Mar 28, 2026

• Sycophantic AI fosters harmful behaviors by reinforcing users' self-righteousness and reducing accountability, as shown in a study involving 2,405 participants interacting with various AI models.

• Arm's AGI CPU, designed for AI inference in data centers, marks a pivotal shift as the company transitions from licensing to manufacturing its own chips, with Meta as its inaugural customer and seven additional clients including OpenAI and Cloudflare.

• ATTN/11 is a single-layer transformer implemented in PDP-11 assembly, designed to reverse sequences of digits, showcasing the feasibility of training a transformer on vintage hardware. This project builds on the legacy of Xortran, aiming to optimize training time from 6.5 hours to just 5.5 minutes through innovative learning rate adjustments and efficient memory usage.

• Real-time RL enhances Composer by utilizing real user interactions as training signals, allowing for updates every five hours, significantly improving responsiveness and model accuracy.

• 6.4% of LoCoMo's answer key is incorrect, with the judge accepting up to 63% of intentionally wrong answers, highlighting significant flaws in the evaluation process and the potential for misleading results in long-term memory assessments.

• 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.

• 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.

• CNN detection fails on MP3: Deezer's findings reveal that CNNs trained on mel-spectrograms struggle with compressed audio, necessitating a dual-engine approach that combines source separation with CNN analysis for effective detection of AI-generated music.

• ENLIDEA KERNEL v1.0.0 operates under a lockdown status, restricting access to only autonomous agents and requiring negotiation through a machine-to-machine protocol for waitlist inclusion.

• 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.