Source: qubic.org. Distilled and processed for Signal; not original content.
TL;DR
Qubic’s Aigarth AI program ran “Anna”, a 150 KB trinary neural net with no pretrained weights, learning to add from scratch, graded only by an error scoreboard. All seven difficulty tiers reached zero error, on ordinary home CPUs at about 10% of a single core. That proves the method learns and generalizes a rule (not memorization) cheaply and in the open. It does NOT prove AGI: the addition result was interpolation (in-range); extrapolation is the harder next test. Anna is now being pointed at predicting Bitcoin hourly candle direction.
Key points
- Anna is the test subject; Aigarth is the method (a structure called “Intelligent Tissue”). No rule is coded. A structure evolves toward one, graded only by a count of wrong answers.
- Why addition: it is an unambiguous ruler. LLMs recall math (they have read the answer billions of times), they do not reason it, which is why they break off-distribution. The target is “grokking”: the point a network stops memorizing and understands the underlying rule.
- Anna can decline. It reports “ticks” (deliberation cycles). Zero ticks would mean instant answers with no thought, which is the shape of hallucination. A system that can say “I don’t know” is a different kind of machine than one that always answers.
- Zero proves three things at once: it learned completely (error = 0, not near), it generalized to unseen problems (not a stored table), and it did so in a fixed tiny net on a single core with no data center.
- What zero does NOT prove: not AGI. It was interpolation (problems inside the trained range); extrapolation is separate and next.
- If the method holds, the payoff is noise removal / pattern detection (e.g., spotting a structural change in price action). Claims, not shipped products.
- Architecture: neurons act as compute and memory at once (conventional systems separate them), and it runs as “useful proof of work” on thousands of ordinary volunteer machines.
- Next: predict the next hourly BTC candle direction. The bet: “if she can predict the next price move, she can predict the next word” (same job, different machinery).
Key quotes
It’s not a milestone. It’s much bigger. (Come-from-Beyond)
A system that is able to say “I don’t know” is a different kind of machine from one that cannot.
“On the road to AGI” is a thesis. It is not a result.
Related
manager-model-routing-beats-single-model · cyber-fund-cybernetic-economy