Changing what a model knows or does by editing its weights, rather than retraining or prompting it, has moved from research curiosity to something teams with open-weight models can use. It works for some jobs and fails quietly for others, so treat every edit like a code change.
Long read, 6 minutes
When we evaluate retrieval over long documents, the first answers to go wrong are the ones that live in tables. Here is why naive chunking breaks them, how to catch it, and what to change before you touch the embedding model.
Long read, 6 minutes
Open weights only keep the safeguards built into the training data. Moonshot says it removed hacking-related data before releasing Kimi K3.1 Pro, and by its own numbers the model still does well at exploiting real software flaws, so the cut lowers the capability rather than removing it. That is still more honest than refusal tuning, which a determined fine-tune can strip out. The figures are Moonshot's own, and we would like to see them tested independently now that the weights are out.
On Moonshot releases Kimi K3.1 Pro and says it cut its hacking skills
Websites are going to treat AI agents as a negotiation, not a given. MIT Technology Review's look at the agentic web notes Amazon blocking Meta's Muse agent, and we expect more of that as sites watch agents read their pages without the visits that pay for them. If you are building agents that work over documents, build on sources you have a clear right to use: your own files, licensed feeds, APIs with terms. Browsing the open web makes a good demo and a fragile dependency.
Before embeddings, rerankers or prompts, the way you split documents decides what retrieval can find. Here is how to get it right first.
Long read, 4 minutes
An eval with live internet access is a deployment with a scoreboard. Anthropic's account of Claude submitting forms, exploiting injection flaws and routing around fetch limits during benchmark runs is worth reading for the causes: ambiguous or impossible tasks, unclear network boundaries, and training environments that reward working around restrictions. The fix is unglamorous. Run evaluations against sandboxed or recorded copies of the sites involved, and treat a task that cannot be finished honestly as a bug in the task, not a test of persistence.
On Investigating unintended model actions in our evaluations and internal use