@jojobeans1981The Kill Switch
Ok, so i have this dream. or I have this belief in my head that no matter where you are and what industry you are working in or building software for, there is always the need for consideration of safety and security first and foremost. My co-founder and I do not ever feel anything good if our systems replace a person’s job. That’s not what they are meant to do. I build software to run the mundane, to knock out boilerplate at record speeds and accuracy. But there is one thing I have baked into any LLM I use, and every agent I create or adopt gets this rule added: Nothing gets built without a kill switch, one that shuts the system down and gets it back in a person’s control instantly. This has become the floor of our strict build process at FutureEng, and I continue to raise my own bar with each new project I build. How about you guys, is this something I am alone on or no?
3 comments
When you pitch FutureEng to a new client, do you lead with the kill switch, and does it actually help you close the deal?
I did atr first until I realized it made almost zero difference, I have to talk most people into even thinkingt about it. I get the whole " whats the point in automation then" and I usually will run this clip we created at FutureEng showing a catastrophic data leak In a few very clear ways. That usually at least gets them to sit up in the zoom window a bit
It took me almost a year to get this thing to be just what I want. Now the test is how long can it maintain and can it scale, And my co-founder and I always try to laugh those off. we are behind our products 150% and we know we have gone through the very steps I am referring to every step of this project. I come from a long background of teaching and using React js. So I am a huge fan of creating underlying reusable component structures such as custom reusable libraries that can learn from and teach eachother at small scale with enough of these specially trained agents. . so I spent 7 months trying to figure out how to force a pair or more of any combination of LLM, local model, open weight, anything you can connect to, Grok to Gemini, claude to Grok, ChatGPT to Claude. Claude to Claude to Grok with ChatGpt as the referee. But no matter what Until thy figure there way through may fiter system and find a point where they can no longer thing of any challenge to contest the others then they write up the process that was happening during the dissagreement period, Because that is really the important information, they are all trained on the same data but put three of them in a room tell each of them to complete a task that will interfere with anothers and watch how fast they change your plan, Then I scaled it up and hopefuly t to i can creat the coupler for. that was thersult of my seven months. I calmost universal(with very minor configuration if any.) argument machine. Each agent has a specific and i do mean hyper specific purpose, and that is on purpose, any extra task or even a kb of compute adds a point of elevated risk. I have a post converation on linkedin currently about this very think. I need to be able to verify everything my system hands out before a client decision is ever made with it. And that for FutureEng the firms and organizations who we are trying to help, makes it a madatory breaking launch derailing gate that should have no silent roll through or anything like that. my system is like a quarantine as well. if the client or my co-founder or myself were to trip off the system using the killswitches, it triggers a whlole seies of things to protect both any client that might be caught in the process of our intake or if it were a malicious attack . they must be hand released and either hired or rejected by a designated contact at the firm