Hermes Desktop
Nous Research announced a public preview of Hermes Desktop, bringing Hermes Agent to a native interface on a local machine.
The native app can expose local files, tools, notifications, and long-running activity outside the terminal. I have not had a chance to fully evaluate it. My attempt to get the desktop app to use the “remote” mode did not work well with the early releases. I have not had a chance to go back and diagnose or check if the newer builds have resolved any of the early issues.
Skills as a supply-chain surface
A user describes adding NVIDIA’s open-source SkillSpector as a GitHub Action that flags prompt injection, data-exfiltration behavior, and other supply-chain risks in submitted skills.
A skill may include instructions, scripts, templates, and references that shape an agent’s behavior. Installing one is closer to adding a dependency than copying a prompt. SkillSpector’s usefulness will depend on the checks it implements and the dangerous behavior those checks can recognize statically. Sounds like something that has no obvious downsides but can at least provide some measure of protection where almost none exists today other than manual verification.
Measuring AI-assisted AI development
Anthropic reported internal data showing Claude accelerating AI-development work and connected the findings to questions about increasingly autonomous successor-model development.
It will be interesting to find out which research and engineering tasks were included, how they measured acceleration and areas where they did not see improvements.
AutoScientist and a two-day training claim
Adaption AI describes AutoScientist as a system that can move from unstructured data through model training in two days while reducing manual pipeline work.
The two-day figure comes from Adaption AI’s announcement. The described workflow appears to cover data preparation and model training, though the post does not say enough about experiment selection or evaluation to judge those parts.
X-Token combined lessons from models with different tokenizers
The X-Token paper looks at a practical problem with training a smaller model from the output of larger ones: the models do not always split text into tokens the same way.
X-Token adds a translation step so the smaller model can still learn from each larger model. That could allow one small model to learn from several model families instead of being limited to teachers built with the same tokenizer.
Agentcookie syncs authentication state between Macs
Agentcookie continuously copies Chrome cookies and command-line authentication state from a daily-use Mac to a second Mac over a Tailscale connection. It does not require a hosted intermediary for that data path and works with agent-oriented tools including Printing Press and Last30Days.
The setup requires one pairing, then keeps the agent Mac’s sessions synchronized without logging in to every site and CLI again. The announcement shows it alongside Printing Press and Last30Days, which is a pleasantly specific multi-Mac workflow rather than a generic remote-agent demo.