Beyond Scale: The 2025 Landscape of Deutsch-Style AGI Efforts

Why Deutsch Matters

David Deutsch argues that true progress—scientific, technological, even moral—comes from explanatory knowledge: testable conjectures born out our our minds, that survive relentless criticism. Deutsch-style AGI will be far different than scaled up versions of the current LLM models. It will be sentient and creative in the same ways humans are.

The use cases for the current LLM's are vast and financially enticing so the investment direction of the "AI Boom" is understandable. But I'm interested in what projects are actually working on the Deutsch-style AGI problem, or at least what projects have the potential to converge on it. Here is a chatGPT o3 survey of the current efforts in this direction. It has no idea what it put together here, but in fine LLM style it has synthesized the relevant information into a useful and insightful table.

I'm particularly interested in the Numenta “Thousand Brains” project. I feel like it's on the right track in some important ways.

Facet	What it means
EK – Explanatory knowledge	Understands reality through causal, criticisable models, not statistical curve-fitting.
CC – Creative conjecture	Can invent genuinely new ideas, not just remix prior data.
CF – Critical feedback	Tests and refines conjectures to produce hard-to-vary explanations.
EC – Error correction	Treats all knowledge as fallible; builds in mechanisms to detect and fix mistakes.
FW – Free-will-like autonomy	Sets or renegotiates its own goals; can refuse or reprioritise tasks.

The 2025 AGI Effort Map

Line of attack / Flagship project	Core idea	2024–25 milestones	Deutsch-facet coverage†	Key pros	Key challenges
Embodied sensorimotor Numenta “Thousand Brains Project”	Cortical columns build 3-D models; thousands vote	Code open-sourced; new non-profit formed	EK: ✔︎ • CC: ✔︎ • CF: ✔︎ • EC: ✔︎ • FW: △	Biologically grounded; ultra-sparse, low power	Needs rich robotic data; scaling columns unproven
Universal-AI mathematicsAIXI, Darwin Gödel Machine (DGM)	Optimal Bayesian search + provable self-modification	Unified curiosity/reward theory; self-rewriting agent jumps SWE-Bench from 20 → 50%	EK: ✔︎ • CC: ✔︎ • CF: ✔︎ • EC: ✔︎ • FW: △	Formal guarantees; first wild-run self-improvement demos	Proofs explode at scale; reward still hard-wired
Symbolic + neural hybrids OpenCog Hyperon, NARS 3.0	Typed knowledge graphs plus neural perception	Hyperon beta due Q4 2025; NARS adds causal abduction	EK: ✔︎ • CC: ✔︎ • CF: ✔︎ • EC: ✔︎ • FW: △	Human-readable reasoning; modest compute	Perception pipeline, small dev teams
System-2 on LLMs OpenAI “Strawberry”, Gemini 2.5	Planning & tool-use overlays on giant nets	Web-scale planner prototypes; Olympiad-level maths mode	EK: △ • CC: △ • CF: ✔︎ • EC: △ • FW: ✖︎	Leverages existing data & infra; ships quickly	Still largely curve-fitting; opaque internals
Open-ended curricula Enhanced POET, evolutionary agents	Tasks and agents co-evolve indefinitely	Code-evolving bots double coding benchmark	EK: △ • CC: ✔︎ • CF: △ • EC: ✔︎ • FW: ✖︎	Generates surprising designs	Fitness ≠ explanations; compute hungry
Neuromorphic & whole-brain sims Intel Hala Point, Blue Brain legacy	Spike-based hardware & biophysical simulation	1.15 B-neuron neuromorph opened to researchers	EK: ✔︎ • CC: △ • CF: ✔︎ • EC: △ • FW: ✖︎	Ultra-low power; test-bed for cortical theory	Programming spikes is hard; leap to reasoning unclear
Safety-first super-intelligence labs Safe Superintelligence Inc.	Alignment-by-construction before release	$2 B funding; formal-verification hires	EK: △ • CC: △ • CF: ✔︎ • EC: ✔︎ • FW: ✔︎	Popperian error-correction at the core; deep pockets	Secrecy slows peer review; proofs still hypothetical

† Legend — ✔︎ strong evidence △ partial ✖︎ weak/missing

What This Table Shows

• Nobody ticks every box—yet.

  • Thousand Brains and the Gödel-machine line cover the most facets, but both still need a robust preference-formation layer to meet the “free-will” criterion.

  • LLM-plus-planner projects sprint ahead on benchmarks but remain weakest on explanatory depth and agency.

• Convergence signs are real.

  • System-2 overlays are borrowing symbolic memories and self-critique tools from projects like NARS.

  • Thousand Brains and Hyperon teams are planning joint robotics benchmarks to anchor sensorimotor grounding.

  • DGM-style self-improvement is being explored as a plug-in for both hybrid symbolic systems and neuromorphic hardware.

• Key open research questions:

  1. Autonomous preference formation – How can an agent develop its own goals instead of maximising a hand-coded reward? (FW)

  2. Public, criticisable explanations – How do we translate internal causal models into statements humans can debate? (EK + CF)

  3. Scalable error-driven learning – Can we keep compute and energy manageable while letting the system learn from every mistake? (EC)

  4. Keeping safety proofs in lock-step – Self-modifying agents evolve; how do formal guarantees keep up? (CC + EC)

Why It Matters

If Deutsch is right, humanity’s biggest breakthroughs will come from ideas that survive criticism, not from larger parameter counts. The projects above are the first serious attempts to code those principles—explanation, conjecture, criticism, correction, autonomy—into silicon. None has cracked the full set, but the diversity of approaches suggests we’re finally exploring the right search space.