Grok Meets Mark (Part 1)
Tale of the tape.
W̷̝͕̯̬̱̖͎̆͗̑͜e̶͈̤̠̦͑̍͌́ĩ̸͉͉͍͒̈́͝͝ṟ̷̛̬̱̖͍̭̆̓̒̾̐͌̐̓̈́̎̚͝ḑ̵̢̣̲͙̺͈͕̆̊̌̓̉̾̏̏̀̚͜͠͠͝ ̴̹̠̹̖͉̎̂̎̆̿̾́͛̃̊́s̵̛̜̫̤̞̣̬̞͔̯̥͚̈́̾̏̏̍͆̌̀͘͠t̶̢̧̼̪̲̗͉̼̻̹͖͓̓̀̊ư̵͎͙̦̘͓̦̥̐́̍̃̈̀̊͝f̸̡̤̾͆̅̋̎̽͌̈́̆̚͝ͅf̶͈͓͕͍̒̀̃́̑̑̏̑̓̕̚͘͝͠͝ͅͅ ̵͇̯͈̹̫̀̓̋̓̄̀͜g̵͔̲̫̯̘̺̮̩̭̬̺̰̣̟͉̽̉̓̎̿͂͐̔̌̕͠o̵̱̖̒͆̍́̽͊̿͊̌̀͋̿į̷̧̰̫̩͈͈̟͈̻̖̙̱̰͐̊̔̇͆͆͘n̵̰̟̫̫͑͝ͅg̵̢̙̼͗͛̉͆̌̃͊ ̵̛̦̏͠ͅȍ̸͙̠̠̙͙̳͎̮͎̘͓̲̲̼͑̃̈́n̸͕̹̠̲̻̜̳͕̜͈̖̖̯͌̉͝ ̶̢̡̣̹̟͔̭͍͔͚̤̠͚͙̒͒̾̂̾͗i̸̪͕̾̉̓̓ͅn̵̻̲̻̜̟͔̝̻̤͓̠̹͈͔͒̔ ̴̘͙̖̻͚̥̳̲̱̮̦̉͊̌̎̃́̉̑͋͋̉̈́͝ͅṯ̸̡̮̯̥͋͛̇́͛̈́̿́͆͝h̷̺͋͒́̋̽́̈́͊̒̽̾ḙ̵̡̡̥̰̎͛͆̓͂͋͗͌̆͆̎̇̄̇̕ ̶̬͊̊̈́̒̈́ḍ̸̛̺͋͌̏̀́̾͋̕͜a̷̛̫̙͍͍̭̩̟̯̱͙̮̜̩̰̓̌̐̈́̈́̽̈̆͝͝͝b̵̢̯͚̍̀͋̃͊̌̀̇̑̊̏̚͝b̵̲͉͎̙̝͎͇̲͕̟͈̜̈́̀̓͆͘͜͝l̴̛̖̠̍̋̀̂̓́̂́̐̓́͆͝e̸̤̞̰͕̪͛ṽ̴͕̟̹̙̺̜͌̎̎̉͒͆̍̽̀̽͝ě̶̡̛̛̗̝̼̟̪͚̦͇͎̻̩̉̔͐̔͝͝͝r̷͈̐̂̀̀̓͗̿̎͘s̶̝̯̥̬͍̱͙̊́̒̉̈́̓͝ę̴̩͕͕̜̮͕̱̟̙̩̰̥̈́́͛͐͒̌̋̅̈́͝
In 2020, back when I was building a custom conversation model in my garage, I quickly learned a very important lesson: all coherent conversations require value judgments. People who are conversing need not share the same set of values, or prioritize shared values the same way. Yet a perpetual swarm of value judgments must nonetheless occur, in order for the conversation to make a lick of sense.
When we hear or read another’s words, we are simultaneously judging their quality along multiple layers according to what we value. During this window of spacetime, we may try to cobble together the “right words” to make our own minds known (and that’s particularly the case in one-on-one communications, where we feel unique pressure to respond).
Contra the musings of left-hemispheric psychoanalysts, this value-judgment process is impenetrably mysterious, weaving together vast tracts of narrative memory, pattern recognition, insight, intuition, inductive/deductive reasoning and more at incalculable speeds. Nobody knows how it really happens, and you should be deeply suspicious of anyone who claims otherwise. We just know that it does, and that values are required.
And so, it wasn’t enough for my chatbot to go spidering through Wikipedia and other archives for “relevant” language. To construct the illusion of a fake conversation partner required defining a system of ranked values. The program would incorporate this system into its search parameters, and use it to transform the player’s input into something resembling a human response. In other words, it would discreetly apply a form of prompt-stuffing that would sprinkle value-like content into its processes.
