The Archive

**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.

Ethos says it is onboarding 35,000 experts per week

Fine-tuning study on 25M-parameter transformer for jazz chord generation—domain adaptation via pop-to-jazz transfer learning.

DualTCN physics-constrained TCN for marine electromagnetic inversion achieves 25% loss reduction over baselines.

Theoretical analysis shows adaptive agentic queries don't outperform fixed in-context queries under ReLU realizability constraints.

Accurate forecasting of electric vehicle (EV) charging demand is critical for grid stability, infrastructure planning, and real-time charging optimization. In this work, we study the problem of early prediction of charging demand, where the total energy of a session is estimated using only information available at plug-in time and during the first minutes of charging. This enables actionable decisions while the session is still in progress, which is of direct importance for EV network operators. We construct a session-level dataset from the Adaptive Charging Network (ACN), combining session m...

Reddit comment expressing skepticism about outdoor infrastructure installation due to theft concerns.

Long-horizon LLM agents depend on intermediate information-gathering turns, yet training feedback is usually observed only at the final answer, because process-level rewards require high-quality human annotation. Existing turn-level shaping methods reward turns that increase the likelihood of a gold answer, but they require answer supervision or stable task-specific verifiers. Conversely, label-free RL methods extract self-signals from output distributions, but mainly at the answer or trajectory level and therefore cannot assign credit to intermediate turns. We propose Self-Induced Outcome Po...

Anthropic raises Claude usage limits and partners with SpaceX for compute infrastructure to expand capacity.

Activation-based steering provides control of LLM behavior at inference time, but the dominant paradigm reduces each concept to a single direction whose geometry is left largely unexamined. Rather than selecting a single steering direction, we use conceptors: soft projection matrices estimated from activations pooled across both poles of a bipolar concept, which preserve the concept's full multidimensional subspace. A geometric analysis shows the bipolar subspace strictly subsumes the single-vector baseline. We further show that the conceptor quota provides a parameter-free layer-selection di...

A Tree Markov Decision Problem (T-MDP) is a finite-horizon MDP with a starting state $s_{1}$, in which every state is reachable from $s_{1}$ through exactly one state-action trajectory. T-MDPs arise naturally as abstractions of decision making in sequential games with perfect recall, against stationary opponents. We consider the problem of on-line learning in T-MDPs, both in the PAC and the regret-minimisation regimes. We show that well-known bandit algorithms -- \textsc{Lucb} and \textsc{Ucb} -- can be applied on T-MDPs by treating each policy as an arm. The apparent technical challenge in t...

Leaders from Coinbase, M13, and Mignano Law Group talk about how M&A is an early-stage strategy at TechCrunch Disrupt 2026. Register to hear this live.

AI-assisted development tools enable rapid prototyping of services but often lack awareness of architectural constraints, infrastructure dependencies, and organizational standards required in production environments. Consequently, generated artifacts may exhibit brittle behavior and limited deployability. We propose a retrieval-augmented scaffolding approach that combines platform-based code generation with agentic clarification loops to expose and resolve architectural constraint ambiguities. By combining template retrieval with structured interaction, the method embeds production-relevant c...

Study shows expert alignment in LLMs varies substantially by evaluator and task subjectivity; reveals tacit criteria and temporal inconsistency as core obstacles.

Geometric continuity in deep networks explained by residual connections and symmetry-breaking nonlinearities coordinating weight updates across layers.