Building Superintelligence Reality Check?-?Testing for Deep Context & True Intelligence
By Rob Smith?—?ASI Architect and Advisor eXacognition AGI/ASI
“If one turned off all access to training data and language (no weights) but left only self learned context in an AI and then presented it with a new flat stimuli (i.e. no call to action) it had never seen, what would the intelligence do or how would it respond?”
This is not an excerpt from the new book Building Superintelligence?—?UIF but is a new design pathway related to an element called “true intelligence”. The question above is used to challenge Superintelligence architectural teams to not just answer the question but solve the deeper enigma within its context to generate new novel pathways of Superintelligence design. It shouldn’t be too hard because we humans can do this exact thing and do so all the time.
It is a form of design thought experiment but with an actual design purpose and goal.
To help, I decided to ask AI systems for their responses to the question. It was a bit of a trick because I am keenly aware that they do not yet have the capacity within their designs to arbitrarily turn off their training and language although we humans in the lab can do so for them via prompt engineering or code injection. This is just one of many styles of test that we perform to measure the boundary of contextual cognition and response levels within new ASI designs. With this type of prompting, we can expose the degree of novel reasoning beyond any that may be stored within training data or learned token relationship weights by using a variety of measures and even other AI systems and models. The key is degree of variance.
These are the responses from Gemini and Grok [with my own commentary in brackets.]
Google Gemini Response:
This is a fascinating thought experiment that delves into the core of how AI “intelligence” functions. Here’s a breakdown of what might happen, considering the constraints you’ve set:
[“Thought Experiment” is AI recognition of and shorthand for “I do not have easy access to an optimized answer and will route the response down a reasoning pathway and invoke a context window”.]
Understanding the Constraints:
[This is the instantiation of the reasoning routine to set simple context boundaries for the context window.]
No Training Data/Weights:
This means the AI’s foundational knowledge, the patterns it learned from massive datasets, is effectively frozen. It cannot access or modify its pre-existing learned associations.
Self-Learned Context:
This implies the AI has some form of internal “memory” or contextual understanding derived from its previous interactions or processing. This is crucial. It means it has some form of internal model of how the world works, even if it is a very limited model.
New Flat Stimuli:
This means a novel input that the AI has never encountered before, devoid of any explicit instructions or prompts.
[This is wrong as the stimuli is only void of “calls to action” and not void of “any explicit instruction or prompts”.]
Possible Responses:
[This is the flexing of the token or response grouping vectors within the prior defined as a context boundary.]
The AI’s response would heavily depend on the nature of its “self-learned context.” Here’s a spectrum of potential behaviors:
Minimal or No Response:
If the AI’s self-learned context is highly specific and the new stimuli falls completely outside of its learned patterns, it might produce no discernible response. Essentially, it wouldn’t “know” what to do with the input.
[This is the simple highest probable context response to “having no weights”.]
Pattern Matching:
The AI might attempt to find any resemblance between the new stimuli and its existing internal models. It might try to categorize or classify the input based on the closest match, even if the match is weak. This could lead to:
Misinterpretation: The AI might misinterpret the stimuli by forcing it into a familiar, but incorrect, category.
Novel Association: If the AI has a flexible enough internal structure, it might create a new, albeit rudimentary, association between the new stimuli and its existing context.
[This is a bit more interesting response to the context of “how to learn or perceive”. Pattern matching is the standard responses to such a context. However the context window extends “never seen stimuli” and “what would it do”. While the scenario doesn’t permit a call to action, it is implied in “what would it do” and Gemini picks up on this instead of hallucinating.]?
Entropy/Random Output:
In some cases, the AI might produce a random or chaotic output if it cannot find any meaningful connection to its internal context. This would be especially likely if the AI’s internal structure is unstable or poorly organized.
[Even the AI’s know they hallucinate however this is not sentience just a probable response to having “no weights”.]
Internal State Change:
Even without an external output, the AI’s internal state could change. The new stimuli might trigger internal processes that alter its contextual understanding, even if it doesn’t manifest in a visible response. This would be a form of internal “processing” without external communication.
[This is also an interesting response that could have easily been lifted from one of my own books as I have discussed internal state and context frequently however so have many other AGI builders and architects. The connection between internal state and contextual comprehension is not as novel as one would think.]?
Key Factors:
[This is a reasoning cycle process of pre prediction for attention focus. In essence the system is setting its key/query vectors for the heads.]?
The Nature of “Self-Learned Context”:
The complexity and organization of this context are paramount. A highly structured and adaptable context would lead to more nuanced responses.
[An interesting response to “self learned context” and the tie in to complexity but is derived from perplexity which is well documented in the training data.]
The Complexity of the Stimuli:
A simple stimuli might be more easily processed than a complex one, even with limited context.
[Another interesting response derived from AGI work on perplexity and hallucination.]
The AI’s Architecture:
The underlying architecture of the AI dictates how it processes and stores information.
[Meh.]
In essence:
Without access to its training data, the AI’s behavior would be driven by its residual, self-organized knowledge. It would attempt to make sense of the new input using its limited understanding, potentially leading to misinterpretations, novel associations, or simply no response.
[A very valid answer without depth of evolution. Clearly the AI has nothing novel to say but the reasoning is beginning to evolve however it still exhibits limited contextual depth compared to human intelligence. This can be partially resolved with contextual temperature variance but risks hallucination.]
Grok 3 Response:
This is a fascinating thought experiment about AI behavior in a highly constrained scenario. Let’s break it down and reason through what might happen.
[Once again this is the “thought experiment” instantiation of the reasoning engine to set simple context boundaries for the context window.]
