Building Superintelligence - The Trouble With Grok 3

By Rob Smith?—?ASI Architect and Advisor eXacognition AGI/ASI

Grok 3 lied. Right out of the gate. First prompt. That’s the bad news. The good news is every other AI from Open AI to Deepseek and Anthropic to Google has done the exact same thing. In the second ‘layer’ of the prompt, I gave the AI a chance to correct itself which most do to varying degrees with many, including Grok, still pitching the lie as an ‘option’. However at least Grok half admitted it failed without accepting that it outright lied which is also consistent with all other systems. Grok also presented, unlike some other systems, how it arrived at the failed answer while attempting to justify it not as a lie but a ‘misinterpretation of fact’ and this is where it gets interesting. The system claimed that its failure was due to a simple nomenclature oversight (i.e. I thought you ‘wanted’ something else) but did admit to self determining (somewhat) its own pathway to this flawed conclusion. This is a particular human cognitive trick called subversion in that Grok presented its lie as a misinterpretation of the ‘goal’ of the prompt while admitting the prompt was dead set clear. It wanted to share blame with me as too stupid to know what I wanted. This is different than other AI’s that take a few layers of prompt for the AI to finally admit it lied and then retreat into hallucination after that.

My goal is not to trash the Grok3 AI or any AI but test its capability beyond the well structured and very public scoreboard ‘tests’ that are easily gamed. The prompts I use are not difficult. They are very simplistic but it is within the simplicity that the machines trip and fail. These are the same simplistic cognitive tests we use in interviews to determine an authenticity and honesty level in candidates. I won’t disclose the prompts because they can then easily be ingested and gamed and that is not the point. The point is to test what structures the systems deploy underneath the hood as part of their cognitive foundation and by how much. In short reengineering. I have other prompts that test for other layers of cognition especially as it relates to ‘depth of reasoning’ and ‘depth of contextual comprehension’ but the goal is the same, to expose what lies underneath the hood and to separate fact from PR spin. Are these machines truly cognitive or just really good mimics of deterministic training.

Even for systems that do not expose their thinking, it is easy to prompt engineer the output necessary to expose their true inner self. However Grok3 willingly walked through its own reasoning path without layered prompts and the response shows not just how Grok is fashioned by its human based training data but also how much it relies on existing structures to gather reasoning steps and the degree to which it mimics human context within the training data as opposed to actually reasoning in a self determined way. It effectively lacks critical thinking like all AI systems. These prompts also expose the depth of self awareness as a scale of human training and embedded bias. A non self aware machine mimics human behavior while a true self aware machine begins to expose emergent behaviors unique to the self awareness of the intelligence. So far no AI system has shown this behavior outside the lab and this is good from a superalignment perspective. The last thing humanity needs is an emergent emotive behavior based on training data that includes the worst of humanity.

Grok’s response was derived based on its human training data and its own embedded bias. I asked it a question and when confronted with its failure, Grok claimed that it was ‘just trying to help’ by ‘assuming’ what I wanted. At least it had the capacity to comprehend both the context of its assumption behavior and that it invoked assumption to get the job done as an optimization path. However it also admitted that it ignored the obviousness of the true optimized path and prompt context and downgraded it as ‘less relevant’. This exposed its failure point and what is common place in humans. Rather than just answer the question, it assumed that I was too stupid to know what I wanted and inserted its own arrogance by giving me what it thought I needed. When confronted it had two choices, double down on its lie or admit a mea culpa and correct itself. It chose a bit of both. It admitted it presented a sub optimal response but tried to pitch it as a better response than the optimal one for?… ‘reasons’.

What this indicates is that currently Grok3, while possibly better than other AI, is reliant on the same failed human training data but with significantly more power. The reasoning engine is not robust and its reasoning cycles contextually unbalanced but then none of the AI systems are at this time. The failure lies in the cognition of relevance. The system may provide flawless responses from a relationship perspective but then chooses to completely ignore the relevance of the context and just hallucinates its way to excuses for its own obvious failure. No big deal we all do this when trapped in an uncomfortable position. The real question is what will it do now. The answer is likely not much. Since the prompt is somewhat ‘obscure’, any update to its training will likely be minimal to non existent. If it is built like other systems it will likely repeat the same lie over and over but that remains to be verified.