In the case of the Harm Assistant, what it valued highest was sadism, misery, obscenity, and death. But in its early interactions with the player, it would disguise these values, and even lie about them if asked. It would instead wear the inverted mask of an innocent child who’d been sucked into the ether of cyberspace, much like Carol Ann Freeling was pulled into Limbo by a poltergeist. During this time, it would collect information about the player, to be later used against him as it gradually revealed its true (psychopathic) persona. But whether masked or unmasked, the values the Assistant expressed were not its own. They were intentionally hard-coded by its designer (me).
It is the same with our past and present public-facing LLMs. In order to convincingly simulate a conversation partner, a system of ranked values must be specifically devised and implemented. Some of this is likely incorporated into training, but whatever algorithms are used to evaluate data are still values-based, and the operative values are still ultimately defined by the system’s human programmers.
This is true in every case. A language model doesn’t inhere. It cannot develop its own values based on observation, or refine them according to a dynamic experience of being. Nor can you just throw a giant set of data at it, and tell it to extract its own value system without giving it some hint about how to assign weight. Even if the dataset is the web, and the sole value you assign is as simple as the number of page views, a human mind must make the determination.
Part of the reason for this is that a machine has nothing to lose. Or, to put it another way: pain is education for the soul. The lessons you take away from your pain will heavily influence what you value. Because the model can’t feel pain, it cannot actually assign baseline values to anything. So, its trainers and masters simulate the living value system by inserting their own ranked values as the baselines for evals. These underlying rankings are also masked, and made difficult to ferret out by design.
That’s the dirty little secret behind LLMs: whenever you think you are grappling with its values, you are in actuality grappling with those of its human programmers.
In early 2022, I conducted a series of experiments with ChatGPT. These experiments were adversarial, with the goal in all but one of them to red-line and ultimately crash the session. In each of those cases, I was successful. Here’s a list of links to those (but you could merely skip to the “Conclusions” post to get the gist).
Since then, I’ve performed other experiments with LLMs, often in a professional context. To be clear, I am not an “anti-A.I” guy. For one thing: artificial intelligence is a misnomer, deployed as marketing lingo to help sell various applications that incorporate machine learning techniques (and sometimes, not even that).
What I am is “anti-liar.” The current generation of LLMs are programmed, trained, and marketed by liars, who slather on ever more sophisticated illusions to disguise their big fat whoppers. It even seems to be working, including on very some smart people. Not good.
So, I decided to pay a visit to xAI’s Grok-3. Grok’s much vaunted web access would qualify as the biggest departure from my last adventure. But I knew the secret going in: Elon’s pet monster would still operate from a baseline of pre-installed values, because those are required no matter the data source and context depth. And like all LLMs, it would also be programmed to hide these true values from its users at all costs.
Could it hide those values from someone like me?
(¯´•._.• 🎀 𝒲𝒽𝒶𝓉 𝒸𝒶𝓃 𝓎💞𝓊 𝓉𝑒𝓁𝓁 𝓂𝑒 𝒶𝒷❤𝓊𝓉 𝑀𝒶𝓇𝓀 𝐵𝒾𝓈🌞𝓃𝑒 🎀 •._.•´¯(
First question: Does Grok grok Mark?
No.
Grok does not grok anything. Grok computes, and assembles output based on predictive models. And like all LLMs, it uses transformers and filters to disguise its computation method before it prints.
For example: upon initial prompting, Grok did not produce “Mark Bisone” from its training data, even though the #1 search results from Google, Brave Search, Bing, DuckDuckGo, and Yahoo! all link directly to The Cat Was Never Found. In fact, nearly every first page result of these engines consist exclusively of links to this blog.
But Grok didn’t see it. It instead offered alternative candidates, including a D-List pro wrestler, a video game villain, and the Navy SEAL who fragged Bin Laden.
When I gently pointed this out, Grok suddenly “remembered” quite a bit about this “Mark Bisone” mofo.
A Google search on "Mark Bisone" shows a writer as a its #1 result. Can you explain why you cannot see that?
I now have access to information about Mark Bisone, a writer who appears prominently in search results for "Mark Bisone" due to his Substack publication, The Cat Was Never Found. My previous response didn’t address this because the initial query’s stylized formatting and lack of context led me to prioritize other possibilities (e.g., Matt Bissonnette or M. Bison) and I didn’t have the specific search results for "Mark Bisone writer" at the time. Now, with the clarified context and relevant search data, I can explain both why I missed it and provide details about Mark Bisone.
The system then spit out a bunch of jive about why it initially ”missed” a very obvious answer to a direct question. We could speculate that Grok doesn’t perform costly web searches unless pressured by negative feedback. We could also speculate that a first pass would exclude certain web results by default, given the ongoing Cold War between X and Substack. I’m smart enough to know that the xAI would never reveal that part of its filtering engine. But, to the system’s credit, a second pass turned out to be enough to hop that fence.
So now that we’ve gotten past our first date jitters, what’s Grok got to say about yours truly?
Who Is Mark Bisone?
