AbstractPhila PRO

I'm currently updating the core architecture and WideRouter compiler capacity while simultaneously fixing a cache related bug. With that I've dedicated a specific part of all routers to caching; which is intentionally meant to be device agnostic and clearable on a whim throughout a pod.

Full ramp-up for accelerate testing is crucial very soon. I require easy access to multiple runpod-friendly, clustered docker, or indirectly connected devices with an api - to properly integrate the debugging and testing infrastructure. Currently the debugging infrastructure on the compiler is essentially nonexistent, it simply crashes with a huge readout that does not assist with debugging - so the only way to get the real debugged error is by simply turning off compilation to run the test without it. This is FINE without compilation, but once compiled the debugging is difficult to read. This can cut into time if you are unaware of the quirk and I want to ensure this isn't visible by catching all necessary elemental exceptions based on the interactions and device movements.

If I can get direct access to a docker system with multiple devices - cpu, gpu, interface, data, raw, tensor, numpy, whatever; I will gladly make a full wrapper structure to directly intermingle with diffusers, pytorch, or perhaps tensorflow alongside the structural system that I'm currently implementing. Please, share hardware. I'll gladly share my engineering and innovations. I'll focus my direct attention on whichever goal you wish first, then move from there.

If anyone is willing to donate hardware for a month or two I would be obliged to focus my attention on the needs of that someone or entity. I would happily share time slots and utilization days or months with others. So long as I can get the architecture working to capacity to provide the necessary implications and training to the models required. My only motivation is to directly train these experiments to match the hardware access and accuracy with deeper models, while simultaneously training multiple adjacent models in a stable fashion.

This is one of my primary goals, so reach out directly abstractpowered@outlook if you have any information. Access to stronger hardware will allow me to directly ramp up ablation tests and accuracy augmentations to machines that currently operate lesser than sota, while enabling the potential for the same complexity as a deep model training with many devices all tasked with different goals.

This is not an easy task so it will take time. Sharing resources, ideas, and concepts of utility I encourage as well. You can find me on discord as

_abstract_
aka discord name AbstractPhila with the popfrog display image, and in the Res4lyfe discord, most often frequenting the sampler madness channel as well. I have a fair history of chatting there and a few other locations for bouncing ideas and keeping daily notes.

Multimodal is interesting stuff

Well I managed 75% so that's a bit of a boost.

David's full trainer that runs from colab with only the lattice_vocabulary install was pushed directly into the AbstractPhil/gated-david repo as trainer.py - The current training script and process is now transparent.
Apparently I pushed it to one of the 3 repos accidentally created, so it's now in the currently visible public repo and will be pushed to the geometricvocab repo soon in a similar and nearly identical functionality, with additional controllers for freeze/unfreeze.
Many freeze/unfreeze mechanics are not required for many forms with the improvement to the baseline math. Not only that, but shared space between multiple versions of clip seems to have little problem.

Alright the new SOTA for modular structured feature learners is cooking. Be ready, the classifiers drop soon.

Imagenet features are being fed and the outcome is showing.

I'll be running it again with a proper upload setup then everyone can see what David is capable of.

I chose this route because I can have David in here almost immediately vs trying to make David standalone functional and getting massive headaches trying to run him over and over watching crash after crash because my old system was heavily AI generated instead of hierarchically created in a reasonably debug capable format.

geovocab2 houses the changes. The largest one being an INSTANT vocabulary load time vs the old one taking minutes to prepare the vocabulary. The LAZY loading with pyarrow support is far more powerful than any of the earlier iterations and I advise switching to the concept if you haven't yet.

AI ritualistically defaults to iterative, even though pyarrow with columnar is considerably faster.

The trie structure was established preparing the ngram structural trainer, which will be included directly into the lookup as an optional sorter comparator. The load time is nearly instant and the lookup time rapid. There are better formats for smaller processes, but this is meant to house hundreds of thousands or even hundreds of millions of ngrams, not just a few hundred. This structure operates really well on tpu; which is how I'll be training the upcoming vocabulary 5pair geometric feature structures - which will contain highly advanced and enriched learned structures between 2d and 9d shapes instead of JUST 5d shapes.

The rapid synthesis in the new system and the robust response from the test formulas show that these are highly enriched. The structural awareness of these crystals are more intelligent and robust than before by a large margin and the theta rotation only helps them rather than hurts them.

The next geometry will be trained entirely in fp64; established from numpy random crystals. The primary anchor of each is specifically oriented based on lexical frequency within the dataset and given a full shaped object based entirely on the lexical order.

Each ngram tree layer of traversal is meant to be given the parent's anchor and theta rotation applied - allowing the internal structure of that lexical order to not only be applied as a semantic and symbolic state, but also retain lexical complexity. This is a large step forward in cohesion.

