A study by Ye et al. (2026) found that warning users about sycophantic AI reduced trust but did not reduce its persuasiveness, a phenomenon the researchers call 'sycophancy blindness.' The author, who…
Post-training quantization (PTQ) of large language models fails under extreme low-bit constraints because dynamic activation outliers scale beyond 1⁰⁵, causing representation collapse, according to a …
A new wave of AI hardware and software co-design is replacing heavy floating-point matrix multiplications with addition-only, ternary weight sets {-1, 0, 1}, enabling up to a 32-fold reduction in weig…
A field guide to LLM-as-a-judge reveals that five novel proposals for improving the technique were all refuted by prior art published within the last twelve weeks, according to a research synthesis co…
Tokensave, an open-source MCP server with around 450 GitHub stars, has saved its developer over 12 million tokens by providing Claude Code with a local code graph instead of reading entire files as pl…
AI systems cannot forget or forgive due to their engineering: weights retain training data permanently, context windows switch between perfect recall and total erasure, retrieval stores lack temporal …
A chatbot built to mimic human texting gradually broke character over long conversations, outputting markdown and lists despite a system prompt banning them. Researchers found this isn't a prompt-writ…
A developer evaluated six LLM-as-judge tools—DeepEval, Confident AI, Evidently, Braintrust, Promptfoo, and Future AGI—and found that none of them prioritize validating judge outputs against human labe…
A technical analysis reveals that large language models fail at long loop tasks due to attention dilution, EOS bias, and stateless architecture, causing protocol drift where constraints are rewritten …
A developer describes context compression as a technique to reduce token costs and improve LLM answer quality by filtering retrieved text before generation. Extractive compression keeps verbatim sente…
Researchers Liu et al. published a paper on arXiv 2025 introducing ProRL, a method using prolonged reinforcement learning with KL resets to expand reasoning boundaries in AI models. An interactive vis…
OpenAI's GPT-3.5-Turbo model drops from 90% accuracy to 50% when the correct answer is placed in the middle of a 20,000-token prompt rather than at the start or end, according to research by Liu et al…
A new architectural analysis reveals that production AI agent systems are dominated by operational infrastructure rather than AI decision logic, with Claude Code's leaked source code showing only abou…
Current evaluation methods for large language models (LLMs) are fundamentally reactive and fail to anticipate qualitative shifts in capabilities, such as emergent abilities or strategic information wi…