What language is the framework in? What "level" is the API? I.e. basic raster operations or complex widget controls?

The framework itself is my personal time waster project, the details of it don't matter that much. Main point is it doesn't exist as far as GPT is concerned. Check out the blog post though, there's a bunch of info and some links.

No thoughts, projections, use cases? To me the fact that the AI can reproduce the likeness of retro software using unfamiliar tools and accessible prompting means it's a matter of time that it'll have a tangible impact on retro computing, one way or another.

For example, though GPT-4 claims that old rendering APIs like S3D and Matrox Simple Interface don't exist when asked to do something with them, you should be able to feed it the API documentation and plausibly get transformations, wrappers, and other content in not many steps. Once the developer experience is in, the end-user experience comes next.

Windblows 95 GPT Edition, eh?

To the extent that Windows was the platform that enabled your experiences (built by developers on top of Windows) in the first place, GPT to me seems a platform that has now and in the future more so the potential to do a similar thing, whatever that thing is.

You'd be better off asking what are the problems that ChatGPT can solve for "retrocomputing" rather than implications since implications can be a lot of things....mostly negative.

vvbee wrote on 2023-04-05, 15:56:

No thoughts, projections, use cases? To me the fact that the AI can reproduce the likeness of retro software using unfamiliar tools and accessible prompting means it's a matter of time that it'll have a tangible impact on retro computing, one way or another.

It would be interesting if you could have it do things to C (or C++) code that are repetitive and tedious but just a bit too complex to try and do with macros or other programmatic transformations of the source code. For example, backporting some C code so that it will compile against Windows 95 headers, or Windows 3.x,etc., or backporting C code to an older C standard to work around missing library functions.

DosFreak wrote on 2023-04-05, 17:53:

You'd be better off asking what are the problems that ChatGPT can solve for "retrocomputing" rather than implications since implications can be a lot of things....mostly negative.

What are among the negative implications?

I'd expect most solutions to come from tooling built around the GPT API other than the limited ChatGPT web chat interface, and I think at that level it's meaningful to look at implications because it should significantly lower the cost (money, energy, whatever) of developer time and so improve the cost/benefit ratio for those activities across the board. This should be especially tangible where you have niches with hobbyists contributing.

jakethompson1 wrote on 2023-04-05, 21:07:

vvbee wrote on 2023-04-05, 15:56:

No thoughts, projections, use cases? To me the fact that the AI can reproduce the likeness of retro software using unfamiliar tools and accessible prompting means it's a matter of time that it'll have a tangible impact on retro computing, one way or another.

It would be interesting if you could have it do things to C (or C++) code that are repetitive and tedious but just a bit too complex to try and do with macros or other programmatic transformations of the source code. For example, backporting some C code so that it will compile against Windows 95 headers, or Windows 3.x,etc., or backporting C code to an older C standard to work around missing library functions.

Transformations should be among the strengths of GPT. In the future, assuming it'll get access to a compiler and can iterate on errors independently, it could be a very fast process.

vvbee wrote on 2023-04-06, 00:35:

jakethompson1 wrote on 2023-04-05, 21:07:

It would be interesting if you could have it do things to C (or C++) code that are repetitive and tedious but just a bit too complex to try and do with macros or other programmatic transformations of the source code. For example, backporting some C code so that it will compile against Windows 95 headers, or Windows 3.x,etc., or backporting C code to an older C standard to work around missing library functions.

Transformations should be among the strengths of GPT. In the future, assuming it'll get access to a compiler and can iterate on errors independently, it could be a very fast process.

Already here 😀

It's called github/Visual Studio Copilot.
https://code.visualstudio.com/blogs/2023/03/3 … /vscode-copilot

Not tried it yet, still on VS2019, however going to update to VS2022 when next appropriate for work and I want to see how it works with some massively complex, multi language projects. As always they perpetually need house keeping and I hear it's very good at that.

If it can write some unit/functionality tests for me that would be great!

If it's on VSCode you can download it for free and give it a go. I haven't tried Copilot myself though.

I wasn't that impressed with GPT-3.5 for coding, but as far as I know they're trialing GPT-4 support now with Copilot X, which should be a noticeable improvement on the Copilot side. No doubt it's the way to program in the future, even that blog post comes across eerily as a human gradually onboarding their replacer bot.

From what I understand you still can't get it to process your entire codebase at once, not enough contextual memory on it to handle it. It's a hardware limitation, but no doubt there'll be software solutions to compress the input or stagger the processing or whatever.

Today OpenAI announced their new model, o3. Based on initial reports it scores about 85% on the somewhat prestigious ARC-AGI benchmark, about matching the average human. GPT-4 scores 5% or so. It's scheduled for release early next year, likely at a high cost initially.

From the ARC-AGI website:

At minimum, solving ARC-AGI would result in a new programming paradigm. It would allow anyone, even those without programming knowledge, to create programs simply by providing a few input-output examples of what they want.

This would dramatically expand who is able to leverage software and automation. Programs could automatically refine themselves when exposed to new data, similar to how humans learn.

If found, a solution to ARC-AGI would be more impactful than the discovery of the Transformer. The solution would open up a new branch of technology.

So the question is again, what are the implications for retro computing?

Why does AI create humans who can't think for themselves?

Artificial Intelligence never beats Natural Stupidity. IMHO. :D

vvbee wrote on 2024-12-20, 19:57:

So the question is again, what are the implications for retro computing?

What do you think the implications are?

feda wrote on 2024-12-20, 23:28:

vvbee wrote on 2024-12-20, 19:57:

So the question is again, what are the implications for retro computing?

What do you think the implications are?

At first there's nothing and then there's the internet where you can discuss, access all information, and buy from anywhere in the world. If you're saying you don't have a particular vision for the future then you might as well expect that AI will transform the hobby the same way the internet did, since why not.

vvbee wrote on 2024-12-20, 19:57:

So the question is again, what are the implications for retro computing?

Writing drivers for old hardware is the only thing that comes to mind.

vvbee wrote on 2024-12-20, 19:57:

Today OpenAI announced their new model, o3. Based on initial reports it scores about 85% on the somewhat prestigious ARC-AGI benchmark, about matching the average human. GPT-4 scores 5% or so. It's scheduled for release early next year, likely at a high cost initially.

Do I read this correctly, they claim that they reached 85% of a human, while current GPT scores about 5% on the same benchmark? If so, I find it very, very difficult to believe this. Especially since none of the previous bombastic announcements proved to be realistic. But hey, hopefully it'll be able to count the Rs in strawberry.

Because if they really reached this level of competence, its implications in retro computing will be the least of our concerns.

It's a third party verified result from a subset of the full benchmark intended to be difficult for AI, and in the subset you might expect a normal human to score in the 50-90% range. To the extent the test correlates with IQ o3's score could be twice the population average. But AI's IQ is alien and won't generalize the same way in any case.