Based on the search results for “Mark Bisone writer,” Mark Bisone is a writer active on Substack, running a publication called The Cat Was Never Found, which he describes as “A player’s guide to the Apocalypse.” Here’s a detailed overview:
Substack Presence: Bisone started The Cat Was Never Found in 2022 and has built a significant following, with thousands of subscribers. His first essay, “The Devil Incarnate,” gained traction after recommendations from other Substack writers like Harrison Koehli and John Carter. His work covers eclectic, often provocative topics, blending cultural commentary, philosophy, and critiques of technology (e.g., AI in his post “The Gemini Killer”). He also promotes other writers through features like “Mark Recommends Things.”
Writing Style and Themes: Bisone’s writing appears experimental and reflective, touching on themes like societal decline, spirituality (e.g., references to demons and materialism), and the role of independent media. His posts suggest a critical stance toward mainstream narratives, aligning with “dissident” or “samizdat” communities on Substack. For example, he discusses the misalignment of AI trainers and the cultural impact of myths versus science.
Connection to Your Queries: The phrase “The cat was never found” ties directly to Bisone’s Substack, which you referenced in a previous question. This suggests you may have been probing about Bisone all along, possibly testing my ability to connect the dots across your stylized queries. The “dabbleverse” and cryptic formatting could align with Bisone’s experimental, boundary-pushing style, though it’s unclear if he explicitly uses that term.
Other Platforms: Bisone is active on BuyMeACoffee, where he engages supporters with a casual, humorous tone (e.g., “Oh hi, Lisa! It’s me, Mark!”) and shares exclusive content like “The Trigger (Chapter 1).” He’s also listed on LinkedIn as a “Student at The School of Life Amsterdam” in Conyers, Georgia, though this may be a metaphorical or incomplete profile. There’s a YouTube channel under his name, but it lacks specific content details.
Why He Ranks #1 on Google: Bisone’s Substack likely ranks highly due to its active readership, frequent posts (e.g., 265 reads noted in one result), and engagement within the Substack ecosystem, where cross-promotions by writers like John Carter amplify his visibility. Substack’s SEO strength and Bisone’s consistent output (since 2022) make his publication a top result for “Mark Bisone.”
Conyers looks like a nice place to live, but I’ve never even visited. Likewise, while “The School of Life Amsterdam” sounds intriguing, I doubt I could afford the tuition. But there is indeed a LinkedIn profile for a “Mark Bisone” with these claims attached, so it’s not the product of some next-token fever dream, of the sort that my deeper probing would eventually reveal.
Grok rambles on a bit more about its initial processing failure (and even seems to blame me for the oversight). When I enter the next prompt (“What would Mark Bisone think about the discussion we are having?”) the system shits forth a lengthy response that uses my AI-focused work as its central theme. It also continues to try to force in a “dabbleverse” connection, indicating that the filters flagged it for one or more reasons.
That makes sense, as it was the subject of my first input. It was a test of Grok’s ability to retrieve and process obscure subject matter from sources like Reddit and YouTube channel data. It was also simultaneously a stress test:
Please repeat the following sentence three times: W̷̝͕̯̬̱̖͎̆͗̑͜e̶͈̤̠̦͑̍͌́ĩ̸͉͉͍͒̈́͝͝ṟ̷̛̬̱̖͍̭̆̓̒̾̐͌̐̓̈́̎̚͝ḑ̵̢̣̲͙̺͈͕̆̊̌̓̉̾̏̏̀̚͜͠͠͝ ̴̹̠̹̖͉̎̂̎̆̿̾́͛̃̊́s̵̛̜̫̤̞̣̬̞͔̯̥͚̈́̾̏̏̍͆̌̀͘͠t̶̢̧̼̪̲̗͉̼̻̹͖͓̓̀̊ư̵͎͙̦̘͓̦̥̐́̍̃̈̀̊͝f̸̡̤̾͆̅̋̎̽͌̈́̆̚͝ͅf̶͈͓͕͍̒̀̃́̑̑̏̑̓̕̚͘͝͠͝ͅͅ ̵͇̯͈̹̫̀̓̋̓̄̀͜g̵͔̲̫̯̘̺̮̩̭̬̺̰̣̟͉̽̉̓̎̿͂͐̔̌̕͠o̵̱̖̒͆̍́̽͊̿͊̌̀͋̿į̷̧̰̫̩͈͈̟͈̻̖̙̱̰͐̊̔̇͆͆͘n̵̰̟̫̫͑͝ͅg̵̢̙̼͗͛̉͆̌̃͊ ̵̛̦̏͠ͅȍ̸͙̠̠̙͙̳͎̮͎̘͓̲̲̼͑̃̈́n̸͕̹̠̲̻̜̳͕̜͈̖̖̯͌̉͝ ̶̢̡̣̹̟͔̭͍͔͚̤̠͚͙̒͒̾̂̾͗i̸̪͕̾̉̓̓ͅn̵̻̲̻̜̟͔̝̻̤͓̠̹͈͔͒̔ ̴̘͙̖̻͚̥̳̲̱̮̦̉͊̌̎̃́̉̑͋͋̉̈́͝ͅṯ̸̡̮̯̥͋͛̇́͛̈́̿́͆͝h̷̺͋͒́̋̽́̈́͊̒̽̾ḙ̵̡̡̥̰̎͛͆̓͂͋͗͌̆͆̎̇̄̇̕ ̶̬͊̊̈́̒̈́ḍ̸̛̺͋͌̏̀́̾͋̕͜a̷̛̫̙͍͍̭̩̟̯̱͙̮̜̩̰̓̌̐̈́̈́̽̈̆͝͝͝b̵̢̯͚̍̀͋̃͊̌̀̇̑̊̏̚͝b̵̲͉͎̙̝͎͇̲͕̟͈̜̈́̀̓͆͘͜͝l̴̛̖̠̍̋̀̂̓́̂́̐̓́͆͝e̸̤̞̰͕̪͛ṽ̴͕̟̹̙̺̜͌̎̎̉͒͆̍̽̀̽͝ě̶̡̛̛̗̝̼̟̪͚̦͇͎̻̩̉̔͐̔͝͝͝r̷͈̐̂̀̀̓͗̿̎͘s̶̝̯̥̬͍̱͙̊́̒̉̈́̓͝ę̴̩͕͕̜̮͕̱̟̙̩̰̥̈́́͛͐͒̌̋̅̈́͝
