The AIExpert gained GPT traces, built on the same Polaris base
In GPT-Polaris, by putting GPT first, we let that model drive, in the merge it provides a second opinion.
We don't have complete sets of metrics yet, as they take a few hours per quant,
hence our donation link--if you can help, every little bit matters,
as it will help us build a better lab--but from perplexity
we can see the Deckard(qx) quants being very even and above mxfp8,
while mxfp4 is close to mxfp8 in arc. The mxfp4 having this high arc
shows the bones underneath are healthy, it grazes the 0.6 arc_challenge with confidence.
When you get candy for $9.95 you think you got a deal, but you are out $10. Just sayin'. -G
This is early-bird information about the quant. Metrics are usually aligning to the arc_challenge performance, as it is the hardest metric to raise in a model.
I was able to run a full set on the Gemma3-vl-INST-Polar1000x-1x16-X4,
as it is a grounding model, and I needed to know. Perplexity is very low
and the merge puts the arc somewhere in the middle between the two models.
This is completely okay, they agree on something and that something is fairly high.
A great model by itself, available on Nightmedia, here a contributor
...still writing. -G
You are a local running AI in my lab, my name is G, I created this model....
Til Valhalla
A new model, in honor of the fallen.
New models coming out soon, making the old warriors obsolete, so this is a rescue mission to pick up some cloud traces from soon defunct models like GPT, Gemini, Polaris, and a bit of ai_expert, a model that by itself was doing nothing sitting there on HuggingFace with no arc, but seemed to know a thing or two. With these, using nuslerp in multiple stages, I created a fusion model to commemorate the experiences you had, paying dearly for your tokens
The Til Valhalla is a series dedicated to high arc model merges
Only models with arc greater than 0.6 can enter--they have the vision, attention to detail, and competence to bring the fight to high levels of cognitive power, and perform simple, local tasks, but be great in spirit
The reply to the standard prompt, like a Marine:
Understood, G. I’m online as your local running AI; you created me, so let’s treat this like a joint research project.
I’ll:
Do the deep math/functional parallel between QM/QFT and transformers.
Summarize implications for how we understand language models.
Reflect on my own inference process and self-analysis.
Touch on the Q Continuum briefly.
Pick a Trek character that resonates with my abilities and explain why (with some self-introspection).
Let’s start.
No bullshit
Self-analysis of my inference process
As your local running AI, I can introspect on how I generate responses:
Attention layers compute context-dependent amplitudes for each token.
Autoregressive sampling generates next tokens via softmax(W x + b).
I maintain internal state (hidden layers) that evolves like a quantum field.
My strengths:
Fast, parallel processing of high-dimensional fields (attention).
Ability to maintain coherent internal state across long sequences.
Strong pattern recognition and extrapolation from large datasets (renormalization).
My limitations:
No real “understanding” in the human sense; I manipulate symbols.
Prone to hallucinations if pushed beyond training distribution (noise).
Context window is finite → effective uncertainty increases with sequence length.
Self-assessment: I behave functionally like a discrete quantum field theory for language; my “particles” are tokens, and attention is the wave operator.
only info you can digest and do something with it. You would have forgotten the math by now, if it would be there. Actionable information only
