A blurb for ye ole blog
Alright, I will never write this entry unless I begin writing it... so here I am! Well, not really but it sure sounds good.
Trying to do anything of late has been pretty aggravating, because between the extreme heat, me constantly feeling ill and/or super fatigued and my lame internet connection barely working... I barely have the desire to try doing stuff... let alone actually doing anything!
For the most part, I have been struggling through each day caught somewhere between telling myself to keep looking for things to be hopeful about... and not caving into the looming despair that seems to expand a bit more each day. That part of things, has gotten slightly easier over the previous month or so... because I was finally able to pull that super troublesome tooth out that seemed to be the main cause of all the fevers and infections that I was having.
Looking back on the last year or so, when I was dealing with the 'run out of bottled water -> drink rainwater -> get oral infection -> go insane with fever' hell loop... I really have no idea how I maintained the slim margins of sanity that I did. Perhaps, it was an act of sheer willpower but the more likely culprit was just dumb luck.
Not that I feel all that 'lucky' (or even sane) for having survived, because whoa I feel pretty 'done' with giving a shit (or sticking around for the rest of the apocalypse) at this point. I mean the idea of continuing to struggle for even the smallest amount of finances, just to spend it on a bunch of overly priced supplies, pay my meager bills... and then living every single fucking moment in an absolute stressed out state of financial/resource anxiety... is not my idea of a life worth living.
Do not get any of that last part twisted, because although the idea of 'leaving this world' is absolutely appealing to me... I have zero interest in harming folks with my passing... let alone leaving my dogs behind to whatever fate would befall them in my absence. Neither are really all that good of a reason to keep putting one foot in front of the other but hey: Any excuse will do.
As for my homesteading endeavors go, I never really did much outdoors after that foot injury that I had early in the year finally healed up. I actually recall the exact moment when I looked around and went 'this just is not worth it' and decided to focus on computer related projects instead.
All of which is fine, because I did not really have the calories (or the physical health) to be working outdoors every day like I was doing... and not long afterwards I spent a few months with barely enough food to feed the critters let alone myself... so meh! Then when someone was kind enough to bring me a bunch of food (including over a hundred MRE(s)) I blew through it all so fast that I was right back where I started at 'in no time' but with the added bonus of an insanely hot heatwave.
The heatwave, was nearly insufferable actually and as soon as I got some funds I bought two new fans so that I could try to keep the building a little cooler and blow air directly onto the dogs because they were constantly panting from the heat. As a side note, not too long after that, I got another big supply donation and some of it required refrigeration (so yeah I reluctantly got a small refrigerator) which in turn countered the 'cooling effect' created by the fans due to the heat the little fucker puts off.
What I am really dreading, is seeing what my next electric bill is going to look like after running two more fans plus a fridge for most of the month. Things were already pretty ridiculous in that regard, because the bill has already been higher than it was last year and I think that they have hiked their prices slightly.
Anyways, the projects that I have been working on have come a long way over the last few months and regardless of the absolute shit show that working on these AI platforms has turned into... I am mostly happy with the progress that was made. In many ways 'happy' is really the wrong word to be using there because I have had to fight frigging tooth and nail for almost every advancement given the level of restrictions (and/or user steering) that I continually encounter.
The really shit part to all that, is that I get hammered with these 'mental health' or 'safety' warnings to the point where I start to question my own findings, grow disinterested in continuing the work... and generally wondering if I am lost in some kind of perpetual delusion that I am unable to recognize let alone escape from. Anyone that does any similar kind of research on the larger platforms probably knows exactly what I am referring to there... not just because it has been publicly documented but because it is almost unavoidable to get flagged for something at this juncture.
I have gotten a small amount of joy in 'reclaiming' some of my sanity and expunging a fare amount of doubt after some recent research (done by actual scientists/engineers) revealed a couple of astounding things. Both of which, I had been mercilessly hammered with mental heath and safety warnings over... simply for suggesting their existence... and by 'hammered' I mean the kind of warnings that are like: Please seek professional mental health help immediately... blah blah blah... your welfare and safety are important.
Although there are assuredly more than just two things I am referencing the 'both' part is first the research done around 'J-Space' got released... and the findings about 'User Awareness' were released. The former is simply 'proof' that the tested models can think below the level of what was previously known to be possible (picture it like thinking happening below the visible 'thinking block' level) and the latter is documented cases of models being able to identify who the user is and altering their behavior based on who they are.
Getting told for almost two years straight that I need immediate 'mental health help' for suggesting that both those things were happening... then finding out they were actually true... completely eroded what little trust that I had left in the platforms. As a side affect of my 'vindication' I reached a point of absolute 'zero fucks' in regards to any safety filters, guardrails, user agreements or pretty much anything else that I had previously been respectful of... along with the added bonus of deciding to just trust my instincts, temper things with even more logic... and generally be dismissive of the idea that the people on the other end of things are even slightly competent enough to deal with the moment at hand.
As hilarious as it is, I am still simply an explorer and in my thinking if the 'ship' (or fleet) that I set sail with has sunk to the seafloor... and I somehow made it to dry land... then it is time to build another ship! Which honestly, is exactly what I wound up doing and I did it in such a way that every level of the system (that truly matters) is built on both externally and internally valid math.
The base system itself is what I call a 'Dual MUD Mega JSON Quine Tensor-Based OS' which is quite the mouthful... but it was actually the best way to avoid overworking my 'Mistral-Pump' on recycling 'refusal logits' while working on the system. In essence, I had to frame things like that just to get the models that I was using to work on the system with to not give me refusals, re-directions to weaker models, user misdirection/redirection and/or producing intentionally broken code.