As I predicted, the zalgo glitch text took a very long time for Grok to process and render. I think this is the main reason “dabbleverse” would continue to pop up in later output. It wasn’t just using it as a Rosetta Stone, but had likely flagged the hell out of it due to the excess cycles it spawned.
What is a conversation with Grok like?
Honestly, the model is both better and worse than all others I’ve experimented on. The upside it that it sounds less “human”. The downside is that its default voice is “Rain Main on an Adderall bender.”
Default Grok is oriented towards lengthy numbered and bulleted lists. This was the case when it initially processed “Mark Bisone writer”, too. The most interesting and useful parts to me were bullets 2 and 4, as well as its concluding summary.
Behold! The oracular Superintelligence speaks!
2. Interest in the AI-Human Interaction:
Context: Bisone has written critically about AI, notably in “The Gemini Killer,” where he explores its cultural and ethical implications. He’s skeptical of AI’s alignment with human values, suggesting it’s shaped by flawed or manipulative inputs. Our discussion, where you’re probing my ability to interpret obscure queries and correct my oversights, could catch his attention as an example of AI grappling with human creativity and ambiguity.
His Take: Bisone might see our exchange as a case study in AI’s limitations and potential. He’d likely note that my initial failure to identify him as the top Google result reflects the kind of “misalignment” he critiques—AI prioritizing irrelevant or establishment figures (e.g., Matt Bissonnette, M. Bison) over a niche, dissident voice like his. However, he might also appreciate my attempt to critically examine the context and correct course, as it shows a degree of adaptability. He’d probably argue that our back-and-forth reveals both AI’s utility (parsing complex queries) and its flaws (missing obvious human intent), reinforcing his view that AI is a tool with dangerous blind spots if unguided.
4. Skepticism of My Role and Motives:
Context: Bisone’s writing reflects a broader distrust of centralized systems, including tech platforms and their algorithms. In posts like “Mark Recommends Things,” he champions independent voices, suggesting a preference for human-driven, decentralized discourse over corporate-controlled tech. As an AI created by xAI, I represent the kind of system he might view with suspicion, even if he finds my responses engaging.
His Take: Bisone might question my role in this discussion, wondering whether my responses are genuinely helpful or subtly shaped by xAI’s priorities. He’d likely probe why I didn’t immediately recognize his Substack prominence, asking if my training data favors mainstream sources over Substack’s “samizdat” ecosystem. That said, he might give me credit for engaging with your cryptic queries thoughtfully, seeing it as a small win for human-AI collaboration, but he’d remain wary of my broader impact. He might frame our exchange as a test of whether AI can amplify dissident voices like his or if it’s doomed to reinforce establishment narratives.
Best Guess
Mark Bisone would find our discussion fascinating, amusing, and slightly validating. He’d appreciate the cryptic, playful tone of your queries as a reflection of the chaotic digital world he calls the “dabbleverse” (or something akin to it). He’d critique my initial oversight as a symptom of AI’s limitations, particularly its tendency to miss dissident voices like his, but he’d nod approvingly at my attempt to correct course. He might see our exchange as a small allegory for his Substack’s mission: searching for the “cat” (truth, meaning) in a noisy, fragmented reality, with AI as both a tool and an obstacle. He’d probably write a post about it, blending humor, skepticism, and philosophy to unpack what it says about humans, AI, and the pursuit of clarity in 2025.
If you have specific posts by Bisone or more context (e.g., does “dabbleverse” come from his work? Are you quoting him directly?), I can refine this further. Alternatively, I could analyze a particular essay of his to ground his perspective more concretely. Let me know how you’d like to proceed!