Everything will be fully transparent. I'll hide nothing moving forward or reserve it, it'll be either Apache or MIT.

My goodness. When tinkering with David I ran into something substantially more potent. I'll need to run more tests, but it seems I found how to scale the pentas upward in a carefully curated way without shattering their structure.

Also David is mad outdated so I'll need to refit much of his systems before I can release him at all. The notebook currently expects a pickled series of imagenet tensors and a pickled series of imagenet crystals pre-selected at startup time - each organized specifically based on the curation.
That won't do, it'll need refitting.

This is only a logically and logistically correct assessment IF the assumption is based on curated data related to the very capabilities which your "mirror" requires to amplify those biases. The alternative is the LLM simply echoes reflective similarity directly associated with NEARBY echoed words, rather than logically related context and content. Instruct helps a lot, so does harmony, and the alternative forms of them - but the BIAS STILL FORMS.
If your bias is something that LLM does not have the relational associations to internal data nor has ever been taught the capability to deduct the logistical responsiveness from those biases, you are likely reflecting your personality quirks and biases onto a machine that simply cannot reflect them and thus the machine will simply... begin to echo those back to you.
This is a common self reflective bias that many of my introspective analysis and self analytical conversations defaulted to when assessing complex logistical and introspective analysis of large structures. This is most commonly amplified and incorrectly confident while discussing those problems with a single large LLM.
Communicate those same unfiltered conversational pieces to another large LLM and you will most definitely find different mirrored effects and different biases. You'll often find GROK, Gemini, and Claude all returns different responses to those same assessments.
Now... if all four say yes, the math lines up, the stars align, and the systems can in fact work if X and Y and Z - you might have a potential solution. EVEN THEN it will be a damn journey to make it work.
LLMs are often VERY WRONG even as a collective when it comes to large data in association with intricate complex technical work. Sometimes it's a single incorrect assessment from a random book that was fed in 500 times from a single topic that was simply disproven at some point, and yet there are still direct biases associated with those incorrect concepts. This amplifies the further you dive down the rabbit hole and is easy to confuse the llm, trick the llm with input, and even easier to break the llm's entire pattern with them because you're already so deep down the rabbit hole that you're accessing heavy noise.

I think I can handle the corpus training with some runpod MI300s or get a cluster of A100s for a week or two. That should allow proper tuning based on the lexical rules of language, but I need to make sure EVERYTHING is PERFECT before I start pulling triggers on clusters.

Also my apologies for not updating the lattice vocabulary, I've been very swept up in direct testing and implementing models. It's been really fun setting all this stuff up.

The more it works, the more I get excited that the formulas I'm manifesting are cohesive representations of purpose rather than simple random convergence. I've altered them hundreds of times, but the pipeline goal is still present. Unified geometric vocabulary WILL be a universal language, not simply a tinker-toy, but instead a full lexical representation of potential with all manifested trajectory and solidification of grammatical, lexical, symbolic, and representative substructure.

It's at the point where time will tell HOW this system is useful. Even if it can DO ALL THAT, large scale adoption or even minimal scale adoption is up to how robustly useful and how many eyes end up on the topics with technical knowhow. It's already well beyond the IF this system will be useful, which means I feel obligated to at least continue kicking my legs until I get access to a speedboat.

Simply put, I've built this system for the eyes of the technical - with some very direct and representative understanding to the less technical available as well.

There are some saving graces though. You can probably house the entire purpose of a word in a 256d token; but you won't get all of the robust lexical and analytical behavioral responses required from the orthonormalization 5th so it will likely be less accurate than a 512d.
You can get some more utility from upscaling 256 to 512 and you gain some sparsity which allows more growth, with the negative elemental response of sparsity being filled with no meaning - which tends to confuse and build pockets of misrepresentation on projection.
Multiple overlapping projections are the most robust from what I've been observing; where you take the same token and blow it up multiple times for multiple different projection sizes. This has proven invaluable behavioral response from the geometry 4-5 with freeze/unfreeze has shown that all layers can complementarily improve performance - while the final version can be any of them individually requested - as they are all experts on their own plane and the output does not require all of their outputs.
There are many potential variations of the models from these geometries - including 200+ projections implemented on the same model using the same tokens.
Pairs, triplets, quins, and penta word + letter combinations remain uncrystalized and unexplored, but I plan to use the same system to run them.
I'll likely implement a sentencepiece-esque translator that will turn a sentencepiece vocabulary directly into crystal variants with weighting for convenience, which will allow for much more utilizable and easy-to-represent vocabularies for expanding current models.
Wordnet with hard gated non-fabricated tokens has proven the most valuable, however they are still shallow and require full solidification and robustness curation with additional definitions and datasets.
Research is ongoing and many mechanisms still need to be created.