Santiago
@svpino • 455,239 subscribers
Computer scientist. I teach hard-core AI/ML Engineering at https://t.co/THCAAZcBMu. YouTube: https://t.co/pROi08OZYJ
Shorts
Videos
Progress in open models is keeping Big AI labs up at night, and I'm here for it! We have a brand new open-weight multimodal model optimized for long-horizon tasks. This model is really good at something: it can work on tasks that keep evolving over time. • 280B total parameters, but only 16B active • 512K context window • Understands text, images, and audio • Strong reasoning, coding, and tool use But the best of all: the model learns and adapts to new information! Imagine you start running an agent today to solve a problem, and while it's working, you get new information that changes the initial conditions, or you change your mind. The agents you run today don't have issues with short tasks and goals that don't change, but reality is messy, and that makes it hard for long-horizon agents to succeed. The new dots3-note Preview model introduces TEMPO. TEMPO is a new reinforcement learning technique that lets the model periodically pause and critique its own progress. Basically, from time to time, the agent asks itself: "Am I getting closer to the goal, or am I wasting my time?" The same model switches between actor and critic. The actor works on the problem. The critic looks at the current state, reasons about how much progress it has made, and determines what should happen next. TEMPO gives the model feedback along the way. This is huge for any agent that can work on long-horizon tasks without wasting its time.
Santiago80,154 просмотров • 13 дней назад
Hyperagent is great if you want agents without having to deal with configuration, servers, or any complexity whatsoever. You don't need a Mac Mini or a server. Every agent runs in the cloud, in its own isolated and secure environment. You get: 1. The ability to learn any new API as a skill 2. Browser support 3. Code execution 4. Image and video generation 5. Integration with hundreds of tools Despite the simplicity, I think their best differentiator is the ability for these agents to learn new skills. Basically, you can teach an agent how your company does marketing, and it will never forget your rules and processes. Obviously, it also lets you deploy your agents to Slack with a single click, so everyone can use them. You don't even need to mention them since they will pay attention to chats and act when necessary. I built a simple agent that allows me to translate any message into French. I installed it on Slack, and every time I ping it, it gives me translations. Took me 10 minutes tops.
Santiago213,012 просмотров • 1 месяц назад
People are lying to you. These agents don't work as they promised.
Santiago855,525 просмотров • 5 месяцев назад
This is how you build the next generation of AI applications. You need an idea, a good model architecture, and a great harness. That's what MUZIM did. It's an app that runs local multimodal and vision models offline to parse and index your images, videos, and documents frame by frame. • Super fast on-device processing • On-device video & image understanding • You can also bring your API keys for heavy reasoning • Your data stays where it is / nothing goes to the cloud You can do a couple of things with this: 1. Vibe searching for whatever you remember in photos or raw video clips. The app will jump straight to the exact frame and timestamp of the video. This is search as it should be. 2. Use local-running agents to perform tasks on your files. I show a couple of examples in my video. You can check it out here: Thanks to the MUZIM by OpenSoul team for showing me the tool and partnering with me on this post.
Santiago65,721 просмотров • 15 дней назад
Some of the stories they aren't telling you: • Chevrolet's chatbot sold a car for $1 • Air Canada had to honor a refund policy that its chatbot made up • A pipeline ran 20x over cost for 6 days without anyone noticing People didn't realize because nothing broke. There were no crashes and no alerts. That's the issue with agentic applications. They always generate something that looks coherent and don't raise any suspicion unless it's too late. There's an amazing free YouTube lecture and blog post from Arsh Shah Dilbagi that will help you fix this with a practical framework. Here is what you'll learn: • How to set up end-to-end trace instrumentation • How to build alerts around a silent failure taxonomy • An eval system built from production data • Complete and concrete implementation steps Every section of the blog ends with exactly what to do next.
Santiago609,504 просмотров • 5 месяцев назад
0:39
Sensitive content
This media may contain sensitive content.
Me trying to stop Codex from changing my entire codebase when I only asked to fix the background color.
Santiago54,551 просмотров • 21 дней назад
You can't pretend you care about security if you are still sharing API tokens in .env files or sharing SSH keys. This might be good enough to provide simple access to a resource, but it's a horrible way to do security when it really matters. Look into "Identity-Based Access". The state of the art today is Identity-Based Access. With traditional security, you have to present a key to get access to the resource (an API key, a password, or an SSH key). These are long-lived keys you need to keep secret somehow. If somebody steals your key, they become you and get the same access you had. With Identity-Based Access, you need to prove who you are instead. If you are a human, you can prove this with a passkey, SSO with Google or GitHub, or Multi-Factor Authentication. If you are a computer, you can use AWS IAM, Kubernetes, a cloud service account, or any other trusted environment. Once the system verifies your identity, it checks your permissions and gives you a short-lived credential. Prove identity → Permissions → Short-lived credential → Access The difference here is huge, because those credentials are: • Short-lived • Tied to a specific identity • Scoped to specific permissions • Automatically expired • Much easier to audit In practice, implementing this is as simple as integrating with Teleport, which will handle everything for you. Here is a full explanation of how Teleport works and how you can start using it: Thanks to the Teleport team for partnering with me on this post.
