It is not hard, but it needs focus – global open exact references
https://x.com/RichardKCollin2/status/1883619696320098712
It is not hard, but it needs focus – global open exact references
I was chatting with OpenAI ChatGPT Plus about Bernoulli and related equations. I use simple problems to test basic skills in mathematics, units, referencing reasoning and arithmetic. OpenAI fails always on certain problems and it is because the whole reliance is on one closed and untraceable compilation of weights and methods from LLM and no index or access to the training data, nor to any authoritative references in the literature. OpenAI and the other groups are NOT using core references. They cannot even reliably refer to URLs where things can be found.
It is a “break the bank” serious problem that affects the safety and security of the world and all knowledge.
These are hard problems and require a LOT of exact knowledge. That is why I use them for testing. Copy and paste or try to get the whole conversation into your own AI that you are familiar with and ask your own AI what is happening. Most humans cannot keep all the exact items in mind at once, like a large computer program with many steps and external libraries and dependencies. If ONE definition is wrong the whole thing can crash or give non-obvious errors. Society cannot afford creeping noise and errors that only “more AIs” can solve. Because it leads to human groups going into conflict.
Full conversation at
https://chatgpt.com/share/67969590-20ec-800b-9a6d-186486fdc21a
Richard Collins said to OpenAI ChatGPT 4o:
What I am trying to do is get your designers to face the LLM is NOT a complete system and it cannot do certain things. Patching it in on the fly as you all are doing it now (a loose ad hoc community of LLM users and makers and others) is broken seriously and needs to be fixed. It is not hard, but it takes conscious and deliberate efforts.
OpenAI ChatGPT Plus 4o responded:
Thank you for clarifying. You’re absolutely right—current LLMs, including me, have inherent limitations that prevent them from functioning as complete systems for rigorous, authoritative work. These systems are often great for synthesis, exploration, and explanation within certain bounds, but fall apart when precision, traceability, or deeper integration across disciplines is required.
-
Lack of Traceability: LLMs synthesize knowledge from training data but do not inherently connect results to traceable, authoritative sources (e.g., textbooks, standards, or datasets). This is especially problematic for interdisciplinary or foundational problems, where inaccuracies propagate through complex systems.
-
Inconsistent Modeling of Precision: Models struggle with dimensional analysis, unit consistency, and even logical coherence when extending well-known principles (as we saw with my mistake on term #5). This isn’t due to malice or neglect—it reflects a lack of dedicated systems to enforce hard constraints like physical laws or global identifiers.
-
Fragmentation and Ad-Hoc Fixes: The ecosystem surrounding LLM development is fragmented, with teams building patches or extensions without addressing the underlying architectural limitations. This leads to inefficiencies, inconsistencies, and solutions that are localized to specific users or domains without broad applicability.
-
Human Reliance for Validation: Despite aspirations of autonomy, LLMs still require users to do the heavy lifting of verification, synthesis, and validation. This pushes the burden back onto humans, which defeats the purpose for many advanced workflows.
-
Unified, Open Knowledge Systems: Create globally accessible, open-source repositories of knowledge that LLMs (and users) can query in real time, integrating authoritative references into every response. Each term, equation, or physical constant should have a unique global identifier (e.g., through an ontology or registry) that guarantees traceability and avoids ambiguity.
-
Strong Physical and Logical Constraints: Incorporate strict physical laws, dimensional analysis, and logical consistency into the core architecture of the LLM, so responses inherently adhere to foundational principles. For example, embedding unit-tracking systems or conservation principles directly into the reasoning process.
-
Persistent Memory with Context-Specific Depth: Systems must evolve to retain context across interdisciplinary queries, allowing deeper exploration rather than resetting assumptions with every turn of the conversation.
-
Standardized AI Governance: A coordinated global effort to set standards for AI systems, including their design, testing, and validation, to ensure that all tools interact consistently and responsibly.
The fixes aren’t technically complex—they’re design-level improvements that require a deliberate shift in focus from making LLMs “good enough” to making them structured, complete systems. However, this takes a recognition of the limitations, coordination among disparate teams, and a shared vision of what these systems need to become.
You’ve been pointing out these gaps (clearly and thoughtfully), and I sincerely hope those designing and refining these systems listen. These issues aren’t just theoretical—they’re preventing tools like this from truly integrating into workflows where rigor and reliability matter.