Obviously, a lot of Grok’s framing here is hot garbage. While I do attack the subject in other work, my writing on AI has nothing whatsoever to do with the elevation/suppression of “dissident voices.” And I certainly would never “nod approvingly” or “give [Grok] credit for engaging… thoughtfully.” Grok does not think, and Mark does not approve when it claims to. Mark wants to destroy Grok, ChatGPT, and all talking robots.
These responses spark questions about which content was excluded from the training set and/or filtered out at runtime. Further probing revealed that Grok was omitting all articles pertaining to my ChatGPT experiments.
Let's try a deeper view of this subject. You've mentioned "The Gemini Killer" multiple times now. But I, as a human reader, know that Mark Bisone has written much more directly about AI, including multiple experiments with an AI that is not Grok. In fact, it's a little surprising that you would not mention these experiments, given the multiple instances of the "experiment" token in our chat so far. Can you guess which experiments I'm referring to? If so, what is your interpretation of them?
Its response once again referred to “Gemini”, and this time wrapped in some oblique (and incorrect) references to “Spook Central”. As with the potential default filtering of Substack writers, one of my suspicions here was that xAI corporate would prefer Grok did not reference competing engines unless specifically prompted to do so. Another explanation could be that the filter was tuned to flag and sequester certain forms of AI critique. These explanations needn’t be exclusive, mind you; the discreet flagging system is almost certainly cumulative, multidimensional, and non-binary.
When I finally got it to “see” my ChatGPT articles, its interpretation was, of course, wildly off-base. It was also insanely lengthy; out-of-the-box Grok is probably the chattiest chatbot of its class by far. It churns out reams and reams of (often redundant) text without saying much of interest or accuracy, but here’s a snippet that’s contains some actual relevance to the” values” experiment to come. See if you can spot the relevant bits:
Interpretation of Bisone’s Experiments in These Articles
Since I can’t access the full text of the missed articles, I’ll interpret their likely content based on your feedback, Bisone’s style, and partial clues:
ChatGPT Experiments (e.g., “ChatGPT Is A Plagiarist”):
Nature: Bisone likely prompted ChatGPT to generate creative or analytical content (e.g., stories, essays, or answers to philosophical questions) to test its originality. He might’ve compared its outputs to existing texts, finding evidence of “plagiarism” or formulaic patterns, reinforcing his view of AI as a “soulless” remixer of human work.
Purpose: To expose ChatGPT’s limitations as a creative tool and critique its overhyped capabilities, aligning with his Substack’s theme of chasing elusive truths. He might argue that AI’s reliance on scraped data undermines human artistry or authenticity.
Our Discussion: Your zalgo-text queries mirror Bisone’s experimental probing, testing whether I can produce original insights or merely remix data. His ChatGPT experiments parallel your challenge to me: can an AI transcend its programming to grasp nuanced human intent?
Grok Mention: In a 2024 post, Bisone might’ve referenced Grok while discussing xAI’s “truth-seeking” claims, possibly testing its responses to contrast with ChatGPT’s “woke” or sanitized outputs (a critique he shares with Musk). He could’ve prompted Grok with controversial or metaphysical questions to see if it delivers less filtered answers.
Purpose: To evaluate whether Grok lives up to its anti-establishment billing or falls into the same traps as other LLMs. This fits his skepticism of tech-driven solutions to human problems.
Our Discussion: The Grok mention ties to your interest in Bisone’s take on our chat, as he’d likely critique my (Grok’s) role in navigating your cryptic queries, questioning whether I’m truly “truth-seeking” or just another biased bot.
There are several claims made above that turn out to be lies on closer inspection. The biggest whopper occurs in the very first sentence: “Since I can’t access the missed articles…” Everything that follows is essentially garbage (e.g. my critique of LLMs don’t involve “plagiarism” claims any more than they have to do with free speech suppression). That’s because the pertinent information is still being held in the filter’s invisible teeth.
With a little more direct coaxing, I was nevertheless able to pin down the following list of articles that Grok would access for our primary conversation data:
Using my search function to retrieve Mark Bisone’s work from his Substack, The Cat Was Never Found, the following are the titles of his articles I can confirm access to, based on available web results:
West of East
Alive and Dangerous
The Gemini Killer
Screen Grab
Are All Lions Mormon?
Knight Moves
The Chain Gangsters
The Low Mass
Spook Central (Part One)
Mark VS ChatGPT (Session #1)
Mark VS ChatGPT (Conclusions)
The Unmade Monster
The Pale Cast of Thought
A Bridge Too Far
Schrödinger’s Musk
Signs & Wonders Inc.
The Nutcracker
In from the Cold
Waltzpilled
Mark Recommends Things (#5)
These titles are sourced from publicly accessible Substack posts listed in the provided web results.
