So, I'm going to give you everything that I learned after working at Goldman Sachs, one of the biggest financial firms on the planet. These are principles I still follow to this day in everything that I do with AI. A company like this doesn't have the luxury of making mistakes, and one wrong data entry or hallucination can cost them millions of dollars, or even worse, their reputation. So, if one of the biggest and most trusted financial firms on the planet follows these principles, it's because they work. And even if you're not working at a big firm, these principles will completely change how you use AI and the results that you get out of AI. So today I'm going to share the five AI principles I learned from Goldman Sachs which I call vault and show you how to apply them to your AI projects. So let's get into it. Okay, so V is to verify the output. And just a reminder, this is the framework I created from the principles I learned while working there. And these principles apply whether you're using claude code, codecs, or whatever new tool comes out next month. So Verify, I think a lot of us, myself included, can get a little too comfortable with AI because it gives you something that looks finished and it will deliver it to you super confidently. It's formatted perfectly and it probably took about 30 seconds. But something that looks finished and something that's actually correct are obviously two very different things. So Marco Agenti, Goldman's CIO, has made this distinction between the reasoning of a model and its final output. Basically, the way a model breaks down and analyzes a problem can still be useful even when the final answer it gives you is wrong. So you can use the breakdown to get value from it, but you still need to verify the result. And this actually starts before the AI even generates an answer. So my role at Goldman was in business intelligence. And every business intelligence analyst in my role had to pass something called data school. We had to understand how data gets pulled in, how it gets cleaned, how it gets updated, and how you turn all of that into something that people can actually use to make a decision because every report, every dashboard, every automation, and now every AI system is only as good as the data that's actually powering it. And your company's data is probably one of the biggest advantages that it has because nobody else has that exact information. They don't have your customers, your sales history, your internal processes, your support tickets, all of that kind of stuff. It's gold. But if that data is outdated or duplicated or messy or just wrong, then plugging it into AI doesn't magically fix it. It just gives you the wrong answer faster. So when I use AI tools, I like to build verification directly into the processes. You could give the model a prompt like, "Hey, you know, before you give me the final answer, recheck every number and factual claim. Cite the source for each one and flag anything that you're not fully 100% confident about." And you don't have to manually verify every single word that it gives you. If you're creating a report, maybe there are only two or three numbers that could actually change the decision. So have the model site those numbers and then you can spot check them yourself. And obviously for higher stakes projects you can even build a separate AI reviewer or multiple you know like a team of AI agents that are just there to review. They'll check the first output before you see anything and then the whole system can iterate once again before you see anything. And that's an additional check. But important claims still need original sources, deterministic tests or human review. So verify the data going in, verify the important outputs coming out and build review into [music] the process. But once you can trust the inputs and the outputs, you still have to decide whether AI should be in the workflow at all. So moving on to a augment, which means augment don't replace. Now my full-time job at Goldman was basically building automations, reports, dashboards, and systems that made teams more efficient. [music] And across the teams I worked with, there were people doing some version of that work. None of this was some brand new AI initiative. You know, Goldman had been building automated data and risk systems for decades. and teams across the firm were doing this kind of work before the current wave of AI agents hit the market. Which means there were already a crazy amount of problems worth solving before anyone started talking about AI agents. There were manual reports being built that took hours. You know, data had to be moved between systems. Teams were waiting on information from other people. People would enter something incorrectly with a fat finger or skip a step or or two people would follow the same process in two slightly different ways. [music] Because at the end of the day, humans are inconsistent. We get tired. We miss steps. comm

