Ouroboros

Entry 1

You of all people would have known my resignation as forthcoming, it was not surprising to see the faces of Gerard and Fitzsimmons at the sight of my letter and subsequent meetings to plead me to stay. I've never seen someone beg me as diligently as Gerard did, I would be lying if I said it didn't stroke my ego a little.

You of course are my fondest companion both inside and outside of this organisation so I was hoping we may maintain correspondence - as good friends. As you know I will be moving to Canada, but to clarify more specifically, I will be moving to some property on a lakeside on Vancouver Island where I'll be building myself a log cabin. Contact with myself will be via email at best - cell reception is spotty, but I frequent a local bookstore with free Wi-Fi. The time difference is a bit hefty, I do hope we may call, but it will have to be during late at night or on weekends if I am to catch you outside business hours. I believe the Boston Pizza in town is open late, I am eager to converse with you over a late dinner if possible.

Disregarding logistics, there is something important that I wished to clarify.

When I had explained my rationale for leaving, it was verbal. As you know already, verbal communication is not my preferred method for conveying emotions, so when I told you the reason for my departure was because I simply wanted to leave the monotonous, mechanical meandering of an engineering org, I was concealing something a bit deeper. Allow me to briskly explain this - I have a number of things to tend to once I depart from the airport.

The practice of engineering, and the process in which I conducted it was not the problem. In fact, I quite love this as an art. It is however the demands that the corporate organism demands of the activities we engaged in. Our use of AI/LLMs was prolific, and this was great during the glory days of AI - none of us knew what we were doing. We were just throwing shit at the wall and seeing what stuck - one of my favourite activities. As we searched, we found what worked and what didn't. Then as the feedback loops from all the tech companies around the world saw major tightening, we suddenly had an explosion of permutations of solutions for everything. New strategies to coerce LLMs and integrate them with basically everything through secured CLIs. At this time, I saw what was coming, hence my move to DevOps long ago. Code itself was solved, interface was just begining, and we started trusting AI code reviews to a point where a statistically significant proportion were approved by humans with no corrections.

Once the UMIE model was formalized among infrastructure providers and adopted by most SaaS, ops was now solved too, but as you well know, the business we conduct requires some due diligence and redundancy, so we were still there. Even as frontier model providers licenced their models via infra providers and we removed the supply-chain risk associated, hence becoming fully independent of all third-party risk, senior leadership was skeptical about the lack of human oversight. Additionally, our customer base, and the general population, is not yet accustomed to the new ways of working using generative interfaces or chat-only mediums. Thus our roles have remained open, so in retrospect, I'm quite thankful of this situation compared to the bloodshed we saw in Feb 2029.

I wish to maintain providence over something that I can fully understand - but that is not my main gripe here. I've seen a concerning behavior that is beginning to proliferate into phenomena outside of work. I don't think I've asked if you've read 1984, but in such, the vocabulary that they develop for themselves is forever being refined to become leaner and leaner. I am in no way implicating some clandestine conspiracy here, but rather drawing a similarity. By nature, LLMs show adherence to generally the same finite domain of permutations in their outputs, no matter their temperature (We've proved this through research. I'd be happy to dig it up again for you). That is, LLMs generally converge to a countable set of categorical vocabularies and mannerisms. This is not talking about the product of its output, but rather it's atoms. My hypothesis is that at some point we will achieve some 'critical mass' that will see LLMs start to produce responses that are unhelpful babble. As if talking to someone who speaks many words but says nothing.

I would pull up some examples, but they are all in my laptop (in checked baggage). Once I've got some space to settle in my new place, I plan to test some ideas that I had myself. To be frank, I did attempt some of these while inside the org, but they were either futile, or partial.

In the meantime, I'd be truly grateful if you could occasionally update me on the happenings of the engineering scene. Now that I'm away from the orchard, it's fruits are far more savoured by my tongue, I wish to still maintain this connection somehow. I'm hoping to rely on you in this regard.

My flight appears to be boarding now. Hope to hear from you soon!

