Sarah Guo, who became Greylock's youngest general partner at 28 and left four years later to found Conviction, staked her firm entirely on AI before ChatGPT launched, with pre-ChatGPT seed checks into…
Cursor rebuilt its agent swarm architecture, using frontier models as planners to decompose goals into subtasks and cheaper models as workers to execute them. In a stress test rebuilding SQLite from d…
Ben Thompson argues that alarm over Chinese AI models like Kimi K3 is economically overblown because tokens are a commodity, while frontier labs sell scarce intelligence. He recommends loosening restr…
Lightning Jar, a web development firm, has switched from Google Analytics to Plausible on its own sites and most client projects, citing privacy, performance, and compliance benefits. Plausible's cook…
Iain Harper argues that the AI industry's promise of safety guarantees is mathematically unsupported, tracing the limits of rule-based systems from Kurt Gödel's 1931 incompleteness theorems through Al…
Study AL found that a prompt-level fence against self-eviction reduced client-side pruning from 5 to 1 pooled across models, but the pre-registered significance gate failed at p = .219 because the con…
A study from barkup-bench found that returning structured error messages to AI models during patch validation does not improve their ability to correct failures. Across three models—Opus, Gemini, and …
A study by barkup-bench found that requiring an LLM to restate a goal before rewriting it does not improve performance: every model complied perfectly 90 out of 90 times but lost every decisive compar…
Barkup-bench's Study AD confirms that the core LLM editing stack validated on claude-sonnet-4.5 and gemini-3.5-flash transfers to claude-opus-4.8, passing every pre-registered gate with 194/200 on the…
A new pre-registered benchmark series, barkup-bench, found that large language model agents silently invent missing information rather than asking for it, but a simple one-sentence permission to ask e…
A pre-registered benchmark series, Study Z, found that large language models (LLMs) follow standing context blocks with perfect accuracy—216 of 216 on facts and rules across three models with zero cro…
Barkup-bench Study Z found that LLM agents (claude-sonnet-4.5, gemini-3.5-flash, and claude-opus-4.8) correctly retrieved planted facts and applied styleguide rules from full 3,300-token context packs…
A month-long study of 23 experiments with over 19,000 scored runs found that the most reliable way to let an LLM agent edit structured documents is a division of labor: the model acts as an executor w…
A new study in the barkup-bench series found that AI models can follow qualitative goals embedded in application memos as effectively as explicit instructions, but models that read the goal's document…
Barkup-bench Studies T and U reveal that stateless LLM agents fail at two common real-world editing tasks: recalling facts stated earlier in a conversation and reading content outside the visible docu…
A pre-registered benchmark series called barkup-bench found that both a 'keep-history' and a 'stateless plus worked examples' recipe for letting an LLM edit structured documents remain reliable across…
A new benchmark series, barkup-bench, finds that LLM agents editing structured data do not need conversation history; two worked examples in the system prompt match or exceed the performance of full s…
Researchers at barkup-bench published two new studies showing that replacing tree-navigation with a search tool dramatically improves LLM agent accuracy and cost when editing structured data. The sear…
The barkup-bench series found that LLM agents editing structured trees fail when conversation history is removed and when models must locate nodes without stable IDs. Study M showed stateless sessions…
Researchers at barkup-bench ran seven pre-registered studies with over 13,000 model runs to determine the most reliable way for LLM agents to edit structured data. They found that giving every node a …