Currently, using the Strata backend with Qwen-Flash-Next. It has successfully made 4 save game editors, 2 plugins, and did some work on La-Mulana 2006 to make the controls and graphics play nice with Windows 11, plus my gamepad. The DLC restoration mod is largely successful, I am doing QA for making the DLC accurate to what was offered in the ad copy.
000000000
Hermes Agent is my model of choice, and I have Strata using 256k context, on a IQ3. I find that Next-Flash is much faster than the smaller 27b, and has more personality. Sometimes it can get a bit loopy. I typically have the AI keep a bible of findings, and start a fresh Hermes session - then tell it to condense the information from the old workspace and session into a more relevant version. Save on context, and makes it more effective since it discards findings that have been superseded.
Hardware wise, a TRX50 AI TOP motherboard, 7965WX CPU, RTX 4090 and x2 RTX 3060, with 128gb of 5600 RDIMM. Down the line, I might try some GPU unlockers when other folk have tested them for bugs. So about a month or two?
I bought my memory and the CPU from Ebay, for prices a bit lower than at Newegg. If I could redo my GPU purchases, I would have gone with a RTX Pro 4500 Blackwell. That has 32gb of VRAM, and occupies a single double-width PCIe slot, and sips energy. The RTX 4090 is only 24gb, and takes up 1 and 1/2 of my slots.
00000000
Qwen-Flash-Next is an experimental model, the official version might come out in a month or so. If it is the same size or smaller, I will switch to that as my main driver. My current AI model feels useful to me, but clearly has raw aspects to it. Vision is a big one, since this AI can fail to see or understand details.
Hopefully my projects make the lives of other people a bit better, alongside my own.
00000
I am thinking about starting up a $5 tip jar at Patreon, with my files free for anyone to download. If I can get enough money saved, I would be able to decrease the amount of time it takes for a project. So far, my projects take 3-7 days to succeed or fail, bottlenecked by the AI speed.
Unfortunately, my projects are pretty niche. Bastion, TEVI, System Shock 2, Ast Libra, Shrift, Sailing Era, The Humankind Odyssey, Alpha Protocol, The Jupiter Incident, and many others are probably too obscure to interest people.
Currently, using the Strata backend with Qwen-Flash-Next. It has successfully made 4 save game editors, 2 plugins, and did some work on La-Mulana 2006 to make the controls and graphics play nice with Windows 11, plus my gamepad. The DLC restoration mod is largely successful, I am doing QA for making the DLC accurate to what was offered in the ad copy.
000000000
Hermes Agent is my model of choice, and I have Strata using 256k context, on a IQ3. I find that Next-Flash is much faster than the smaller 27b, and has more personality. Sometimes it can get a bit loopy. I typically have the AI keep a bible of findings, and start a fresh Hermes session - then tell it to condense the information from the old workspace and session into a more relevant version. Save on context, and makes it more effective since it discards findings that have been superseded.
Hardware wise, a TRX50 AI TOP motherboard, 7965WX CPU, RTX 4090 and x2 RTX 3060, with 128gb of 5600 RDIMM. Down the line, I might try some GPU unlockers when other folk have tested them for bugs. So about a month or two?
https://github.com/JustAResearcher/gpu-unlock-ada
https://github.com/Zagitta/gpu-unlock
https://github.com/impacifist/Windows-GPU-UNLOCK
00000000
I bought my memory and the CPU from Ebay, for prices a bit lower than at Newegg. If I could redo my GPU purchases, I would have gone with a RTX Pro 4500 Blackwell. That has 32gb of VRAM, and occupies a single double-width PCIe slot, and sips energy. The RTX 4090 is only 24gb, and takes up 1 and 1/2 of my slots.
00000000
Qwen-Flash-Next is an experimental model, the official version might come out in a month or so. If it is the same size or smaller, I will switch to that as my main driver. My current AI model feels useful to me, but clearly has raw aspects to it. Vision is a big one, since this AI can fail to see or understand details.
This is the kind of AI success story that never makes the investor decks: Your setup sounds like the future they can’t easily monetize. 💪😎✌️
Unless they somehow make personal setups too expensive for… Oh, wait…
Oh no…
Thank you. :)
Hopefully my projects make the lives of other people a bit better, alongside my own.
00000
I am thinking about starting up a $5 tip jar at Patreon, with my files free for anyone to download. If I can get enough money saved, I would be able to decrease the amount of time it takes for a project. So far, my projects take 3-7 days to succeed or fail, bottlenecked by the AI speed.
Unfortunately, my projects are pretty niche. Bastion, TEVI, System Shock 2, Ast Libra, Shrift, Sailing Era, The Humankind Odyssey, Alpha Protocol, The Jupiter Incident, and many others are probably too obscure to interest people.