Mar 29, 2026

• TurboQuant offers a revolutionary two-stage algorithm that compresses KV cache in AI models, achieving a 6x reduction in memory size without sacrificing accuracy, thus addressing the growing memory demands of large language models (LLMs).

• Vibe coding, a term coined by Andrej Karpathy, involves deploying AI-generated code without thorough review, leading to significant failures, including a 6-hour outage at Amazon that resulted in 6.3 million lost orders.

• Meta and Arm are collaborating to create the Arm AGI CPU, a groundbreaking data center processor tailored for AI workloads, enhancing performance and efficiency beyond traditional CPUs.

• TurboQuant achieves 3.2× memory savings by compressing model weights with near-optimal distortion, offering a drop-in replacement for nn.Linear that maintains performance.

• The thermodynamics of computation is fundamentally flawed due to its reliance on idealizations that ignore thermal fluctuations, leading to erroneous conclusions about non-dissipative processes at molecular scales.

• A benchmark was created to evaluate LLMs on 28 physics laws, generating adversarial questions that exploit common cognitive biases and unit confusions, ensuring a rigorous assessment through symbolic math rather than subjective judgment.

• PentaNet introduces a pentanary quantization scheme {-2, -1, 0, +1, +2}, enhancing model capacity by 47% compared to ternary quantization, while maintaining zero-multiplier inference benefits through simple bit-shifting.

• Litellm versions 1.82.7 and 1.82.8 were compromised, allowing a malicious .pth file to execute on every Python process start, scraping sensitive data like SSH keys and API credentials without requiring imports.

• The first open-source implementation of Hebbian fast-weight write-back for the BDH architecture enables the model to rewrite its own decoder weights during inference, utilizing sparse activation codes as addresses, which was previously unimplemented publicly.

• New hafnium oxide memristors could reduce AI energy consumption by up to 70% by mimicking the brain's efficient neural connections, enabling both storage and processing in a single location.

• The implementation of TurboQuant in Python introduces a novel approach to quantization by utilizing random rotation to achieve optimal 1D quantization without the need for calibration data or dataset-specific tuning, making it applicable across various contexts.

• LVFace, utilizing a Vision Transformer (ViT) backbone, reportedly outperforms ArcFace in facial recognition tasks, particularly in scenarios involving partially occluded faces like those wearing masks, as evidenced by its first-place finish in the MFR-Ongoing challenge.