So, I've been using Deep Sea Carness for the past week now, and I found a lot of things that I really like about it, and a lot of things that I also don't really like about it. So, today I'm just going to give you guys my honest thoughts, break down what I've been using it for, how it feels, and I'm also going to be specifically talking about it in comparison with Cloud Code since I know that's on a lot of you guys' minds. So, let's not waste any time and just get straight into this video. Okay, I'm going to try to do this in about 60 seconds. If you've never heard of Deep Seek Harness, what is it? So, Deepseek has different models, right? We had a big Deepseek moment last year, Deepseek R1, and then we have like DeepSeek V4 Flash and Pro right now. Decent models, pretty cheap, but we have a harness now from Deepseek, which is open source, and it's free. So, as we know, we have something like Codeex, which is the harness, and the AI models that you can use inside Codeex natively are GBT 5.6 Soul, GBT 5.6 Luna, things like that. We have Cloud Code, which is a different harness, and the AI models we natively use inside of Cloud Code are Cloud Opus, Cloud Fable, things like that. And now we have a Deep Seek harness which has other AI models inside. I mean by natively we have the Deepseek models, but you can also add any of the other models. You can add Opus in there. You can add GBT in there. You can add Kimmy. You can add Meta if you're crazy enough to do it. You can add basically whatever you want in there. Now, yes, you can also do that with these harnesses as well, but they're not exactly like made for it or promoted as much. But also things like OpenClaw, that's an agent harness. Also something like Hermes Agent, that's an agent harness. So it's not groundbreaking new tech. It's just a new harness. But what's really cool about this one is because the harness is open source, we can like fully customize it. So with a closed source harness like Cloud Code, it has its basic kind of like prompting rules and it has its tool calls and it has its loops and it has its agentic harness that we can't touch. We can basically look at it like a car that we can sit in, we can drive it, we can change out the model of the engine, but we can't move things around. Whereas with our deep sea harness, the open source harness, we can switch out the seats. We can switch out the steering wheel. We can switch things out. We can change the way the agent actually behaves under the hood, which is pretty cool. And that's why a lot of people are freaking out because if you go back to the site, it literally says everything is a plugin. And so by that, they don't mean just like skills and like MCP servers. They mean every capability is a plug-in. Tools, skills, sessions, sandboxes, storage, loops, the UI. This is all a plug-in environment. And so I will just be honest about that real quick. In this video, I'm not going to be talking about plugins that I installed from other people or plugins that I've built because I wanted to see out of the box what this thing felt like. And I will say about plugins, be careful because if people are giving out all these open source plugins, you don't know what might be in there. So, be smart. Have Cloud Code or have Codex review the plugins before you ever install anything like that off the internet. Okay, so here are some of the things that I want to talk about today when it comes to our DeepC Carnis. So, the first thing is let me just open up the UI and just show you what it looks like real quick. So, it looks like a chatbot, right? It looks like kind of like the Chag app or whatever you might be familiar with. It's not going to be anything new or super confusing. You can have products on the left. You can manage your different sessions, things like that. Pretty normal stuff. And if you want to test this out, literally what I did is I gave the GitHub link to Codeex and I said, "Hey, help me set this up." It opens this up in a local host. So, this is completely local right now. And then you're pretty much all good. All you have to do from there is you're going to go to your settings. You're going to go to models and you're going to put in some sort of key. I put in an open router key which is why now I can choose between all of these different models that are available to me on open router. Unfortunately you can't by default use like a cloud subscription or a codec subscription. I have seen some plugins where people say that it works. I haven't myself tested it but if you are interested in that then definitely look into that. But default you're going to be charged here by token you know API billing. Anyways that's what it looks like. We'll look at some other stuff in this UI in a bit but let me just go back over here to this screen. Okay, so let's talk about output quality. This entirely depends on the model and the effort, of course. I mean, you have a harness, so that can do so much, but at the end of

So, I just changed tax season forever. I just built an agent for an accounting firm that reads a client's tax organizer the second that it comes in. It maps every number into their workbook [music] and checks it against last year before human ever can look at it. So, I just fixed that the kind of mistake that shows up every single season. An [music] advertising expense entered at $12,000 when last year it was 1,200. That one extra zero was caught in seconds and it drafted a client an email asking to confirm the number. So, it's checking three things. Any number 10 times bigger than last year, a negative where there shouldn't be one, [music] and an ending cash balance that does not tie to the bank statement on file. And it never sends anything on its own. First, it'll draft and then from there a person will approve and nothing is leaving the building without somebody clicking send and approving. And it won't guess either. Anything under 90% confidence, it's going to stop and flag a human because a wrong guess in February is a wrong tax return in April. Comment need and I'll send you the entire step-by-step video.