding in machine code just completely a waste of time and impractical.

Absolutely. My argument is that you can expand that argument to a point where the abstraction level of developing reaches the level of human language. And this is not new either; to a meaningful extent low/no-coding platforms had the same ambition. What we discovered is that they still require knowing the correct language to use, and apply system thinking, and understand the core principles that make a sound implementation. I don't believe that has changed with the arrival of AI, and that means using AI still requires engineering skills.

You can't code a website in machine code just as you can't render a graphic onto the screen using "POKE" anymore. It's just not possible. Your example really does not apply.

You can, which is what happens through the layers of abstraction in the end - all code you write translates to machine code by the time it's executed. But as you said, writing it by hand is impractical, which is why we've found more efficient ways to generate it.

[AlexDev404]

This relates directly to the thing I think we need to agree to disagree on. He did unimprove it, using AI.

I honestly don't know if you're trying to make a joke or not because that's just silly now.

Well, I think we have another key point of disagreement. Any average Joe can prompt an AI, but creating a prompt with all the content, guardrails, context, etc. to make it produce what is needed is a skill in my experience.

You don't need any of that anymore these days. And that's my point.

So the issue for you is that AI is not "rule-based"? That feels like a rather arbitrary objection to the technology to me. If only for one reason: ask two human developers to code something, and you get two different implementations. That has always been the case, and you wouldn't reject the output of either developer because of that fact.

Exactly. And when you ask two developers to code, it's their code and not yours right? Or do you automatically own their code?

Absolutely. My argument is that you can expand that argument to a point where the abstraction level of developing reaches the level of human language. And this is not new either; to a meaningful extent low/no-coding platforms had the same ambition. What we discovered is that they still require knowing the correct language to use, and apply system thinking, and understand the core principles that make a sound implementation. I don't believe that has changed with the arrival of AI, and that means using AI still requires engineering skills.

No it does not anymore. And it will probably require less and less of it in the coming months.

You can, which is what happens through the layers of abstraction in the end - all code you write translates to machine code by the time it's executed. But as you said, writing it by hand is impractical, which is why we've found more efficient ways to generate it.

I'm sure you knew what I meant.

[rbergen]

So essentially what you're saying is that if you hire someone to write some code for you and you ask them to make changes, it's your code? Because that's what I'm getting from you.

It's not "my" code, but it is our mutual responsibility. And in my role I am accountable for a lot of code I haven't written myself, and I take that accountability very seriously. Furthermore, in a professional context, actual code ownership in the legal sense is a contractual aspect in the end, as I'm sure you know.

That may have been the case back then, but it isn't anymore. A good developer writes secure code, a good developer ensures his or her program is user-friendly, and a good developer treats their code with respect with his or her successor or collaborators in mind.

And a good developer still makes sure the code they are responsible for meets all those requirements, to the extent they apply. Using AI to generate output in any way does not exempt anyone from any responsibility concerning their output. If I made the impression I think one should blindly use whatever comes out of any form of automation then I'd like to correct that now, as that has never been my conviction.

When you enlist AI to write code for you, you are waiving all those promises, and you are no longer a programmer - you're just someone with an idea.

As just stated, I don't think it's that black and white. And actually, if anybody feels they can't use AI in a way that satisfies the requirements in terms of responsibility and accountability, they indeed shouldn't use it. My statements that using AI correctly is a skill very much boil down to exactly this aspect, amongst others.

I'm very sure you understand what my point was, in the same sense you did with my second example about both those things being impractical or just impossible to do these days since people have to keep up with demand or with systems being completely abstracted to the point of no return. On the other hand, AI is just doing things you could've just done yourself.

I argue again that that distinction is arbitrary. Until weaving was automated it was done by hand, and made sense up to that point. Introducing automation made new developments possible that cannot reasonably be done by hand. And yes, the same applies to modern software development. All of those methods of automation originally did something one could do by hand, and then enabled things that couldn't. I think that's where this new form of automation is heading as well. And in certain, largely non-LLM, fields of AI that has already been happening for quite a while.

Btw, the less you exercise your knowledge, the more you forget how to do certain things. So it's at your own detriment.

I never said I think engineering knowledge is unnecessary - in fact, to the contrary -, but the abstraction level of knowledge that is required shifts, over time. I've been developing software for about 40 years now, and I couldn't tell you the x64 CPU instruction set if my life depended on it.

No. Because the AI is creating it based off of the data it has learned. It's like having someone write an email for you based on what you want.

