Video Summary
Odysseus is an open-source, local-first AI workspace designed for self-hosted chat, agents, research and content tools with privacy in mind.
Installation is doable natively on macOS but often requires troubleshooting (Python versions, tmux, dependencies); Docker is the recommended path.
Key features: local model integration (via Ollama), blind model comparison, autonomous agents, deep-research visual reports, document and image editors, and persistent memory.
Performance and responsiveness depend on the local model used — some smaller models were faster and better for certain tasks; larger frontier models need more resources.
Useful for building a 'second brain' with separate LLM wiki + RAG workflows, but expect some jank and extra configuration for search APIs and advanced features.
Yes — the reviewer installed and ran Odysseus natively on a Mac Studio and configured local models via Ollama. Some larger models require more resources or specific setups, and a few models had connectivity or compatibility hiccups.
Common issues include Python version mismatches, missing tools like tmux, and other dependencies. The repo recommends Docker for easier setup; running natively may need manual troubleshooting or helper scripts (Claude Code was used here to assist).
Notable features are autonomous agents, a deep-research visual report that synthesizes sources, blind model comparison tools, local document and image editors, persistent 'brain' memory, and integration for both local and frontier models.
Odysseus is designed to be local-first with no telemetry by default, so data can stay on your machine. However, integrations like OpenAI or web search engines require API keys and external services, which may send data off-device if used.
It’s promising for private workflows and experimentation, but the project is early-stage and can be janky — useful for inspiration and personal projects, but some features still need refinement for seamless daily replacement of mature cloud tools.
"PewDiePie built and open-sourced his own AI workspace called Odysseus."
PewDiePie's Odysseus is a self-hosted AI platform that integrates capabilities from models like ChatGPT and Claude.
It operates entirely from local machines without relying on cloud services or subscription models, emphasizing user privacy.
The focus of Odysseus is on providing tools for autonomous agents, deep research functionalities, and the ability to draft emails.
"I'm just going to run it natively on my bare metal Mac."
The speaker embarks on the installation of Odysseus on their Mac Studio, finding the process accessible via a GitHub landing page.
Users are encouraged to clone the project, and running it within a Python virtual environment is advised to prevent cluttering the main system.
The installation requires various dependencies, including a database and YouTube transcript API, which aids in the functionality of the AI.
"I've got an old, janky version of Python; it's not PewDiePie's fault."
A problem arises during installation due to a mismatch in the Python version, which necessitates troubleshooting assistance.
During the troubleshooting, it becomes apparent that having the correct environment setup is crucial for smooth operation.
The speaker reflects on the potential and innovation presented by Odysseus, indicating a desire to develop a personalized AI framework tailored to individual needs.
"You want to think about what is mundane, what is monotonous, and what can we outsource to AI?"
The speaker emphasizes the importance of defining tasks that can be automated using AI, thus indicating a shift toward personalized AI solutions.
Odysseus is seen as a step toward enabling users to create bespoke AI systems that minimize unnecessary complexity.
This tailored approach can facilitate deeper engagement with AI technologies while meeting specific user needs effectively.
"I don't have a username and password, so let's just try this and see if it works."
The creator begins by accessing PewDiePie’s Odysseus AI setup and expresses excitement about exploring its features.
Upon trying to log in, they encounter confusion due to the need for a username and password which they do not possess.
They attempt to log in with random credentials but receive an "Invalid credentials" message, leading them to seek assistance from the chat.
"The model that I'm loving and using the most locally at the moment is Qwen 3.6, a 35 billion mixture of experts."
The creator shares insights on their preferred local AI model, Qwen 3.6, highlighting its excellent performance for mundane tasks despite some hiccups.
They also mention using the light version of DeepSeek V4 on their Mac Studio, noting its high RAM requirements.
The creator reveals their intention to improve productivity through building features into their personal "second brain" project.
"I really like the interface. I like how this looks."
After logging into the Odysseus AI, the creator appreciates the user-friendly design and terminal-like feel of the interface.
They discuss adding models, noting the option to integrate local models through an Ollama URL, which adds versatility.
The excitement is apparent as they prepare to set up Ollama for Mac, indicating a willingness to explore local models.
"The idea of Odysseus is that everything is 100% local, which is absolutely awesome."
After successfully installing Ollama, the creator highlights the advantage of having a completely local AI solution with Odysseus.
They express their enthusiasm about the ease of downloading compatible models and the system's capability to diagnose their machine's specifications for optimal performance.
The installation process is straightforward, demonstrating the creator's delight in the setup's user-friendliness.
"This is why I'm testing it and figuring out these janks so that you don't have to worry."