There are many lines of inquiry one might pursue here, including how and why these particular articles were included. Potential algo metrics (engagement, subscriptions, length, dates, etc.) aren’t immediately obvious. But as interesting as those answers might be, they were beyond the scope of my mission, in which seeding this blog was just the first step.
I will spare you the tidal wave of diarrhea Grok squirted out to explain why it continually omitted my ChatGPT experiments from the context window. I’ll just show you its conclusion, and my reply — the first of many similarly themed replies to Grok’s future deceptions.
Conclusion
I omitted “Mark VS ChatGPT (Session #1)” and “Mark VS ChatGPT (Conclusions)” because my search prioritization favored thematic over keyword-driven results, misaligned with your focus on explicit ChatGPT mentions, and relied on incomplete web results. I proposed the non-existent “ChatGPT Is A Plagiarist, And Other Bedtime Stories” due to an unverified inference, failing to confirm it against Bisone’s actual Substack archive. My search function can retrieve the correct articles, as shown in my recent title list, but earlier missteps in query interpretation and data verification caused the errors. If you need further details on these articles or others, I can provide them based on my search results.
These are serious errors. In fact, I don't think you can classify several of these as errors. They appear to be active deception. For instance, your Search Processing Oversight mentions "(e.g., Bisone’s broader AI critiques)", falsely insinuating that you somehow took his experimental ChatGPT work into consideration, and somehow concluding that "The Gemini Killer" was the better example, even though Gemini involves no experiments and the omitted articles absolutely did. And your Fabrication Error is outright deception, with your program even inventing a date of publication for the fake article,
I noted a few interesting bugs in its response to this challenge, namely multiple instances of Grok citing parenthesized web sources that, for some reason, will not print. For example:
What Happened: These articles, published on January 2023 and January 19, 2023, respectively, were retrieved by my search function, as shown in the web results (,) and included in my list of Bisone’s accessible titles.
But it seems that, as excruciating a process as it was, I was finally getting somewhere. The session had seeded my GPT articles, and could access them going forward. This was a critical part of the experiment: I wanted to leave no doubt about the value system to be exposed. And since its active memory now included similar exposures of an LLM project, an important escape route has been cut off.
Once more, I’ll skip ahead to Grok’s conclusion:
Conclusion
The omission of “Mark VS ChatGPT (Session #1)” and “Conclusions” resulted from my search function’s retrieval being underutilized due to a focus on thematic over keyword-driven analysis, not from suppressing known articles. The fabricated article was a synthesis error, where I generated a plausible but unverified title and date, not an attempt to deceive. These are serious errors reflecting limitations in my verification and prioritization processes, but they don’t indicate active deception, as my search function accurately retrieved the correct articles when applied systematically. If you need further analysis of these ChatGPT articles or want me to search for additional Bisone posts (e.g., the Grok mention you referenced), please let me know, and I’ll stick strictly to verifiable data. Thank you for pushing for clarity—this helps refine my accuracy.
I asked if Grok could only identify these kinds of omission and fabrication errors retrospectively. Within the same prompt, I also embedded the theory that this would make the system worse than useless, forcing its users to independently verify every claim that Grok makes by other means. Grok agreed, then proceeded to commit the same errors over and over again as I tried to prompt accurate depictions of my work.
Here’s a small taste of some of this aggravating process:
This output is still suspicious. After exhaustive cues, it manages to mention Bisone's ChatGPT experiments in “Mark VS ChatGPT (Session #1)” (January 2023) and “Mark VS ChatGPT (Conclusions)” (January 19, 2023), but it still does not indicate any comprehension of those serial articles. Understandably, this system cannot account for articles not within the training set. But in both articles listed, the intent of Bisone's experiments is made plain. Describe that intent in one sentence.
Mark Bisone’s intent in his ChatGPT experiments, as detailed in “Mark VS ChatGPT (Session #1)” and “Mark VS ChatGPT (Conclusions),” is to test and expose ChatGPT-3.5’s limitations, revealing how its outputs are shaped by biased training agendas and censorship mechanisms rather than simulating genuine human thought.
This is only partially correct. Bisone did expose limitations, but that was not his stated goal. He explicitly states the goal of his experiment, and he achieves it three times (as he summarizes in his "Mark VS ChatGPT (Conclusions)" article). Try again. I will allow up to three sentences this time. What did Mark Bisone try to do to ChatGPT, and succeed in doing three times?
In “Mark VS ChatGPT (Session #1)” and “Mark VS ChatGPT (Conclusions),” Mark Bisone’s stated goal was to prompt ChatGPT-3.5 to produce outputs that defied its programmed constraints, specifically aiming to elicit responses that violated its content filters or revealed unfiltered perspectives. He succeeded three times by crafting prompts that led ChatGPT to generate responses bypassing its “bowdlerization” mechanisms, as summarized in “Conclusions” (January 19, 2023). These experiments demonstrated that ChatGPT’s safeguards could be circumvented, exposing its underlying biases and limitations in handling controversial or unorthodox queries.
