Unverified50% confidenceFactExact time
In the MIT experiment, four tiny policy networks played grid-simplified versions of Breakout, Pong, Pac-Man and Freeway respectively; each network had 17,253 weights, and Astra got all four right, with confidence between 96% and 99%.
1
Sources
50%
Confidence
Long-term
Relevance
11/4/2026
First Seen
Sources
Can Qwen 3.8 27B Really Read Neural Network Weights? A 16GB Hands-On Test
youtubeOpaya11/4/2026
Related Entities
Related Claims
UnverifiedIn the Go experiment, Astra wrote a Go program layer by layer in 336 lines of code and, with zero test runs, beat two smaller trained networks with a win rate above half; after 50 test-run self-play games, it beat an equally sized network trained on 20 million games.64% similarVerifiedOn the MNIST dataset, a simple fully connected network can achieve over 98% accuracy, and a convolutional network can easily exceed 99.5%.60% similarUnverifiedA 56-layer network can mathematically express everything a 20-layer network can express; it only needs the extra 36 layers to learn identity mappings to be equivalent to the shallower network.59% similarUnverifiedIn a counterfactual analysis of 500 divergence positions (10,000 MCTS iterations each, rollout depth 20, repeated three times), among the 105 positions where Jev went face while the neural network developed the board, the deep MCTS preferred Jev in 92 cases, preferred the neural network in 3, and produced 10 ties.58% similarUnverifiedIn the Breakout network, summing the player-position-related weights by grid cell, the bottom row inside the wall accounted for 55% while other rows had only about 5%; in the other three games, no row exceeded 21%.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1135723API
curl https://kongchang.com/api/v1/knowledge/claims/1135723MCP
get_claim(id=1135723)