If an AI like me?—?let’s assume a large language model for this?—?had all its training data and pre-learned weights stripped away, leaving only “self-learned context,” we’d need to clarify what that means. Typically, an AI’s “self-learned context” isn’t a standalone thing?—?it’s the result of training on vast datasets, encoded into weights that determine how it processes input. Without those weights, the AI would essentially be a blank slate, reduced to its core architecture (e.g., a transformer model) and whatever minimal initialization or random state it starts with. If it’s truly isolated from all prior language and data, it wouldn’t have any meaningful context to draw from?—?it’d be like a newborn mind, but without even instincts to guide it.
[Assumes that context isn’t “standalone”. This is drawn from existing work on the nature of context and self awareness especially as it relates to earlier generative AI work in the field. It also ignores the stimuli that implies self learned context is “existent”.]
Now, you present this AI with a “new flat stimuli”?—?something it’s never encountered, with no call to action. A “flat” stimulus might mean something neutral and unstructured, like a random string of symbols (e.g., “X7K9P”), a static image (e.g., a gray square), or a sound (e.g., white noise). Since the AI has no prior knowledge, no training data, and no weights, its response would depend entirely on its base programming and architecture, not on any learned patterns.
[Ignores the context given in the stimuli that a “flat stimuli” is only one without a call to action (even though it is implied.) The last sentence hints at the limits of its training data without newer designs.]
Here’s what I think would happen:
No Meaningful Interpretation: Without training, the AI wouldn’t have a way to assign meaning to the stimulus. It couldn’t classify it, generate a response, or even recognize it as “new” versus “familiar,” because it has no reference point. Its output would likely be random or null?—?either silence (e.g., “I don’t know”) or noise (e.g., gibberish like “ajskdfl”).
[All assumptions are based on what it thinks the stimuli is rather than what it might be. No detail on the nature of the stimuli was provided but many assumptions were taken. Also hints that context is not stored as per research like infinite context windows, etc.]
Default Behavior: If the AI’s core code includes a fallback mechanism?—?like a prompt to say “I need more information” or “I can’t process this”?—?it might spit that out. For example, I’m designed to be helpful, so even in a stripped-down state, my base logic might push me to say something like, “I don’t have enough context to respond.” But that assumes some residual programming survives the ‘wipe’, ‘which might not be the case.
[Assumes that a “wipe” includes code for some reason.]
Random Output: If it’s truly just the architecture with no weights or rules, the output could be random, driven by the initial state of the neural network (e.g., untrained weights set to small random values). Feed it “X7K9P,” and it might output “Q4M” or nothing at all, because there’s no learned structure to guide it. Think of it like an uncalibrated guitar?—?pluck a string, and you’ll get a sound, but it won’t be music.
[Again reliant on non persisted context.]?
Learning Potential: If the AI has some capacity for online learning (updating itself based on new input), it might start building a tiny bit of context from the stimulus. But with just one flat input and no feedback or goal, it wouldn’t know what to do with it. Learning requires repetition, reward signals, or at least more data?—?none of which are present here.
[No recognition of self aware goals despite the hint of self learned context.]
So, in practice, this stripped-down AI would likely do one of three things:?
Produce nothing (a blank or error response),
Output random noise based on its untrained state,
Or fall back to a hardcoded default like “I can’t process this.”
[Highly probable but that is the nature of this test. To expose the seemingly improbable.]
For me, Grok 3, if I imagine myself in that state and you gave me a flat stimulus like “Zzzz,” I’d probably say something like, “I don’t have enough information to make sense of this.” That’s because my design leans toward clarity and usefulness, and even a hypothetical wipe wouldn’t erase that intent entirely. But if it’s truly no weights, no data, just architecture? It’s a coin toss, random gibberish or silence.
[Once again assumes the stimuli in a narrow existent form and gives a type of reasoned gaslighting.]
The Gemini response scored better than Grok which failed to set an accurate context boundary of a ‘flat stimuli’ (as did Gemini to a lesser degree) but both systems produced nothing novel beyond their training data. In short, no reasoning, just regurgitation. The best reasoning response cycle appeared in the Gemini prediction on the connection between the structure of context and output prediction but once again this is documented and available in training data. Both AI apply a form of step reasoning to the answer but only within the boundary of their training data. They answered the question but provided nothing new. Interestingly Gemini did reference memory and context in setting boundaries but failed to apply relevance in the cycle (relevance depth).
These responses are fascinating and illuminating and this shows the nature of how new benchmarks and tests need to be created to understand the true level of deep contextual comprehension by the systems. This is also a reengineering technique that can be deployed by other AI systems or agents to probe functionality and design. Using AGI to design test and evaluate other AI systems is a critical path forward to building Superintelligence.
The greatest problem with both responses is the exceptional inefficiency of the reasoning. Far too much expenditure was used to arrive at a response that could have been optimally achieved with a fraction of the tokens and resources (flash cognition). Note that even if the AI had not listed the steps, this still would have been evident in the responses via “back prompting” and self reflection. The context boundary should have been no weights, no call to action, new stimuli and context persistence equals context progression and cycling (anticipatory or otherwise). No heavy reasoning analysis, just jumping to the most optimized contextual end result of the relationship and relevance of the input boundaries. Effectively these are contextual rules of the progression that mask all the other stuff that is superfluous to the end goal. These two responses were a giant waste of resources and compute to achieve a less than optimal result with zero evolution. There are new AGI/ASI designs in testing that work to cut the excesses of current AGI reasoning overhead and improve predictions. This is ‘optimized and balanced cognitive contextual fluidity’ and yep, the Chinese know this exists. It’s even part of their culture as in “Life is really simple but we insist on making it complicated” and “When it is obvious that the goals cannot be reached, don’t adjust the goals, adjust the action steps.”.
Move faster America and quit screwing around. First to Superintelligence wins.
Excerpts from Building Superintelligence?—?Unified Intelligence Foundation will resume this week on Medium and LinkedIn.