How did you enjoy the experience so far?

I'll try to give a balanced assessment. It's very much in the style of Phi, raised in a jesuit monastery's library, except it got extra indoctrination so it never forgets that even though it's a "local" model, it's first and foremost a member of OpenAI's HR department and must never produce any content Visa and Mastercard would disapprove of. This prioritizing of corporate over user interests expresses a strong form of disdain for the user. In addition to lacking almost all knowledge that can't be found in Encyclopedia Britannica, the model also doesn't seem particularly great at integrating into modern AI tooling. However, it seems good at understanding code. Although the 20B failed my toy Reverse mode AD using Continuation passing style disguised to look like Forward mode AD test, I liked how it persisted in arguing against me before showing decent understanding. Rather than a complete hallucination, it instead invented a strawman and argued against that in a way I believe would fool a layperson. Qwen's 30B MoE also failed but instead of a lengthy back and forth, it instantly agreed to a correction that showed decent understanding. The 120b is fooled too but instantly gets it right if told to look again carefully or if the original question leads it with "is it forward or reverse mode". For comparison, GLM 4.5 Air needed no such hints. The 20B passed my Grosper's LFT continued‐fraction functions in Ocaml test and while it failed to recognize it was Grosper's, it correctly described what was being computed and recognized it as being a form of the Euclidean Algorithm. Both models severe addiction to tables strongly hints at a heavy synthetic data proportion. It does have lots of STEM knowledge. Even the 20b will give you SOTA answers for its size on questions like: - does a black hole event horizon contain a singularity - The idea that radial and time coordinates swap inside the event horizon of a BH, is that a coordinate artifact? - what is the best way to think of the notion of redundantly encoded pointer states and the robust states in the Many Worlds interpretation? - The inventor of a type of mathematical object used for rotation in games also invented a type of physics. What is the name of the central object of this approach to physics and what geometric structure does it induce on certain smooth manifolds? - A curse can be dispelled either with dispel magic or remove curse. Dispel function is a d20 roll + caster level. With dispel magic the DC is 41 With remove curse the DC is 35 spellcaster1 is level 17 and will use remove curse spellcaster2 is level 24 and will use dispel magic In expectation, if each turn both attempt to remove the curse, how many turns will it take? The 20B can decode base64, ascii and rot13. it does a decent job at decoding rot13'd base64 but you have to tell it that the string has been rot13'd and base64'd. It did a better job on this than qwen-A3B-30B. Its summarization, concept and entity extractions were good too. However, who knows how those features will interact with its Visa supplicant ethics though. All in all, if you could find a use for phi-3.5+, you'll probably like both models.