The Archive

Live coverage of Anthropic's Code w/ Claude 2026 event keynote and announcements from Simon Willison.

Fully convolutional neural network for chemical-mechanical polishing modeling in IC manufacturing using white light interferometry.

Systematic review of jailbreak attack and defense methods for LLMs with critique of narrow evaluation metrics like attack success rate.

Adaptive deep learning framework for angle-of-arrival based outdoor localization in 5G/6G networks with flexible training strategies.

Flow matching method for few-shot vision-language model adaptation using polar decomposition to decouple radial and angular feature dynamics.

While citing web forums and discussion boards can help users find answers to more niche queries, this design choice could also prove chaotic.

Analysis of car-following deceleration behavior using NGSIM trajectory data identifying gap-closing rate and visual looming discriminants.

Resource modeling and pipelined hybrid parallelism system for efficient large-scale Mixture-of-Experts training on HPC platforms.

Genesis AI, a startup that raised a $105 million seed round to build foundational AI for robotics, has unveiled its first model, GENE-26.5, but also a demo showcasing a set of robotic hands performing complex tasks.

Up to 3x the speed with no loss of quality—is it too good to be true?

Study of relation hallucination in vision-language models under rotation and noise perturbations with evaluation of augmentation and preprocessing defenses.

Preference-based self-distillation method for on-policy training that moves beyond KL matching via reward regularization to improve reasoning stability.

Automated incident alert: elevated error rates across multiple Claude models on 2026-05-06, status tracking post.

Match Group said that it's slowing its hiring plans for the rest of the year because AI tools "cost a lot of money."

**TL;DR:** My last post about testing TinyGPU attracted some interest. This is the follow-up. The Blackwell card is detected and the driver loads, but NVIDIA's GSP firmware fails to boot through TB5 (known issue, I'm working with tinygrad on it). While debugging that, I went down a rabbit hole and discovered that Apple's RDMA subsystem accepts Metal GPU buffers for zero-copy network transfers — something nobody has documented. I also found hidden `ibv_reg_dmabuf_mr` symbols in Apple's libibverbs that suggest GPUDirect RDMA might be possible on macOS without any kernel modification. Here's eve...

Empirical study finds predictive neural encoders systematically fail to learn causal representations, achieving 49% causal fidelity despite high prediction accuracy across 2695 configurations.

Proposes detecting structural hallucinations in diffusion models via local intrinsic dimension analysis as instabilities on model-induced manifolds.

User reports Qwen 3.6 27B in Hermes agent harness successfully handles junior IT tasks, signaling maturity of local model + agent systems.

Single-pass hallucination detection method for LLMs using attention head KL-divergence without sampling, validated across multiple model families.

HEDGE: generative model for hypergraphs using structured stochastic diffusion with two-sided heat operator to preserve higher-order interaction structure.

CuBridge: LLM-based framework for generating and reconstructing high-performance CUDA attention kernels with improved correctness and efficiency.

Graph-SND: sparse-graph generalization of System Neural Diversity metric for multi-agent RL, reducing quadratic-time computation to O(|E|) with unbiased estimation.

Position paper argues embodied AI deployment in sensitive environments creates systemic privacy crisis requiring fundamental privacy-utility trade-off design.

Apple has agreed to pay $250 million to settle a class action lawsuit for overpromising the arrival of Siri's AI features.

Decentralized learning framework where heterogeneous nodes train learned neighbor-trust policies for collaborative inference deployment in IoT.

Spatial regionalization method using minimum description length principle to partition time-evolving domains without pre-specifying region count.

Uno-Orchestra: unified LLM multi-agent orchestration policy that jointly learns task decomposition and worker selection via RL, benchmarked on 13 suites.

Reward models fail to capture socially desirable preferences across bias, safety, morality, and ethics—exposing hidden alignment failures in LLM training.

I was working on a project, I got hungry went to eat and take a shower while also having this be my break, came back, session was at 0%, typed to claude that the animation of the CSS needs to be slower and more subtle, he changed it, 45% usage. Nowhere did it warn me that possibly cache was cold or that I would be consuming a lot of tokens to CONTINUE a chat that I didn't close on the same PC. So now I have to slow down my work and wait for this 5 hour cycle to end to properly speed up my progress.

SLYP agent discovers Windows COM privilege-escalation race conditions via agentic binary exploration and generates debugger-verified proof-of-concept exploits.