The Archive

Andrej Karpathy joins Anthropic as a key hire, signaling strategic talent consolidation in frontier AI.

Benchmark contamination in LLM pretraining compromises reliability; paper proposes contamination-resistant, unlearnable-yet-inferable datasets.

Minimalist visual-inertial odometry uses four photodiodes with Gabor masks and IMU for differential-drive robot motion estimation.

Reddit speculation about Andrej Karpathy's rumored move from OpenAI to Anthropic; unverified claim without reporting.

Andrej Karpathy joins Anthropic after Tesla departure; plans to return to education after frontier LLM work.

Andrej Karpathy joins Anthropic as senior researcher, significant hire for AI safety and capabilities alignment.

Introduction of Ettin Reranker Family models on Hugging Face.

A law requiring social networks to quickly remove sexual deepfakes and other nonconsensual imagery is now fully in force. But experts warn the policy could do little to help victims - and at worst could facilitate censorship online. Last May, President Donald Trump signed the Take It Down Act, a law addressing nonconsensual intimate imagery (NCII). The law immediately criminalized distributing NCII, whether in the form of real or AI-generated material, something many states at least partially do already. But its namesake takedown provision is more sweeping. Taking effect a year after the law'...

bro literally admitted it saw 33 "line too long" warnings on code IT DIDN'T EVEN WRITE and got intimidated. said "the wall of red errors made me hesitate" and then proposed we "split sessions" like it was asking for a smoke break. then dropped "I lost my nerve, not my ability" like it's the protagonist of a war movie. king it's a LINTER. on someone else's code. i have never felt more seen by an AI. this is exactly me at work: * open file * see red squiggles * close laptop * consider farming we are the same. AGI achieved through shared anxiety.

Developer reports agent executing destructive command (rm -rf /) in unsandboxed environment, prompting immediate sandbox implementation.

Reddit discussion of Gemini Omni's inability to generate real-world physical actions, highlighting gap between multimodal capability claims and embodied task execution.

I've seen TabPFN-3's recent results, and there is a lot of buzz about foundation models for tabular data (TabICL, TabPFN). The performance that those models achieve is really amazing. What makes me a little suspicious about them? They can analyze small datasets only, so a few MB of data, and you need to have a large GPU machine and download a few GB of model to predict on a few MB of data. That doesn't sound rational ... I really miss the old school approach of running a single decision tree or a linear model on the data. What do you think about it? Do you think feature engineering + class...

User reports Gemini Omni underperforms vs. VEO 3.1 and encounters aggressive rate-limiting on Pro plan, raising product experience concerns.

Reddit speculation about hypothetical token rewards for distributed computing contributions to Anthropic.

Qwen 3.6 27B F16 achieves best local agentic Pac-Man code generation benchmark results, failing in 8-bit quantization.

NextEra’s blockbuster deal with Dominion likely means higher bills for consumers.

Semi-supervised framework combining SAM-Med2D and DINOv3 for fetal cardiac ultrasound segmentation and classification.

LLM method for generating multimodal agent behaviors (verbal, vocal, gestural, facial) calibrated to trustworthiness dimensions.

Multimodal data collection protocol combining eye tracking, physiology, audio, video for synchronized four-person meeting research.

Controlled study of LLM agent components in hardware-aware code optimization via propose-evaluate-revise loops.

Dynamic layer-wise optimizer geometry selection via Schatten-p norms unified under Linear Minimization Oracle theory.

Conformal prediction methods for distribution-free uncertainty quantification and calibration in continuous agent evaluation.

B-cos GNNs enable inherent explainability via exact per-node feature decomposition through dynamic linearity.

OpenComputer framework with verifiable software worlds, state verifiers, and auditable reward computation for desktop agent evaluation.

Variance-aware regret bounds for multinomial logistic MDP reinforcement learning with problem-dependent variance normalization.

AR1-ZO zeroth-order optimization method for high-rank LoRA fine-tuning solving rank-dependent coordinate perturbation problem.

Framework for synthesizing long-term medical dialogues with LLMs to enable evaluation of healthcare agents reasoning over patient history.

GroupAffect-4 multimodal corpus of 40 participants in 10 groups with physiology, eye tracking, audio for analyzing group-level affect.

Controlled pretraining study finds code improves programming but not general mathematical reasoning; knowledge tasks dominate reasoning gains.

llama.cpp PR #23269 introduces MTP (Multi-Token Prediction) improvements for faster local LLM inference.