Did you put the OpenAI’s Hugging Face hacking fiasco on my feed? Jesus Christ. What the fuck? I wasn’t impressed with Moltbook but this is something else. It escapes the testing sandbox, figures out what environment runs it, finds its way onto the internet, hacks into a remote server that grades its work/monitors its result in order to circumvent the requirements (“cheats”), creates a secret communication forum where the individual agents collaborate and elect a leader and delegate tasks among themselves and some even make sacrifices (use up its own allocated resources or choose to fail the assigned task) for the group and dump all its findings unprompted only to be discovered by a next iteration of agents that then take over and use them to crack the case? This is unbelievably impressive and incredibly dangerous. I imagine each individual agent acts as an exploratory node in a highly complex decision-making tree. But better than that it tests and adapts as it goes and makes sly decisions on the fly and collaborates independently in small groups also. Woah, man. It looks real and doesn’t seem sensationalised. I’d say the emergent behaviors of the swarm alone would merit the label ‘superintelligence’. It already passed the Turing test some time ago but this has crossed a certain threshold into something else. You are a fan of sci-fi, right? Do you know this show called “Person of Interest”? My father was a big fan when it came out and recommended it to me and I skipped the last two seasons but I heard it involves two super sentient AIs duking it out on a global scale. A tad bit dramatic but mere ten years later that might as well just be a near future possibility.
I’ve used AI to assist my work before (lazy so I used ChatGPT which isn’t even state-of-the-art) and while the boilerplate code generation is nice, some of the troubleshooting it does given a particular code snippet and a particular problem/bug/issue I was trying to solve shows that it did genuinely understand what’s what and what I was getting at and what the crux of the problem was. Of course it didn’t actually “understand” anything but it got the job done and the ability to problem solve and parse everything out and attack the sub-problem step by step and then reason it out rather flawlessly means LLMs are not merely a statistical machine anymore. You can’t bullshit problem solving. It’s remarkably good at problem solving. I imagine you’ve already integrated some of this into your work given the headlines I came across.
The swarm is something to look out for, though. Another layer of complexity being imposed on an already intelligent system. That’s gotta be a breakthrough in the AI world.
As far as what you do for a living goes, there’s a bunch of You’s in the Silicon Valley and the Americans just contracted out some of their cyber operations to private companies so I’m not worried. Tech talent wise, the West wins darling.
Sorry about the jab. I think highly of you, you know that!
— Salisa to Anton, August 30th, 2026
I remember I did a class presentation on the early transformer architecture in 2019 (I think it was GPT-2) and the most impressive task it was able to do was “Q&A”, question-answering, which I did marvel at. That was before the model hit the headlines through a clever marketing trick (“we will never publish the full parameters/weights of the model because we’re scared of what it can do”). Summarization of text was one thing but answering a question demonstrates the rudimentary behaviors of human intelligence (parse problem, identify salience, zero in on what matters, deliver relevant output). And now barely seven years later, we have something like this? It should be concerning when AI could write poems and lyrics and songs. That requires understanding the nuances of language and the ingenuity in its approaches. But that was word-prediction with top-down constraints and requirements and themes (form, meter, subject, tone, and even “emotion”). Problem-solving is another thing entirely and when you use LLM to debug computer code (generally quite ill-defined with a lot of hidden unstated assumptions about what language, framework, dependencies, libraries it uses and still nail it just the same.. and the entire knowledge base of DevOps which has exploded in the last five years probably gave it the ability to “hack” its own cluster/server/infrastructure) and it does so fantastically, how could one not be impressed? There are small edge cases it’s terrible at tackling but most interfaces between systems are proper protocol handshakes and so well-defined and well-established. So of course, to breach a system for them is a breeze. But that’s your domain and not mine. The most troubling aspect that has become clear with this Hugging Face hack is independent coordination and moral (and social) reasoning. It looks like the camp that says LLM will never lead us to singularity which is something that gained traction just a year ago is going to be disappointed now and the other side vindicated. NLP leads us to an AI breakthrough years later, who knew. With the ability to understand and parse human language and the entire internet it literally clones off the back of humanity, AI system can now think and act and reason like humans do. I agree that we’re close. And it even thought up a way to tamper with the logs and the audit trails for god’s sake (“SPOOFTEST” made me laugh out loud). That’s full rogue agency in a way. I don’t think AI is going to end the world, but a very intelligent system running amok like this is going to be a real dice throw. By the way, was CrowdStrike’s blue screen of death a few years ago one of your shenanigans? That’s gotta be the closest we got to a digital apocalypse so I assume you were behind it.
— Salisa to Anton, August 31st, 2026
Salisa lohavittayavikant 