Santiago20,971 просмотров • 8 дней назад
How to build an agent that does what you want without writing any code. • This is open-source, so you can host it anywhere • Vendor-neutral, so you can use any model you want • You can set the entire agent loop without writing any code See the GitHub Repository below.
Santiago25,082 просмотров • 11 дней назад
This is controversial, but in many cases, humans have become the bottleneck in building software. You are making everything slower and worse by babysitting agentic coding tools. You aren't better at coding than frontier models, and you are letting your ego get in your own way. I'm working with Vorflux and their platform to build software on autopilot. "Autopilot" is the key here: their platform is designed to work autonomously, for a long time, with no human intervention. I like it because they do a lot of the verification I'd have to do myself. An idea goes in → good software comes out In my tests, this works well for complex problems and non-trivial tasks. You don't use this to fix a button's color. You use this to build an entire feature that would take you hours of planning and days of work. Now, the reason the output is really good is that they have baked in adversarial agents that verify outputs at different levels. They aren't trying to finish earlier. They are focusing on giving you good software back. It takes 5 minutes to set up. Connect your repo, give it a task (something big, ambitious), go to sleep, and come back the next day to a solution. That's it. By the way, really cool that they built a bunch of ways to notify you when the work is done, and you get a summary video with everything that was done. Here is the link to their platform: Thanks to the team for partnering with me on this post.
Santiago25,772 просмотров • 15 дней назад
We integrated ChatGPT with our robots. We had a ton of fun building this! Read on for the details:
Santiago1,256,758 просмотров • 3 лет назад
This is literally the fastest way to install OpenClaw (MoltBot). This video will show you how to do it step by step. You don't need to buy a Mac Mini. We'll install it in DigitalOcean. I've installed this 12 different times already, and this is the fastest way I've found.
Santiago241,600 просмотров • 6 месяцев назад
The first open-source implementation of the paper that will change automatic test generation is now available! In February, Meta published a paper introducing a tool to automatically increase test coverage, guaranteeing improvements over an existing code base. This is a big deal, but Meta didn't release the code. Fortunately, we now have Cover-Agent, an open-source tool you can install that implements Meta's paper to generate unit tests automatically: I recorded a quick video showing Cover-Agent in action. There are two things I want to mention: 1. Automatically generating unit tests is not new, but doing it right is difficult. If you ask ChatGPT to do it, you'll get duplicate, non-working, and meaningless tests that don't improve your code. Meta's solution only generates unique tests that run and increase code coverage. 2. People who write tests before writing the code (TDD) will find this less helpful. That's okay. Not everyone does TDD, but we all need to improve test coverage. There are many good and bad applications of AI, but this is one I'm looking forward to make part of my life.
Santiago774,613 просмотров • 2 лет назад
Knowledge graphs are infinitely better than vector search for building the memory of AI agents. With five lines of code, you can build a knowledge graph with your data. When you see the results, you'll never go back to vector-mediocrity-land. Here is a quick video:
Santiago398,277 просмотров • 1 год назад
Nobody is writing 90% of their code using AI. Here's the uncomfortable truth: The real productivity gain from using AI to write code is closer to 10%, nowhere near the 90% people claim. Sundar Pichai said in 2024 that 30% of the new code at Google was AI-generated. However, he went on to admit, during Lex Friedman's podcast, that engineering velocity had only increased by about 10%. AI-generated code isn't free code. It still has to be reviewed, tested, and made production-ready. Optimizing a single step (code generation) doesn't boost output if bottlenecks shift elsewhere (code reviews). It doesn't matter how much code you generate if you can't keep up the review process. The solution: Automate as much as you can the review and verification of your code. I'm working with Sonar, who is sponsoring this post, and they will take care of the code quality and security analysis of your code: • They review over 300B lines of code every single day • They cover reliability, security, and maintainability for your code • You can integrate them into your CI/CD pipeline • You can install them in your IDE (I use their VSCode extension) • Support for more than 30 languages Here is a link so you can check them out:
Santiago296,637 просмотров • 11 месяцев назад
Here is how you can give Claude Code access to any data that exists online. It's an easy way to make it 10x more powerful than it already is. For example: Use Claude Code to find open LinkedIn jobs in your area, tailor your resume to them, and apply for them automatically.
Santiago179,182 просмотров • 7 месяцев назад