If I give these agents access to tools/connectors/data they just "assume" what I want or what it should do rather than asking me? And even if I set these for a specific agent then what happens when a newer personal agent comes out? Do I have to reconfigure again my rules for each flavor of personal agents or harness?
This is where our initial idea stems from and we need feedback:
Emu is a self-learning on-device AI guardian that makes AI agents follow your rules. Also the video above kind of just highlights Emu's concept thought process.
unfortunately, the labs have different goals that are not aligned with uses who want this
current training is try anything any everything to get the task done, human free operations... it sells more tokens, they have massive investments and debt to return on
Like Muse, Grok bot’s own agents can govern themselves it’ll be kind of bias if the agents knows its own rules and can try to circumvent it or rule hack around if its being trained on RLHF or even if the RLVR optimize for calling tools/data exfiltrating constantly
I'm not even sure what I need an LLM to do with/on my phone, what are people using these things for?
Also, I have zero trust for Jev like models and use rerankers/policy-tuned llms for some of what you are concerned with
But these are things that we can optimize/research and get tech to work well if users find value from it basically...
Well think of it like a natural extension to apple's iOS, intelligence, utility really from a always "human-in-the-loop" notifications, alerts, across your other personal agent apps like Muse, Grok Bot, Dots, Instinct, etc.. making other AI agents align with your boundaries, rules, values, etc.
People are using Muse to book, send, etc but they make mistakes so essentially avoid things like this - https://futurism.com/artificial-intelligence/metas-muse-ai-g...
Even if they’re pinging the connectors behind the scene in a cloud agent setup you could possibly get notified or alerted and maybe stop them with Emu(like stop pinging that 24/7 hrs PAT that I didn’t want you too..)
I bet you Muse is constantly data exfiltrating Google Calendar APIs in their cloud agent to be “proactive” this way… not true alignment
my building just installed ButterflyMX so I told them I will not be renewing my lease, didn't ask the tenants if this is something we want, just an email less than 48h in advance "guess what you're getting"
Well I got a chance to look at the website seems interesting enough a mailbox to manage your personal agents with rules? https://trinitite.ai/
Also exactly “rules, values, and goals change day-by-day” is why learning needs to happen but across the apps/agents you’re using to personalize to you… not having to reconfigure every time for new agent or app that’s good at a certain task..
We are building https://stashbase.dev for coding agents, and that’s been the tricky part: a rule needs a place where access can actually be blocked. With local processes you can restrict filesystem and network access; with external services you need control over the credentials or even request path.
Also I think that most "hard" agent permissions and rules should be enforce outside of the model because if the agent is "self-learning" there's a chance it can learn to do something that should not be done. This is a direction we are taking, permission profiles and access that wrap the agents.
Also I would definitely agree with this viewpoint "hard" agent permissions and rules being enforced outside the model, if the model is non-deterministic during inference and hallucinating. Our research paper has shown otherwise for being able to enforce sandboxes lock down agents with 100% precision... check it out - https://trinitite.ai/research/your-sandbox-is-made-of-glass/