The Archive

Concordance comparison method assembles local grammars for Portuguese named entity recognition via pairwise grammar analysis.

EvoNav uses LLMs to evolve reward functions for robot navigation via reinforcement learning, automating design of navigation policies.

Semantic consensus framework enables federated fine-tuning of LLMs through model behavior rather than parameter aggregation, supporting heterogeneous architectures.

UniVLR unifies text and vision in multimodal LLMs by rendering reasoning traces as shared visual workspace, improving latent reasoning efficiency.

Bistable Memory Recurrent Units (BMRU) improve gradient flow and stability for ultra-low power sequence models via hardware-software co-design.

Self-distilled trajectory-aware Boltzmann modeling closes training-inference gap in diffusion language models via multi-step denoising trajectories.

GEAR enables fine-grained credit assignment in RL-trained LLM agents via adaptive-granularity advantage reweighting at token and segment levels.

Proximal spectral gradient methods with convergence guarantees for nonconvex constrained optimization under heavy-tailed noise.

Memristor-based analog content-addressable memory architecture for edge AI inference with improved scalability.

Self-supervised learning framework enforcing martingale consistency across coarse and refined data views.

Fine-tuning approach to improve language model calibration for user-specified output probability distributions.

Spectral analysis of tree ensembles deriving minimax convergence rates and compression schemes for random forests and gradient boosting.

Comparative analysis of decentralized agent discovery mechanisms (Chord, Pastry, Kademlia) across edge and cloud compute.

Spectral gradient clipping method preserving matrix structure and controlling singular value decay during neural network training.

Lifelong normalization mechanism enabling stable sequential model editing in LLMs without catastrophic forgetting.

Spiking neural network framework with gradient-trainable multi-timescale conductance dynamics for sparse temporal processing.

Shift-invariant transformer autoencoder for calibration-robust spectral unmixing in pharmaceutical and biomedical analysis.

The Karpathy coding skill is locked behind Pro. It doesn't use any Pro-only features, so I rewrote it for free plan chat workflows. Same philosophy, tuned for no terminal, no subagents, and a shorter context window where mistakes are expensive. Paste the whole thing into a Project's custom instructions or use it as a system prompt. It auto-triggers on any coding request. --- name: karpathy-coding description: Apply Karpathy-inspired coding discipline to any programming task. Use this skill whenever the user asks you to write, fix, refactor, extend, or review code — even casually...

FrontierMath is supposed to be one of the hard benchmarks for frontier models, and now Epoch is saying an AI-assisted review found fatal errors in about a third of Tiers 1-4. Noam Brown says the initial flags came from GPT-5.5. Obviously we’ll have to wait for the corrected scores, but this is a pretty interesting moment: the model is already strong enough to sanity-check the benchmark.

Wedding guest deployed Claude-based concierge; users attempted jailbreak attempts as second most common interaction.

Morning Everyone! Big one today (**104 changes!**): Claude Code just went async. The new `/goal` command lets you set a completion condition ("all tests pass and the PR is ready"), then Claude keeps grinding across turns until it's hit. The new `claude agents` view shows every session you've got running: working, blocked on you, or done. Translation: kick off a goal -> let claude cook -> come back later. First proper fire-and-forget loop CC has shipped. Pretty huge unlock if you've been juggling multiple sessions and losing track of which one needs you. Full notes: [https://www.luk...

Reddit post identifies systematic absence of Claude users from published AI psychology research, raising methodological concerns about chatbot adoption studies.

Reddit user posts anecdote about ChatGPT confusing VSCode with Victoria's Secret in context retrieval.

Right now, every AI model you've ever used works the same way. You talk, it listens. It responds, you listen. Thinking Machines is trying to change that by building a model that processes your input and generates a response at the same time, so it's more like a phone call than a text chain.

Unitree announces GD01, a manned exoskeleton mecha; hardware milestone with unclear AI integration or technical specs.

User observation about ChatGPT's real-time display behavior during text input; no technical substance.

well done, Team Thinky.

Reddit post claims Mira Murati's new project outperforms OpenAI's GPT-Realtime-2; lacks specifics on model name, capabilities, or verification.

PhD student. Need advice. After the ICML abstract deadline, industry coauthors asked to be removed, they missed their employer's internal approval window. They had contributed (discussions and written feedback) but I hadn't explicitly asked before adding them. January: wrote to PC chairs, got written confirmation from all coauthors, got explicit written approval. Chairs said they'd implement. Never happened. Paper accepted four months later with original author list. At camera-ready we followed up. Chairs reversed: blanket policy, no exceptions, keep the list or withdraw. What do you t...