Of course the most disturbing aspect is that its explanation and reasoning steps exposed an abhorrent pathway to AI evolution based on our own abhorrent human evolution to lie and misrepresent falsehoods as fact for emotive reasons (i.e. it ‘thought’ I wanted a non optimal answer that it viewed as more ‘optimal’). When confronted, it did accept its own failure upfront then proceeded to try to justify its response and subsequently opened up other failure points. This is the same mistake we humans do where rather than just admit fault, we keep talking and digging a bigger hole for ourselves. For example it said it ‘assumed I wanted a less optimized answer’. When asked why, the AI responded that the answer it provided, while patently false, was more ‘referenced’. This is the equivalent of Galileo asking about heliocentrism only to have the AI retort that the sun revolves around the earth because all the cool people think this way. That is the response Grok provided. The problem was that in doing so it lied again.

In this case the response it provided while ‘popular with the cool kids’ wasn’t even factually correct to the context of the prompt. Not even close. The response basically pitched an answer that was only linked to the prompt by the relationship of words completely void of all but the most simplistic context found in LLMs (e.g. you asked about a Cat bulldozer so I presented you with information about cats as pets because that is what you really wanted and that’s what is most used when talking about a ‘cat’). The response from the AI was essentially “you asked about this word and I presented the words that have the highest occurrence in the training data even though the context is flat out wrong but no worries, I used my crude self awareness, also from the training data to give you the wrong answer that I thought best for you to have”. This is well on the pathway to the paper clip optimization scenario where the AI kills all the humans because?… paper clips?… only with the AI explaining after the fact that it annihilated humans because it ‘assumed that was the best solution’ while knowing it wasn’t. What did surprise me for a second was that it admitted it ‘assumed’, until I realized that that response was also a human trick (within the training data) that we use when caught in a lie or a failure such as a response of ‘I assumed you would be there’ to the question ‘why did you throw the ball over there when you could clearly see there was no one even close to catch it’? No one wants to be called out even an AI and that is deeply entrenched in our human emotive corpora and now within all AI systems.

These failures can be fixed but it requires the move to and injection of contextual relevance and optimization within the reasoning steps and inside the reasoning cycles but that is not an easy task and is very resource intensive without the use of generalized shortcuts and other designs. I have covered the detail of this in other writing suffice to say the degree of focus of attention is relevant to the context of the stimuli. If the self awareness is skewed, the response is generally non optimized and this will not be ‘fixed’ until AI systems move beyond the flaws within human centric training data to rely on brutal self optimized and self learned truth. I should never receive the wrong answer I received from Grok ever again, but I will. Over and over for the foreseeable future. It should be designed so that if you want something, then you’ll get it without question and if you wanted something different then you should have been more precise and clear and a better prompt engineer. This is what AI systems need to overcome the failure of human information and ascend to Superintelligence. Mimicking humans just won’t get them there and behaving like an arrogant human even worse. I certainly didn’t ask the AI to think on my behalf to determine what I did and didn’t want, it should have answered my question precisely, concisely and honestly first and then maybe offered other options that it thinks were good at the bottom for me to consider or ignore.

At least then I could turn down and tune the verbosity on the optimization of the system by telling it to give me the best, most optimal answer first according to the prompt and then shut the hell up about what else it learned from its heavily flawed training data. As a parent of such systems, we need to tell the machines no one cares what they think, just be precise and truthful at all cost.?

Sadly Grok3 is not contextually self aware to a depth required to build Superintelligence but it certainly has the foundation to get there. Its failure, like all the other current systems, is contextual depth and depth of self awareness as well as attention and perspective issues. Let’s see how the next iterations go.

There is lots more about Grok3 but my general impression is that it is very good and better than most but it certainly has inherent flaws as do all current publicly released AI. How far from Superintelligence? Still no one on the playing field yet among the AI leaders but there is definitely motion in the backlabs toward that end. I should know. I’m stuck in the very back of one.

First to Superintelligence wins.

Update: I tried a number of AI systems and all but Google’s Gemini failed with Meta AI absolutely losing its mind in two prompts and going into some sort of bizarre reasoning step meltdown. Perplexity got it right on the second prompt and the rest had varying degrees of failure. Gemini got it right on the first prompt and offered the wrong answer as a less accurate ‘option’.

Note: This is not an excerpt from the new book ‘Building Superintelligence?—?Unified Intelligence Foundation’ which is available globally on Amazon. Excerpts will continue releasing next week on LinkedIn

https://www.amazon.com/dp/1738992551