Which is what many senior managers are indeed asking certain members of their staff to do, in their name. That pattern has always been there, but you object to it now being implemented by a machine.

It's no longer in your tone, your voice and your thoughts. It's in theirs.

This is not necessarily true. With sufficient input and context, modern models can write in someone tone of voice. Again, like humans do.

Additionally, again, these systems you are describing are rule-based, not self-attentive, and most importantly -- have way more respect for the planet compared to all the processing power, water, and energy using AI takes up currently.

The latter is what I expect to be a second core objection you have against AI. And that is a fair point, but a different conversation. You started attacking someone for, very openly, using AI for something others give them credit for. If your core objections are against the technology in itself, then I think that's the conversation you should actually aim for.

Nobody should ever make themselves fully dependent on AI generated code. Because then as I mentioned earlier, dependence creates forgetfulness.

...amongst other reasons. I'm happy to see there is something we can agree to agree on. :)

What you're describing is a big ask for all of humanity to lean into becoming fully reliant on AI while completely forgetting how to write code in the first place. What you're doing is effectively begging for an apocalypse.

I'm not arguing to become fully reliant on AI, nor am I "begging for an apocalypse". In fact, very much not. AI is, or can be, a powerful tool, and using it comes with great responsibilities, as I have indicated multiple times in this conversation. It is up to us, the experts in our field, to figure out what responsible use looks like, on the shorter and the longer term. I am convinced that rejecting the technology outright is not the answer to that question.

[LWFlouisa]

I'm curious to take to look just to see how Dave did it. Grandpa was a really big fan of Pinball.

[rbergen]

@LWFlouisa There are actually (more or less) 2 angles to this. The conversation you added your comment to was about how Dave used AI to create part of the solution. I.e., it concerns AI being used to support - or, in case of the dashboard, basically take care of - the implementation of what's in this repo. (After that conversation took place, I read a Medium post on that topic that really resonated with me - unfortunately I can't share it because it's a paid member writing.)

The other angle, which I personally think is more interesting, is how Dave used AI to master the game by setting up and training his own model. That effort was also supported by AI, but still took Dave a year to complete, and not in 15-minute installments either.

[LWFlouisa]

Ah like figuring out a special method for training on your own specific game type? So in theory each implementation could be completely different.

[rbergen]

That's basically exactly where this is right now. After addressing the Tempest challenge, Dave went on to Robotron, which required a fresh model with a fresh set of rules/boundaries.

In a kind-of tongue in cheek exchange I did raise the prospect of generalizing the "learn to play a game by its rules" premise to something that could be applied to all (MAME-emulated) games, but we may be a few individual games away from that becoming an actual project. :)

[FrankRizzo890]

And FWIW, I previously wrote non-ML based bots to play individual games similar to what Rutger is describing. (Written in 100% LUA, and are visible on my GitHub repo).

…

------ Original Message ------ From "Rutger van Bergen" ***@***.***> To "davepl/ArcadeAI" ***@***.***> Cc "Subscribed" ***@***.***> Date 3/20/2026 8:50:30 AM Subject Re: [davepl/ArcadeAI] Excellent Project! (Issue #10)

rbergen left a comment (PlummersSoftwareLLC/ArcadeAI#10) <#10 (comment)> That's basically exactly where this is right now. After addressing the Tempest challenge, Dave went on to Robotron, which required a fresh model with a fresh set of rules/boundaries. In a kind-of tongue in cheek exchange I did raise the prospect of generalizing the "learn to play a game by its rules" premise to something that could be applied to all (MAME-emulated) games, but we may be a few individual games away from that becoming an actual project. :) — Reply to this email directly, view it on GitHub <#10>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AC3E6XCJ2XHXVEL4RFHYCBL4RVSENAVCNFSM6AAAAACV3LTZNKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DAOJZGE4TEOJZG4>. You are receiving this because you are subscribed to this thread.

[rbergen]

You're wrong, certainly but not only on the effort part. Proving that would require me sharing conversations that took place between friends, and I won't do that.

I appreciate the "respectfully", though.

[AlexDev404]

You're wrong, certainly but not only on the effort part. Proving that would require me sharing conversations that took place between friends, and I won't do that.

I appreciate the "respectfully", though.

I also wouldn't want you to disclose personal information.
I'm sorry Bergen. You are right. I was hoping you did not see my comment before I got to delete it.

All this while I didn't actually believe you were being genuine, and even moreso his friend.

And all this while I've misguided myself into thinking that Dave didn't really do anything, when in fact he did so much.