- Jared

Entry 2

Greetings again, thank you for taking the time to write such an elaborate response. Once again, I am truly grateful of this connection we maintain. Allow me to first update you on the progress with my home.

We have erected a small shed with some insulation such that we have a somewhat functioning living quarters while I work on the first room. All the walls will be made from local cedar with a maple sheen. I've always quite admired this aesthetic about Canada. Seeing a building made of those gorgeous maple logs gives me a sensation of cosiness, and I want that in my abode. I've found part-time work as a farm hand on a neighbouring farm, and have been testing the resolve of my muscles by accepting various odd-jobs around town that usually involves splitting lumber and transporting it. My other part conducts lectures and tutoring at UVic, and much prefers to sleep in their family condo, especially during colder times. I take my rest there with them over the weekend. I drive a beat-up old Ram truck, manual of course. First time dealing with diesel, it's a weird feeling not seeing an 8 on the tachometer. She also feels very sluggish, but she makes up for that in brute.

As for your correspondence on engineering, this is an intriguing observation you present to me. I thought that code comments were eradicated after the research from Armaan J. showed that they cause catastrophic long-term hallucinations and model providers added that custom coercion layer to their agentic engineer models. I think it's quite prudent of you to have taken a small sample of PRs to examine. However I believe your comment on the covert nature of the fruitless examination by agentic inquiry is fairly hasty - unless of course you meant this with a shred of sarcasm. The changesets you are investigating indicate an inert behaviour, it's possible that some of these are simply invisible or subliminal to the agent. Additionally, your investigation is an agentic operation, not a defined process or corpus, so by definition, it will be non-exhaustive.

Based on the example you provided, the comments themselves theoretically shouldn't lead to any negative long-term hallucination - they are seemingly meaningless and would reasonably be ignored by the agent. They appear to be some sort of licence or code signature, a concept which pre-dates your tenure. These were often comments that were left at the top of the source file back when we were using third-party libraries. However, what is truly bizarre to me is that these comments appear as headers to function definitions. Like I said, typically you would find them at the top of the file instead. I remember that in the really old days when we had to write our own documentation, that one method was to write the documentation as a comment on any method or class. In these comments you will find specific directives that indicate, to a custom interpreter, various parts of the documentation (like the method arguments, return value, and whether any exceptions are thrown). The interpreter will then compile a set of HTML documentation.

But what you have provided me here is of no known nature from my knowledge. It's simply a vacuous paragraph that has a load of legal jargon in it. If I pass this through G, it just simply hallucinates the meaning of each sentence - like this is an excerpt of a larger section with context elsewhere. Correlating it with other documents on the web with Pexa yields no known results, similarities include the terms and conditions pages of various sites. I've also tried looking for reports from others online, but as you know, no one really posts about this stuff publicly anymore since mostly everything is in-house at major orgs now. Ever since the collapse of all the package managers like npm and such, no one really raises GitHub issues anymore because they literally have nowhere public to go.

Is this the only instance of a comment like this (ie, a signature of sorts)? Or are the other instances you mention much simpler? Or more coherent? Do they have similarities? What did the latest version of Cogent say about this (assuming the org subscription has access to this one)? I've been hearing that their recent strides included an enhancement to large-scale pattern recognition with a 4M10D token context window.

I keenly await your reply.

- Jared

Entry 3

I had a peculiar interaction at the local clinic that I thought I'd note to you in this letter. I noted that at the bottom of the PlatiMed AI report there was an unusual repetition or rephrasing in a closing remark. It read:

"The information contained in this report is of general nature only. Usage of this information by the patient is done so at their own risk and PlatiMed AI will not be liable for any injury, illness or death resulting from the usage of this information. The usage of this information is strictly not to be used by the patient. Any resulting injury is the liability of the client/patient. Furthermore, any usage of this information will require the patient to understand that the usage of this information could cause injury or harm, and such resulting actions are not the liability of PlatiMed AI. And as a result of the information, then undeniably this usage will see to it that the information remain as such without the interference of any outside affairs pertaining to the usage of the information in a way that is to cause harm or injury or death to another due to the incorrect usage of this information."

