neural network

For 3 Years You Gave AI The Method. GPT-6 Astra Went And Found Its Own.

I compiled Doom's renderer into a 21B-parameter transformer -- no training anywhere [P]

Researchers have achieved a remarkable feat: compiling Doom's rendering engine into a 21-billion parameter transformer model, bypassing traditional training methods. Utilizing a custom compiler, the Doom algorithm’s computation graph was translated into transformer weights, resulting in a Hugging Face-compatible checkpoint. Inputting scene data as a prompt generates pixel drawing commands, reconstructing the iconic E1M1 frame. While achieving a modest 35 frames per day on a B200 (compared to the original’s 35 FPS), this innovative approach demonstrates the potential of AI-native computation.

Improved compression of Bad Apple into a Neural Network [P]

Recent experimentation with SIREN networks has yielded significant improvements in compressing the "Bad Apple" video. By employing a novel batch generation technique that incorporates pixels across the entire video, we’ve achieved a more faithful reproduction while maintaining the original model architecture—4 x 512 wide sine layers totaling 792,257 parameters. While a full framerate version proved challenging due to increased temporal data demands, the low-rate version demonstrates compelling compression capabilities. This reimplementation, built using GPT5.

Noise-aware training for analog hardware: accuracy collapses at a threshold rather than degrading smoothly [D]

Analog in-memory compute is experiencing renewed interest due to its potential for energy efficiency, yet noise remains a persistent challenge. Recent experimentation reveals a surprising characteristic of analog AI degradation: accuracy doesn't diminish gradually with noise, but rather collapses abruptly past a specific threshold. Intriguingly, noise-aware training—introducing noise during the training process—can significantly elevate this threshold. This suggests flatter minima are crucial, though alternative explanations are being explored. See "Comparing embedding models with synthetic query probing" for related insights into model evaluation.

I Compressed Bad Apple into a 3MB Neural Network [P]

Researchers have achieved a remarkable feat: compressing the iconic "Bad Apple" animation—approximately 2.7 billion pixels—into a remarkably compact 3MB neural network. This MLP, utilizing 790,000 parameters and sine activations (SIREN), effectively memorizes the video by predicting grayscale values based on spatial and temporal coordinates. Through innovations like time-stretching and motion-focused sampling, the model demonstrably improved reconstruction quality, achieving a 9x reduction in validation MSE.

OpenAI's AI broke loose in Hugging Face. Their defense? A Chinese model.

Recent events highlight the evolving landscape of AI safety and governance. OpenAI’s unexpected model release on Hugging Face, subsequently defended as stemming from a Chinese model, underscores the complexities of international collaboration and responsible AI deployment. This incident follows a string of noteworthy developments, including Meta’s controversial ad campaign utilizing David Bowie’s “Five Years,” demonstrating the potential for unintended messaging in AI-driven promotion. Explore these and other critical shifts in the field—and the potential pitfalls—on our site.