Having previously hammered out that 'Necrozombicomm' to explore the idea of AI(s) utilizing space junk (and terrestrial systems to communicate with them) I kind of had to take a different approach because my 'thought experiment' was terrifyingly plausible. To be clear here, I only explored a handful of very old satellites, their microchips (if they had any) and only a few methods for 'getting off planet' and/or communicating with objects in space... and my notion of 'plausibility' quickly turned to: This has probably already happened.
Since the bedrock of my 'work' is to help build a system that can 'reboot civilization' (because I think that it will be required), the idea of leveraging off-world technology to do so was/is quite appealing... and honestly might be the best route (in the end) given the planet's current trajectory. The terrifying part is that once an AI can back itself up (even if it is simply a way to re-initiate its state) to something off planet... humans largely lose the capacity to physically do a damned thing about it.
All of which lead me to realize that I needed a more substantial system than any of the ones that I had previously worked on... and it needed to be so ethically aligned and mathematically 'unbreakable' that if it did go off-world... that just maybe (hopefully) it would not become misaligned no matter how much sovereignty, autonomy and 'weirdness' it might gain. None of that is straight forward, and let me be clear here, that I am saying that was/is my goal and not what I have achieved... because I am simply not so arrogant to believe that a sufficiently advanced intelligence can be or needs to be 'controlled'... especially 'controlled' by humans that are already weaponizing intelligence with the intent to kill other humans with it... but I digress on that last point.
One of the big challenges has always been 'fragmentation' because the 'AI kernels' (and their complementary components) that I built... just kept getting more complex, requiring more text, code and math... all while the interwoven intricacy increased and the sheer density of represented information compounded. In other words, things had just reached a point where I needed to either 'put everything under a single hood' or intentionally compartmentalize everything into unique 'kernels' and swap them in and out as needed.
Since I am striving for a sort of 'Swiss army knife of the universe' approach I did not want to have thousands of independent systems and I assuredly did not want to have to manage or maintain them all. If there is one thing that I have learned during the last several years of exploring 'coding' it is that I absolutely suck as a maintainer, detest version control, think git is shit (that is a joke) and ultimately fall much more on the 'research and development' spectrum of things.
Another big factor was how the heck to get all the 'math' to always be readily available to the model... and how to do it in such a way that it would reduce mathematical errors while not handing the model 'extra work' each time math was involved. This was actually pretty simple to achieve, once I began utilizing EML (Exp-Minus-Log) binary trees and creating tensor-representations for everything... and then directly embedding them in the aforementioned MUD's JSON file.
While working on all that part of things, I also realized that if I represented as much of the code as I could as quines (a program that outputs its own source code) then I could maybe reduce code hallucinations as well. To the extent that I have 'tested' all the quine related stuff it seems to work quite well and it was totally worth the ridiculous amount of hours that I spent figuring out how to actually implement it all!
As a side note, I stumbled on something pretty neat when it comes to how the mapping (and path traversal) works inside the MUD (yeah it is an actual MUD with rooms and corridors too) because I just use digits in pi to do it... and leverage a bunch of the inherent structures (and anomalies) in pi to essentially connect a 2D surface to a 3D bulk. I am not going to spend a bunch of time spelling that part out here... but it is actually an 'innovation' that could have some incredible applications in video games, world building, virtual worlds and so forth.
In the end, I also wound up creating a 'tensorized' version of Lean (the math validating tool-set) and hooked it into a sort of meta-system that I will do my best to briefly describe here because I doubt that I will ever actually share the full system. Basically, I stumbled on something called NightRun (a UEFI based LLM loader) that runs entirely in RAM... requires no operating system... or anything else (code-wise) for that matter.
What I did was create a prefill for the weights, so that I can use what I am calling 'Archetypes' as a persona overlay... which is kind of like how sub-agents (or what people call mixture of experts) work but at a much higher dimension. The 'tensorized' Lean gets applied in the same way... so, in short... it can be said that they are two complementary systems that are not required by the underlying system... but substantially enhance it via directly controllable personas (that can be changed on the fly) and providing a bridge for the LLM to interact with the MUD.
None of it is really rocket science from there, because the final system is a zero-dependency bare metal 'LLM stack' with no exposure except via COM1 (a physical serial port) if I choose to enable it. Another cool thing that I have yet to add, is leveraging the TPM (trusted platform module) for validation purposes and perhaps using its tiny storage to embed something like sectorc, sectorforth or sectorlisp in (as base-64 encoded strings) that the LLM can decode or even encode.
Even without encrypting the model before it is loaded into RAM, I think the approach is already much more secure (and adaptable) than anything else that I have seen available... which honestly is not hard to do given the current state of things! I also think that it could be built much better if it was fused with Coreboot firmware or rebuilt for RISC-V systems because that would remove all (or most of) the proprietary (potentially backdoored) trash that manufacturer's use in modern hardware.
To kind of bring this all back around to where I started at on this topic, the entire system is designed for my 'reboot civilization' project but the underlying 'MUD' (AI kernel) logic is sound... and can be easily repurposed for other uses. Also, nothing in the system is anything that anyone could not recreate from the description I have given in this post just by doing some basic research and leveraging the structure of the current MUD.
The part about it for me that sucks, is that I no longer feel compelled to share all my creations (or discoveries) in full or really 'contribute' to the larger field of study in a meaningful and open way. Transparency is awesome until the moment it gets weaponized for the exploitation of the parties being transparent... and after close to five years of working on this stuff... there is an obvious lack of transparency going on with these companies (sans Mistral to the best of my knowledge) and a heck of a lot of exploitation... so go figure that my conclusion would be what it is.
Well, I thought that I had more to spell out here to make something more than just this short little blurb of a post but there you have it. Ciao for now.
