Blog · Category
AI news
14 posts on AI news, newest first.
Cursor can now reach the agent on your laptop. Your teammates still cannot
Cursor's Remote Control reaches a local agent from your phone. What the agent writes is still on one machine that has to stay awake. Here is the other half.
Oct 8, 2026 · 6 min readMistral Large 4 scored 59.9% on business workflows someone else had already written down
Mistral Large 4 scores 59.9% on 657 business workflows an eval specified for it. Your team's workflows are not specified anywhere, and that is the gap.
Oct 7, 2026 · 6 min readGoogle open-sourced the embedder. Your notes app kept the index
EmbeddingGemma 2 is Apache-2.0 and runs on a laptop. It only helps your team if the notes are files you hold and the search index is a cache you can rebuild.
Oct 7, 2026 · 7 min readAgent memory fails by throwing things away, not by searching badly
A streaming-recall benchmark found every failure in a bounded memory was an eviction, not a bad rank. Retrieval cannot reach what was already dropped.
Oct 7, 2026 · 8 min readThe team's vault is the one MCP server that should not run on a laptop
MCP's own docs now say a local stdio server inherits your whole account, and that folder allowlists enforce nothing. What that means for a team vault.
Oct 7, 2026 · 8 min readYour notes outlive the model that reads them
Anthropic retired Sonnet 4.5 and OpenAI set shutdown dates in one week. Model churn is scheduled. Your team's knowledge should not be on that clock.
Oct 6, 2026 · 7 min readNobody keeps a list of what is subscribed to your team's notes
OpenAI shipped MCP Events at DevDay. A subscription is a standing grant, and the draft defines no way to list one. Your notes server has to keep that list.
Oct 6, 2026 · 6 min readOpen weights, rented notes: what Beam's Apache-2.0 licence does not cover
Reflection AI will ship Beam's weights under Apache-2.0. That licence covers the model, not the knowledge it reads. The corpus is the half nobody self-hosts.
Oct 6, 2026 · 8 min readWhen your team's knowledge is a person, an AI has nothing to cite
Stack Overflow's 2026 survey: 93% need source attribution to trust AI, and 72% get work answers from a coworker. A colleague is not a citable source.
Oct 6, 2026 · 7 min readResearch says a model's context belongs in a file. Keep one your team can open
Context Language Models let a model edit its context as a file, and beat every external harness. That file is scratch. The durable one is a note you own.
Oct 6, 2026 · 7 min readBefore you give an agent write access, take a checkpoint
BCG found 42% of companies expect agents to decide without approval by 2030 and 5% have the controls. The first thing an autonomous agent rewrites is your docs.
Oct 5, 2026 · 7 min readMods made Claude Code's instruction loader replaceable. Your team's conventions never fitted one repo
Anthropic turned Claude Code's AGENTS.md loader into a replaceable mod on 1 October 2026. The question underneath it: where should a team's instructions live?
Oct 5, 2026 · 8 min readStart your team’s brain
Free and open source. No account needed.