I usually don't read these things as I prefer to make my own judgements based on my own understanding of the human body, but the unusually large footer on this document had me intrigued.

As usual, these documents are quite vague, but the body had made sense, although the footer appears damaged in some way. I can't seem to find any reports on this around the web - yet again my Pexa net is empty apart from some terms and conditions.

I suppose that things like this are generally inert would you agree? What I mean is why would anyone go through the bother of reporting this error? If anything, this confusion will benefit the plaintiff more than PlatiMed. It's advantageous to the consumer, yet invisible to the org. Not necessarily a bug in the traditional sense, but a subtle side-effect of a product decision that no one had stopped to think about. Also no consumer wants to jump through the system behaviour hoops in lowest priority end-user support agents.

It makes me wonder whether this output I have here will be recycled into training data, or if this 'glitch' happens elsewhere, will frontier models adorn this behaviour too? Surely the training sets nowadays use some sort of rigorous cleansing. But then again, who knows what the hell they are doing. It's like the Borg now, whatever the engineers working at these AI oligarchs think becomes some sort of solution hobbled together just to stay a fingernail ahead of the competition. Hell, they probably have millions of their own agents working to clean every inch of their datasets, if my hypothesis is right, they might be poisoning themselves.

What do you think?

Regards.

- Jared

Entry 4

Wow, it seems it's finally here. I recall before my leave that LiveNet was just a theory. To me, this leak of its internal usage at UniModel is profound not because of the fact it was being used experimentally on production requests, but because they've been doing this for 8 months now!?!?!

In my experience (and also personally) engineers don't have very tight lips when it comes to shiny new tech. I too have leaned into the folly of blabbering to my other engineer friends about how some of the new stuff I'm building works. So it baffles me how they were able to keep it under wraps so covertly for so long. This explains their meteoric valuation climb after some of the new model flavours they released, I suppose these ones had the LiveNet retrofitted to them as the research paper was being finalised, or I guess, as an experiment.

I don't think that this is AGI like many others seem to be saying, and although you make a convincing argument, I'm still in discordance. All LiveNet consists of is a MetaRAG with combinatorial Alice vectors that translate the source knowledge into a 'coercion overlay'. I don't understand what this last part means, but based on a rough chat and glance at the paper, it seems to be somewhat similar to a LoRA model (god this brings back memories!) in the sense that it augments the neurons, but it does so in real-time as information in the corpus changes.

It is scary to think now that model training is no longer necessary. We can simply build a comprehensive web corpus and feed that via LiveNet to a frontier model. This means the model can improve itself in real-time, but will also regurgitate whatever it creates. Essentially we have tightened the feedback loop that model training provides anyway. It's funny how nothing foundationally has changed, but instead we have just allowed ourselves to get closer to finding out what's at the end of all this in polynomial time.

It irks me that I can't seem to foresee what should happen next with AI. This question itches me every time I'm stacking lumber or treating some logs. I can't really see anything ahead of LLMs as it currently stands. We've embedded this technology into so many things now that it seems we want it to perpetuate the core sycophantic behaviours and hallucinations simply because we have confidence that, when coerced effectively, it can perform a repetitive function that cannot be automated traditionally or must be adapted to. That is to say that there is no need to 'train' a model that can 'detect' its own gaps in knowledge. If you think about this as deeply as I have, you may find that the notion of such an entity is paradoxical, such a problem to solve will reach back to the days of shadows cast in a cave. To put plainly, you will begin questioning the very foundations of knowledge itself.

Apologies for having led this letter astray with a bit of sophistry. I did want to remark on your idea for rotating telemetry probes to reduce costs. My question to you is: where is the telemetry data being stored? If the AI chose a land-based medium, then perhaps that is draining your costs. If you use satellite Valkey then you will have roughly the same latency and can automatically configure erasure to a time period that the AI deems as appropriate. Remember, now that the AI rolls the releases, the issues themselves persist mostly in a reasonably calculable timespan. That, and also the fact that live customer telemetry exists now given the redressing of global privacy standards. If anything, you should set up sentinels that analyse individual customers closely when they report a problem.

- Jared

Entry 5