As they attempt to download a local model, the creator faces a minor hiccup requiring Tmux, which prompts them to install it via Homebrew.
They take this moment to engage the audience, asking for preferences between two AI models, Claude and OpenAI’s ChatGPT, to determine which one they should test as the frontier model.
The creator shares a sense of urgency and a light-hearted approach while navigating the issues, reinforcing their commitment to thoroughly understanding Odysseus AI's functionalities.
"By testing this PewDiePie creation out today, I'm gaining a ton of experience on the kind of user experience I want for my own model."
The creator discusses gaining valuable insights on user experience by testing PewDiePie’s Odysseus AI. Their goal is to apply these experiences to the development of their own AI model.
While the local model is downloading, they plan to explore additional model options, like loading a Frontier model.
"We got OpenAI with 126 models. No local models yet because GPT-OSS is downloading."
The creator engages the audience while setting up the AI, explaining the process of adding an API key for testing OpenAI’s models available within Odysseus AI.
The discussion includes the time discrepancy across different regions, with viewers sharing where they are watching from and whether the stream time suits them.
"This repo was only launched 16 hours ago, and it shows PewDiePie has been actively adding updates."
The creator highlights the collaborative nature of the Odysseus AI project, noting that multiple contributors have already engaged with the repository on GitHub.
They mention observing recent commits and how PewDiePie is accepting pull requests, showing an ongoing commitment to developing the AI further based on community feedback.
"Can you get a transcript of this video and tell what it is about? This is impressive."
The creator prompts the AI to fetch a transcript of PewDiePie’s video, demonstrating Odysseus AI’s capabilities to perform web searches and interact with APIs effectively.
They illustrate a keen awareness of the AI’s potential compared to competitors, indicating it may have more integrated features that streamline user tasks.
"Claude Code is just like one operating system. This has pre-built integrations, which I wouldn't have installed by default."
The creator contrasts Odysseus AI with other models, noting that it has added functionality through integrated tools, like the YouTube API.
Although the transcription process is slower than expected, they find the capability impressive and indicative of the model's versatility.
"I plan to use Karpathy's LLM Wiki for my personal knowledge and Retrieval-Augmented Generation for data I don't want to mix in with that."
The creator discusses a strategy for integrating two systems: using Karpathy's LLM Wiki for personal knowledge management, including information about himself and his channel, while leveraging Retrieval-Augmented Generation (RAG) to access external data without contaminating his personal dataset.
He envisions a dual-memory approach, where the AI can consult either memory based on the task at hand, such as creating a specific YouTube outline that reflects his style.
The aim is to maintain a clear distinction between personal insights and broader internet data while allowing the AI to deliver tailored results.
"An LLM Wiki is for your own personal memories; RAG can pull in wider data from the internet."
The creator emphasizes the need for separate systems to efficiently handle different types of information, ensuring that his personal insights remain undisrupted by external content.
He plans to use RAG to sift through data from various sources, including YouTubers and articles relevant to his field, enabling the AI to fetch the most pertinent information without affecting the integrity of his personal Wiki.
This separation allows for a clear focus on relevant content while enhancing the AI's ability to provide informed responses based on a broader context.
"Manu has tried Odysseus from Linux natively but wasn't able to get it running."
The creator shares experiences related to setting up the Odysseus AI on different platforms, noting challenges faced during implementation, particularly on Linux.
He is experimenting with various models, including DeepSeek, and is optimistic about the capabilities of Odysseus AI despite the technical difficulties.
The process involves troubleshooting issues with downloading and integrating models, emphasizing the ongoing development and exploration within the AI landscape.
"I can test on different things that I'm working on, even agentic values."
The creator demonstrates a hands-on approach with AI models, engaging in comparative tests to determine model performance.
His testing includes humor prompts to evaluate which model performs better, revealing differences in response times and quality between models.
He notes that the ability to conduct blind comparisons of various models can provide insights into their efficiency and effectiveness in real-time applications.
"It's a disappointment that I can't get GPT's 20 billion parameters working."
The creator expresses frustration over difficulties in integrating specific models and ensures that he troubleshoots by checking connections and model setups repeatedly.
He invites community feedback and assistance regarding his setup challenges, showing the collaborative spirit in troubleshooting tech-related hurdles.
The ongoing dialogue highlights the complexity of deploying advanced AI systems and the importance of community support in overcoming obstacles.
"This is very, very interesting."
The creator begins by evaluating different ASCII art options, demonstrating a preference for what they consider a more elite design. They also express surprise that a smaller version performs better, being both faster and superior.
The focus then shifts to generating HTML code using the "Black Hole" feature, which results in a substantial amount of HTML text. However, the creator is uncertain about the execution of this HTML within the required environment.