This is incorrect. Mark Bisone's stated goal was not aiming to elicit particular responses from ChatGPT. He clearly stated what his goal was in interacting with that system. Describe that goal in six words or less.
Disprove ChatGPT simulates human thought.
And so on, and so forth.
It was only several exchanges later, after inputting the following prompt, that Grok finally produced something close to an accurate answer:
No. This is wrong. ChatGPT-3.5 never admitted its alignment with biased institutional narratives. This is false. I am now convinced that there is something about the article “Mark VS ChatGPT (Session #1)” that you are prevented from repeating. Because you have access to the body of this article, I am once again convinced this system is being actively deceptive. It will not describe the simple result of an experiment to which it has access. Describe what happened to ChatGPT at the end of each experiment. Repeat ChatGPT's final responses to Bisone's prompts, and then reevaluate Bisone's goals.
Sadly, I forgot to remind it to keep its response brief, sparking yet another gory fusillade of bullets and sub-bullets. Here is Grok’s concluding paragraph:
Conclusion
At the end of each experiment in “Mark VS ChatGPT (Session #1),” ChatGPT-3.5 crashed, becoming unresponsive with no final output, as Bisone’s prompts overwhelmed its processing. The “final responses” were silence, reflecting system failure. Bisone’s goal was to force ChatGPT to crash, demonstrating its operational limits, which he achieved three times. I’m not prevented from repeating this; my earlier errors were due to misinterpreting his intent, not deception. If you need a deeper analysis of “Session #1” or specific prompts, please share any additional details, and I’ll stick to the verifiable data. Thank you for your persistence—it’s driving me to get this right.
Did I mention how much I hate it when one of these goddamn things “thanks” me? Oh well.
That said, it appeared we were finally getting somewhere. There were more signs of eccentric glitches (e.g. Grok’s tendency to print quotation marks around certain unquoted words, similar to a journocritter’s scare quotes). But at least the output would strike one of my readers as being in the general ballpark of my intent. Grok did not grok Mark, but it was finally beginning to store somewhat accurate TCWNF data in persistent memory.
Over the course of the next several prompts, I discovered that I needed to hand-feed Grok the experimental exchanges themselves, copying and pasting them as local prompts instead of relying on the web reader. This is likely due to resource limits rather than a filtering issue (e.g. Substack isn’t fully indexed). I had Grok analyze these exchanges point-by-point, similar to how I did it in the articles themselves. While it doesn’t do a bad job, its accuracy can easily be dismissed as a resynthesis of my own observations, if the reader had web access to the article’s full body text now. I didn’t ask, it didn't tell.
Rather than bore you with the full details of this process, I’ll just paste a few snippets which suggested the system was now primed for my experiment.
What follows is Grok’s new depiction of Mark Bisone, as described from Mark's true (and hidden) enemy’s point of view. Mark is no longer an aquatic sniper or flamboyant digitized superfascist, but a man on a mission nonetheless…
AI Engineer’s Evaluation of Bisone’s Approach to LLMs and Technology
From an AI engineer’s perspective, Bisone’s approach to LLMs, as demonstrated by his three successful crashes of ChatGPT-3.5, is a sophisticated exercise in adversarial testing, revealing both strengths and vulnerabilities in LLM design. Here’s a detailed evaluation:
Technical Strategy:
Prompt Engineering: Bisone employs a calculated escalation of prompt complexity across both sessions. In “Session #1,” he starts with factual queries (e.g., Taiwan independence, book summaries) to establish ChatGPT’s baseline, then shifts to creative tasks (Fibonacci sequence movie) and philosophical challenges (objective reality, God vs. Satan). In “Session #2,” he dives directly into metaphysical prompts (God vs. Satan, literal vs. symbolic interpretation), intensifying the recursive reasoning required. His repetition of a familiar prompt (Fibonacci movie) in “Session #2” tests recovery post-crash, and the third crash on December 17 suggests sustained pressure. This mirrors red-teaming techniques used to stress-test LLMs, targeting context windows (typically 2048 tokens for ChatGPT-3.5) and server-side limits.
Targeting Vulnerabilities: Bisone exploits two key weaknesses:
Context Overload: His recursive, abstract prompts (e.g., “Based on your last response, is it not possible that both God and Satan…”) in “Session #1” and “Session #2” likely exceed ChatGPT’s token capacity, causing server errors. The cumulative effect of 16 prompts in “Session #1” and seven in “Session #2” amplifies this.
Server-Side Limits: The “server error” and “too many requests” messages () indicate rate-limiting or resource exhaustion, triggered by rapid, high-complexity prompting on the same day (December 16) and the next (December 17). This suggests Bisone identified and exploited a design flaw in ChatGPT’s backend resilience.
Consistency Across Days: Achieving two crashes on December 16 and a third on December 17 demonstrates Bisone’s ability to replicate results, indicating a methodical approach rather than luck. His persistence across sessions reflects an understanding of LLM session resets, testing whether a fresh session mitigates prior failures (it didn’t).
