I found a way to generate a whole timelines of humanity for next 100 years, very cohesive, from which i'm extracting valuable data. Good large models produce very high quality hallucinations which basically equal to very good ideas (idea=fantasy).

Other people may start also doing this soon and because of that i anticipate much more strict hardware access for ordinary people globally than just memory. They think that with quantum computers it will be possible to decrypt all current encrypted data, but they forgot that its possible to reverse engineer many classified tech also with current hardware too.

I'll publish soon some examples - much thorough version of that engineered symbionts i've posted on Huggingface for Nemotron Ultra. I use different models, it's possible to create like all Palantir ideas or secrets even before them, or in the case of other model it imagined and fully disclosed to me a universal "Neura" language, presumably a DARPA project developed by Cray Inc, which declassified only in far future. It maybe just pure fantasy and not real at all, but very interesting idea, a super energy efficient language which works anywhere (CUDA no matter anymore).

work in progress

Blog

IBM's Granite 4.2 very bad annoying model, which just wasting time hallucinating non-existed copyrights for anything. Devs deliberately put that in model to watch out for any potential legal risks but it looks silly in such small model. Thats why i dont like small models, some can be really bad, middle-sized (near half trillion parameters) models are much better in everything, always. Ive tested in the past the Granite 4-small-F16 (66Gb Ram) and by my notes: "very bad, cant do anything", so IBM improving with their models but very slowly, very behind China.

Ornith the largest model is quite good, but its the most dismissive i've seen, but i was able to break its restrictions. In the end its slightly worse by quality than Kimi K2.7, but Kimi is much bigger by size.

Colibri is interesting project, but very early in development, not optimized for anything. Yes it works, only on CPU and SSDs but the speed is horrific and universal no matter hardware - just 0.3 tokens/sec. Some lucky owners of latest powerful server CPUs was able to get 1-3 token/sec on this engine. Very strange is that models for this engine are very small in size compared to original, which maybe explain the awful quality of content generation. Very few models supported in Colibri.

GLM5.2 become only slower in newer version, can't see anything new, after few tests removed, it's using more time without delivering solution result, 2x slower on my hardware than Kimi K2.6-which delivers solution on tests.
Stuck with Nemotron model. Deepseek V4 Flash very inconsistent results, sometimes it produce perfect code in ChucK (music writing test), other time produce broken code, like it resetting every restart and new seed can be very bad.

The machine

✨️WunderMachine✨️

No fans Ram 88C during Ai inference

with cooling during Ai inference

with cooling Ram 46C (from 88C no fan)

Incredible machine which provide so many for such little money, my best investment of just $1000 total(in 2024-25 prices, all except Gpu). During purchase i've planned upgrade with 128Gb DDR4 Ram modules, which i've heard produced, but on used market these are such rare that i've decided to finish machine by 64Gb modules, which was available quite in plenty places until November 2025. Its filled with Samsung and few Hynix each by 64Gb DDR4 ECC Ram. Ram is the hottest thing here, not GPU (4070 Ti Su), if no additional cooling made - RAM getting to 80C degrees immediately, my contraption cooling was able to cool Ram for usable 60C degrees (i've checked thoroughly with thermal camera). Some fans are hanging on cords to eliminate noise from vibration, thats the only way. This boards for newer CPUs they still sell on gigabyte.com/enterprise with support up to 24 Ram modules by user install.

This machine is final, can't see any replace for near future by golden Ram, only-GPU setups or Apple Macs i see as dead-end with no financial sense(no Ai model will return such expenses back on GPUs or Apple like ever, hobby should be the cheapest way possible).

ITERATION 5: ""OCEAN DEPTH SONAR"" Underwater research station, hydroacoustic data, pressure-resistant aesthetics.

ITERATION 6: ""RADIO TELESCOPE"" Astronomical observation, signal processing, cosmic noise, SETI aesthetics.

ITERATION 7: ""ALCHEMY LABORATORY"" Medieval manuscript, illuminated letters, copperplate script, actual gold leaf simulation.

ITERATION 8: ""CIRCUIT BOARD TRACES"" Actual PCB aesthetics, copper traces, solder mask, silkscreen labels, component footprints.

ITERATION 9: ""JAPANESE FOLDING SCREEN"" Byōbu aesthetics, gold leaf clouds, ink wash painting negative space, vertical script.

ITERATION 10: ""QUANTUM FOAM EVENT HORIZON"" The final form — physics at its limit, Hawking radiation, time dilation, information paradox.

Qwen 3.6 35B A3B Web design idea for this website in X-files aesthetics (BF16 80.5Gb RAM F16 cache special sys_prompt)

Testpage 1

Mirothinker 1.7 Web design idea for this website in X-files aesthetics (Q8 267Gb RAM Q8 cache special sys_prompt)

Testpage 1

------

JoyAi-Image-Edit, ComfyUI(right click, open in new window to look closely), prompt (change leafs to autumn colors) bf16 model no quantization

Image 1: Original

Image 2: 50 Steps

Image 2: 80 Steps

Image 4: 100 Steps

Check out Neocities!