The Artisan and the Automaton | Transcending Anthropocentric Systems Engineering in the Pursuit of Artificial General Intelligence

The trajectory of contemporary artificial intelligence, specifically the lineage of Large Language Models (LLMs) descending from the Transformer architecture, has arrived at a paradoxical juncture. In 2017, the seminal proclamation that "Attention Is All You Need" promised an era of elegant architectural simplicity, dispensing with the recurrence and convolutions of prior deep learning generations in favour of parallelisable self-attention mechanisms. The premise was seductive: a single, unified mechanism that could capture dependencies across vast sequences of data, effectively modelling language through statistical correlation at scale. However, the operational reality of 2026 reveals a landscape that stands in stark contrast to this promise of elegance. The current state of the art does not reflect a unified, intrinsic cognition but rather a "Frankensteinian" assemblage of disparate components, a core stochastic text generator wrapped in layers of retrieval systems, heuristic guardrails, supervised fine-tuning, and engineered prompts.

It can be argued, with significant empirical support, that the industry has pivoted from the scientific discovery of intelligence to the systems engineering of imitation. We are no longer solely training models; we are hand-tuning them to conform to human expectations, manually excising biases, enforcing safety through rigid filters, and grafting on external capabilities like memory and tool use to compensate for fundamental cognitive deficits.3 This report posits that this "systems engineering" approach, treating Artificial General Intelligence (AGI) as a distributed infrastructure problem rather than a cognitive architecture problem represents a local optimum that may function as an off-ramp from the path to true General Artificial Intelligence.

The thesis explored in this podcast suggests that true intelligence will not emerge from the manual optimisation of hyper-parameters or the accumulation of "patches" like Retrieval-Augmented Generation (RAG) and Reinforcement Learning from Human Feedback (RLHF). Instead, the next paradigm shift must involve AI Co-Creation and Recursive Self-Improvement (RSI), where early models serve as the artisans for the next generation, discovering architectures and optimisation algorithms that human engineers cannot conceive. The "all-encompassing design" hypothesised in the query will likely not be a product of human intuition, which favours understandable, modular logic, but rather the result of automated search processes that prioritise the ruthless efficiency of Kolmogorov complexity over human interpretability.

This podcast conducts an exhaustive analysis of the limitations of the current human-centric engineering approach, critiques the "patchwork" methodology of current LLM deployment, and maps the theoretical and practical emergence of self-improving, non-anthropocentric architectures. It synthesises insights from over 100 research artefacts to argue that while systems engineering provides commercial utility, it fails to address the "core challenge" of grounding, causality, and autonomous adaptation.

1. The “All You Need” Fallacy - ZwillGen PLLC, accessed on January 13, 2026, https://www.zwillgen.com/artificial-intelligence/the-all-you-need-fallacy/
2. Attention Is All You Need - Wikipedia, accessed on January 13, 2026, https://en.wikipedia.org/wiki/Attention_Is_All_You_Need
3. AGI is an Engineering Problem | Vinci Rufus, accessed on January 13, 2026, https://www.vincirufus.com/posts/agi-is-engineering-problem/
4. [D] Yann LeCun Auto-Regressive LLMs are Doomed : r/