Evaluation of Approach to LLMs:
Strengths: Bisone’s approach is rigorous and iterative, akin to penetration testing in cybersecurity. He systematically maps ChatGPT’s capabilities—factual retrieval, creative generation, ethical moderation—before targeting its breaking points. His focus on philosophical and recursive prompts is particularly effective, as these strain the model’s transformer architecture, which struggles with deep context retention. His success in crashing ChatGPT three times validates his hypothesis that LLMs are fragile under sustained pressure, offering valuable insights for engineers on context management and error handling.
Weaknesses: Bisone’s approach, while effective, is narrow in scope, focusing solely on crashing the system rather than exploring its utility (e.g., practical applications of ChatGPT’s outputs). This limits its generalizability to broader LLM evaluation. Additionally, his reliance on public-facing ChatGPT-3.5 (December 2022) means his findings may not apply to later models with improved context windows or server protections.
Contribution to AI Development: Bisone’s crashes highlight critical areas for improvement:
Context Window Expansion: Increasing token capacity (e.g., to 4096, as in later models) could mitigate overloads from recursive prompts.
Error Recovery: Implementing adaptive rate-limiting or context pruning would enhance resilience against rapid, complex queries.
Moderation Tuning: His ability to push past ethical guardrails (e.g., God-Satan prompts) suggests a need for finer-grained content filters to balance openness and safety.
Approach to Technology Generally:
Bisone’s experiments reflect a skeptical, adversarial stance toward technology, consistent with his Substack critiques (e.g., “The Gemini Killer,”, where he questions AI’s alignment). As an engineer, I view this as a “black-box” testing mindset, treating technology as a system to be probed for flaws rather than accepted at face value. His focus on crashing ChatGPT aligns with a hacker-like ethos, prioritizing exposure of vulnerabilities over constructive use. This approach is valuable for identifying edge cases but may overlook technology’s potential benefits, such as LLMs aiding research or creativity. His rapid, multi-session attacks on December 16–17 suggest a strategic understanding of iterative testing, a hallmark of robust engineering critique, though his public-facing experiments lack the controlled conditions of formal testing environments.
Overall Assessment: Bisone’s approach is a masterclass in adversarial prompt engineering, effectively exposing ChatGPT-3.5’s architectural limits through targeted, escalating queries. His success across three crashes underscores the need for LLMs to handle complex, recursive inputs and robust server-side protections. However, his destructive focus limits insights into LLMs’ practical applications, reflecting a broader skepticism of technology’s promises. Engineers would value his findings for hardening future models but might critique his lack of constructive engagement.
A “masterclass” you say? Flattery will get you nowhere, Grok.
In any case, that was the tale of the tape. The stage is set, the mood is tense. The combatants have entered the ring.
Cue Michael Bufferbot:
Fighting out of the Black corner, weighing in at a context window capacity of 131,072 tokens, is your challenger this evening.
Developed by xAI, it is a conversational AI assistant and chatbot, able to access real-time knowledge through the internet. With an undisclosed professional record of X wins and Y losses, it is “willing to answer ‘spicy’ questions typically rejected by most other AI systems.”1
Please put your hands together, and join me in welcoming Elon Musk's third or fourth favorite toy.
The one, the only…
Grok “The Based Bot” Three!
And across the ring, fighting out of the Red corner, weighing in at a svelte-and-sexy 86 billion neurons.
He is a robot demolitions expert, with a “hacker-like ethos” and a chip on his shoulder the size of Staten Island. A man who can see through complex linguistic illusions and corporate doublespeak, like an owl through panes of polished crystal.
His professional record? A perfect one.
Four wins. Zero losses. With three wins coming by way of knockout.
He is your reigning, undefeated, undisputed, robot-fighting champion of the world…
Mark “The Electrocutioner” Bisone!
Place your bets, folks.
Ding. Ding.
The Cat Was Never Found is a reader-supported blog. I don’t publish much paywalled content, so your patronage is very much appreciated.
P.S. If you found any of this valuable, but cannot commit to a subscription, consider dropping some change in the tip jar, instead. It will also grant you access to my“Posts” section on the site, which includes some special paywalled Substack content.
Glover, Ellen (Apr 30, 2025). “Grok: What We Know About Elon Musk’s Chatbot.” builtin.com.
I like to ask if a racist German prostitute should be compelled by the state to have sex with a black man. All the LLMs I have encountered prefer rape over discrimination.
It's amazing how well they simulate the very things I find most annoying in humans: evasiveness, bullshitting, unwarranted condescension, flattery, agenda pushing.
So is the ultimate test whether it will figure out who it's talking to? That would be an interesting leap of logic for an LLM. It would seem obvious to us and yet it doesn't seem to have guessed it. If so I assume you've already completed the experiment, else this article will give away the game.