But I've realized I've made a significant misjudgment and miscalculation in the thing I did and said.

I am so sorry. It most likely is meaningless at this point now, since you, and probably anyone else reading has probably already received my previous reply in their email, but I still want to apologize out of remorse and respect for you and Mister Plummer.

I recently have watched his entire video from start to finish, watching what he does and did, and with this I have deep respect for the work he has done in this codebase.

He is a brilliant man, and someone who I now think I've ruined my chances at ever meeting in life.

I feel even worse now that he even explicitly brought the topic up in his video, and took his time to explain his entire procedure from beginning to end, but instead of watching his video before I posted, I instead resorted to leaving a negative comment on his codebase.

I'm really sorry for what I said, and so I understand and respect if now you and/or Dave now have a negative perception of who I am now.

[rbergen]

First off, I'm choosing to take that at face value. Which may be naive, but so be it.

Second, "Mr. Plummer" is Dave and I'm Rutger, please.

I have met Dave a few times. I don't think I'm qualified to assess if he is brilliant or not, but I can say that the tests say I'm doing pretty well on the intelligence front, and Dave outclasses me without making an effort. He combines that with a focus and persistence I don't even have a chance of matching.
It's a powerful combination. (Dave would say there are things I can in fact add to that with my wiring, but I won't make those statements on his behalf.)

Going back to this project: it has been a journey of exploration, again at a number of levels. And you're right that that shows in commits that took place.

Stepping back a bit, these are uncertain times in the software industry. To an extent, I feel like we're all on Columbus' ship, sailing West to end up East. That journey didn't go exactly as planned, but it did lead to a pretty interesting set of discoveries. I think the kind of exploration this repo represents is valuable, now.

[davepl]

This is a very odd thread. There may be so many layers of sarcasm and nuance going on here that it’s lost on me, so I’ll just take it at face value. While it may well be entirely true that I am overconfident and have no idea what I’m doing, I will submit that you can enlist in ArcadeAI, run the scripts in the Tempest folder, and watch the ML play Tempest at a level of complete mastery. On your own machine without me there. So I’d temper the rest of the comments with that reality. I’m just starting with Robotron, and it currently gets me to about level 13. Given that the expert bot can reach 9, that’s not much better than a programmatic bot yet. So I have no material progress in Robotron with which to defend myself yet, but I remain optimistic. At the end of the day, I wrote a two-pass assembler in a machine language monitor when your mom was in high school, so proving myself is pretty low on my priority stack. Forgive me if I do not do so, as I am fully happy being considered a pretender in the ML space as long as I’m having fun and learning. If amateurish attempts combined with irrational perseverance that result in success offend you, I am guaranteed to annoy you. It’s what I do. Watch. Dave

…

On Mar 20, 2026, at 11:50 PM, Immanuel Daviel A. Garcia ***@***.***> wrote: AlexDev404 left a comment (PlummersSoftwareLLC/ArcadeAI#10) <#10 (comment)> Respectfully, this thread is entirely BS, quite honestly. As I've mentioned several times before, Plummer knows nothing about what he's doing in this project. Training an AI is no easy task, and not something you can just vibe-code your way into. There is no sort of "assistance" going on in this repository - and it's quite obvious from the comments and pull requests that David Plummer is overconfident and has no idea about what he is doing in this codebase. His commits are in no way atomic, and are filled with random junk and text files which quite obviously states that he is NOT programming anything himself. And most obviously, he is just pasting junk into an AI and expecting it to program it for him as seen by these two pull requests: PULL REQUEST - #4 <https://github.com/davepl/ArcadeAI/pull/4>image.png (view on web) <https://github.com/user-attachments/assets/f6e1a93e-89f2-4d91-be76-fd98887fcfdc> image.png (view on web) <https://github.com/user-attachments/assets/f58bf1ba-42ec-4948-b2d7-1c0d909394df> PULL REQUEST - #3 <https://github.com/davepl/ArcadeAI/pull/3>image.png (view on web) <https://github.com/user-attachments/assets/81ded053-1dbd-4f59-9acf-73356b3e6925> David Plummer is not an ML engineer and has no business fooling around with AI with things he has never once studied in his life himself. He is wasting his time and as such, everyone else's. You don't let children play with fire, just as you don't let people wield such tools, they have no business wielding. He is getting nowhere right now, and as such, will get nowhere in the future. He pulled down the vibecoded "dashboard" for one good reason, and one good reason only which we all know the answer to - he has no clue as to what he's doing - so probably couldn't even fix the dashboard if he so sadly tried. The other angle, which I personally think is more interesting, is how Dave used AI to master the game by setting up and training his own model. That effort was also supported by AI, but still took Dave a year to complete, and not in 15-minute installments either. There is no "mastery" going on here. The AI can't even play the game properly. You are being fooled. He hasn't even uploaded a single model file to HuggingFace since Adam. This is complete bullshit. And he's bullshitting his way into something which he could do easily if he had just taken the time to learn the subject at hand. If he even knew a single clue (instead of being clueless as to what he was doing) he would know that the easiest and most probable way to fix his ML model was to lower the learning rate and raise the gradient instead of throwing his training data at AI in hopes of "something being wrong in the code". As I've said earlier in this thread - You don't learn by having AI do the work for you - You learn by putting in effort. — Reply to this email directly, view it on GitHub <#10>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA4HCF2PRVPZTIRYUIEYHFL4RY3URAVCNFSM6AAAAACV3LTZNKVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHM2DCMBSGYYDSNJUGA>. You are receiving this because you are subscribed to this thread.

