ARTIFICIAL JUST INTELLIGENCE
THIS IS THE NEXT STEP
The explicit and implicit promises of the LLM developers to create a global super-duper intelligence that will know everything that humanity knows, is capable of solving any problems, effectively replacing humans everywhere, have fully justified the opinion of the most rabid skeptics. Instead of what was expected, the public received a completely useful tool - but at the same time requiring human intelligence to detect nonsense regularly generated by the system, monstrously expensive to develop and operate.
What’s next? Authorities argue about what the new route to General Artificial Intelligence will be, listing the skills/abilities required - for example, “What is (Real) AGI?” and “AGI Checklist” [https://petervoss.substack.com/p/agi-checklist] by Peter Voss.
Developing a “full-scale AGI” that implements all the rational requirements for such systems is obviously an extraordinary task. And the main obstacle is that the requirements list is available, but the ways to implement them remain unclear or are not publicly discussed. At the same time, it is apparent to those involved in real-world development that it is unrealistic to bridge the gap between LLMs and the desired AGI in a single leap.
If, however, we adhere to the declared goal of developing AGI, it is reasonable, as a first step, to create a system that, while much less perfect (and therefore much less useful), will demonstrate the presence of all the required capabilities. Such a system - we will call it Artificial just Intelligence (AjI) in the future - will define a set of subsystems, the combination of which can demonstrate the minimal implementation of all required functions. Further progress can then be made by improving each component independently of the others (engineers know how much easier this makes things).
Naturally, this path immediately raises two sets of questions: what constitutes the minimum implementation of each required function, which existing components can be used, and which need to be developed from scratch.
The current (fourth) series of our notes describes a feasible (in our opinion) approach to implementing such an AjI. Stay tuned to our channel!