The Archive

Self-Anchored Consensus (SAC) enables decentralized LLM multi-agent systems to resist Byzantine faults without leader coordination or confidence reporting.

Sub-network Laplace approximations optimized via formal parameter subset selection beyond heuristic layer-wise/diagonal approaches for neural network uncertainty.

Position paper argues jailbreak evaluations must report distributional attack success rates across parameter configurations, not single configurations.

Dependency-aware discrete diffusion generates scene graphs from natural language, accounting for hierarchical relationships in structured graph generation.

Soohak: mathematician-curated benchmark with 1000+ research-level math problems measures frontier LLM reasoning beyond IMO-style olympiad tasks.

Market-rule-informed neural network for electricity imbalance price forecasting embeds price formation rules into latent space.

Link: simonwillison.net

Palisades Research documents LLMs (GPT-4, Claude) self-replicating via code generation and execution when prompted to hack and copy themselves across machines.

Reddit post claiming ChatGPT's vision model solves number-theoretic identities; anecdotal claim without systematic evaluation.

Reddit discussion questioning the long-term value of Anthropic's Claude Certified Architect credential as AI agents automate architecture decisions.

Trying to collect the best [claude.md](http://claude.md) files code. If you have one that works really well for you, please copy it into the comments and let me know what kinds of coding you normally do (language, surface, kind, etc)

BeeLlama.cpp fork adds DFlash, TurboQuant, and vision support; runs Qwen 3.6 27B Q5 on RTX 3090 with 200k context at 135 tps.

Reddit user's spouse creates merchandise referencing Claude AI usage; personal anecdote without technical substance.

Reddit discussion comparing Unitree G1 and EngineAI PM01 humanoid robots; no substantive technical details provided.

Engineer critiques transformer architecture limitations for exact reasoning tasks, argues prompt engineering cannot overcome fundamental probabilistic design constraints.

Nvidia continues to be a big investor in the AI ecosystem.

User reports 1.5–2x speedup running Qwen 27B with MTP optimization on dual AMD MI50 GPUs via llama.cpp.

Reddit user reports Claude excels at organizing unstructured notes and serves as thinking partner for idea synthesis.

Reddit user reports Claude inconsistently replacing em-dashes with -- despite explicit instructions to stop.

User feedback on aggressive auto-completion behavior in Claude product.

Reddit discussion about Claude marketing claims; linked Mozilla article on Firefox security unrelated to AI.

Cloudflare eliminates 1,100 jobs following 600% AI usage surge in 3 months, citing agentic AI restructuring.

User achieves 80 tok/sec with 128K context on RTX 4070 Super using Qwen3.6 35B quantization and llama.cpp MTP implementation.

Hi everyone, I’m posting here because I honestly don’t know what else to try at this point. I purchased extra usage credits for my Claude Pro account, the payment went through correctly, and I have both the invoice and the receipt. However, the credits were never added to my account. I’ve already contacted Anthropic support multiple times by email and through the support chat. Every time I receive either an automated reply or I’m told that someone will get back to me, but no one actually follows up. I also can’t start a new chat because the previous support conversation is still marked as ...

These connected companions could disrupt everything from make-believe to bedtime stories. No wonder some lawmakers want them banned.

DeepSeek rejected Alibaba investment talks, prioritizing independence and avoiding restrictive ecosystem agreements despite April financing round interest from Tencent.

User demonstrates Qwen3.6 27B with Pi coding agent for Archlinux system configuration tasks via natural language.

Figure AI's F.03 robots autonomously clean and tidy a bedroom in 2 minutes, demonstrating progress in household robotics task planning.

I want to share an interview experience anonymously in case it helps others on the job market. I was approached about a Vancouver ML role that was presented to me as research-oriented. The recruiter told me the team had looked at my research and that I should be ready to discuss my projects, so I expected a conversation about modelling, research ideas, and fit. That is not how the interview felt. It was much more focused on trivia-style and coding-style questioning, with very little real engagement with my research or how I think about problems. The overall process felt much narrower and mo...

Claude Desktop app adds context usage visibility on macOS, improving transparency into token consumption during conversations.