[jrobbins-LiveData]

I almost feel guilty posting my original compliment and gratitude! I meant it, because Dave's project is really interesting and enlightening to me. I spent more quarters than I'd care to count back in college on both Tempest and Robotron (the dread "brain wave"!) and I look forward to your ongoing work on it. It is a great way to see a variety of AI and ML techniques applied to a nicely bounded (and FUN!) problem domain. Thank you, @davepl!

I have no idea why expressing my gratitude for this excellent work provoked so much animus.

Maybe the folks so inclined could, instead of spewing unsubstantiated invective, put up their own repos and videos showing their "more worthy" projects, and let us be the judge?

[rbergen]

@jrobbins-LiveData Thank you for circling back, here. I feel I should apologize, because it can't be denied that I contributed to this "hijacking" of a very straightforward compliment and expression of gratitude, as you say. My comments weren't made with bad intent, and I guess they just show that I have buttons that I respond to when they are pushed.

[AlexDev404]

While it may well be entirely true that I am overconfident and have no idea what I’m doing, I will submit that you can enlist in ArcadeAI, run the scripts in the Tempest folder, and watch the ML play Tempest at a level of complete mastery. On your own machine without me there. So I’d temper the rest of the comments with that reality.

I don't think anyone here really has the machine power to recreate the model from scratch using these scripts, let alone run multiple instances of the emulator simultaneously. Additionally, environmental factors create different results, since the training results aren't deterministic. That's why it would be great if you could upload your results to HuggingFace. You actually don't have to - but just saying.

I’m just starting with Robotron, and it currently gets me to about level 13. Given that the expert bot can reach 9, that’s not much better than a programmatic bot yet. So I have no material progress in Robotron with which to defend myself yet, but I remain optimistic.

I believe that you can do it, and sorry again for my criticism.

At the end of the day, I wrote a two-pass assembler in a machine language monitor when your mom was in high school, so proving myself is pretty low on my priority stack. Forgive me if I do not do so, as I am fully happy being considered a pretender in the ML space as long as I’m having fun and learning.

I understand what you mean, but I don't think arrogance is the answer here. Because at the end of the day it isn't when you did it, it's how you did it. Competence scales relatively, not absolutely.

If amateurish attempts combined with irrational perseverance that result in success offend you, I am guaranteed to annoy you. It’s what I do. Watch.

Looking forward to it :)

[AlexDev404]

First off, I'm choosing to take that at face value. Which may be naive, but so be it.

Yeah I was actually being honest throughout, especially in my last message.

Second, "Mr. Plummer" is Dave and I'm Rutger, please.

My bad, sorry.

I have met Dave a few times. I don't think I'm qualified to assess if he is brilliant or not, but I can say that the tests say I'm doing pretty well on the intelligence front, and Dave outclasses me without making an effort. He combines that with a focus and persistence I don't even have a chance of matching.
It's a powerful combination. (Dave would say there are things I can in fact add to that with my wiring, but I won't make those statements on his behalf.)

I could imagine. I admire those qualities. Thanks for all your patience throughout here.

Stepping back a bit, these are uncertain times in the software industry. To an extent, I feel like we're all on Columbus' ship, sailing West to end up East. That journey didn't go exactly as planned, but it did lead to a pretty interesting set of discoveries. I think the kind of exploration this repo represents is valuable, now.

That's true. It makes it hard to differentiate between what's real and what's not sometimes. But hopefully in the future we can know what's for certain.