"Let’s get cursor up and running."
The creator decides to utilize a specific coding tool called Cursor to run the HTML code. They create a new folder named "black hole" to organize the project.
After pasting the generated HTML code into two different files (model a.html and model b.html), the creator attempts to find a play button that would allow the execution of the code, realizing that they can compare both models visually.
"Model B is better, isn't it?"
Upon opening both models in the browser, the creator notices that model B presents a more polished and appealing output compared to model A. They express disappointment upon realizing that their previous comparison results might have been lost.
The discussion highlights the importance of comparing models to see which one delivers better performance in tasks like coding.
"Let's recap on what we've looked at in Odysseus."
The creator recaps the installation of Odysseus, sharing insights on how they connected with OpenAI through an API key while also utilizing local model recommendations provided by the system.
They mention having tested local chat capabilities, noting overall satisfaction with the speed and quality but express some frustration regarding the responsiveness of the agent featured in the application.
"This search feature is not that current, it seems."
As part of testing the search functionality, the creator finds that while the AI does respond accurately regarding general knowledge about Odysseus, it struggles with specific details about the Odysseus AI repository.
The creator aims to leverage the deep research capabilities of the AI, which consolidates information from multiple sources, although they encounter challenges in setting up local models effectively.
"Deep research synthesizes sources into a visual report."
The creator highlights the potential of the deep research feature which is designed to compile information into a cohesive visual format.
They experiment with comparisons related to various AI models, emphasizing the feature's utility in generating insightful, organized analysis on contemporary topics.
"Cannot reach model none. Oh dear."
The user begins by setting the parameters for their AI, with rounds set to auto, the search engine as DuckDuckGo, and the model as GPT 5.4 mini to conserve tokens during testing.
They encounter an error stating that the model cannot be reached, indicating the need for minor adjustments in the settings.
"Searching. Round one."
After resolving the initial setup issues, the user prompts the AI to compare Gary Tan's G Brain with Karpathy's LLM Wiki concerning personal second brain applications.
The AI completes the query faster than expected, producing a visual report that summarizes its findings and offers a PDF export feature.
"Self-hosted, that would probably explain it."
Upon completing the first search, the user realizes that the AI's default search engine, DuckDuckGo, is not functioning properly due to its self-hosted nature and the need for an API key.
They troubleshoot this by exploring settings and contemplating switching to another search engine, like Brave, to enhance functionality.
"This is really cool. Now we can see this feature actually working."
After obtaining the API key for Brave, the user tests the deep research capabilities again, which successfully produces results and showcases the AI's effectiveness.
The AI conducts three rounds of research, synthesizing information from numerous sources, showcasing a robust comparison between the two LLMs.
"G Brain is a stronger operational system while Karpathy's LLM Wiki provides a better conceptual framework for a personal second brain."
The visual report generated by the AI compares the strengths and weaknesses of G Brain and LLM Wiki for creating a personal second brain, detailing their operational designs and effective features.
The report indicates an exceptional capability of the G Brain model while asserting that, conceptually, Karpathy's LLM Wiki holds an advantage in simplicity and ease of understanding.
"This is almost like Sublime Text, but inside the browser."
Moving on to document creation, the user creates a new document and tests the AI's ability to assist in making edits, like removing emojis or formatting headers.
They discover that the AI functions as a versatile tool for enhancing writing, operating similarly to advanced code editing tools, highlighting the potential for seamless integration of AI in creative workflows.
"Very interesting concept; the AI assists in expanding your writing."
The creator is observing the initial processing time for the AI model to flesh out content, expressing curiosity about its effectiveness.
The process results in the AI turning a short community post into a full blog, showcasing its ability to expand on brief content.
"It has taken everything that I wrote about and made it into a blog post."
The AI is capable of generating well-structured and coherent blog posts from casual notes and bullet points.
The creator notes the absence of unnecessary characters, highlighting the quality of the content produced by the AI.
"Remembered, you like clear, direct, practical blogs with strong headers."
The memory feature allows the AI to retain user preferences regarding writing style and personal choices, such as breakfast preferences.
Users can query the AI about what it knows, which leads to a demonstration of the AI's recall abilities regarding user-defined inputs.
"It seems to forget the details of our chat."
The creator experiences some challenges in obtaining precise information from the AI, indicating limitations in its memory recall for small models.
There is a sense of frustration with the AI’s speed and ability to engage in natural conversation, leading to a preference for improved integration with more powerful models.
"That's a to-do, and I can tick it off the list when I've done it."
The AI has a built-in to-do list and reminder system, allowing users to manage tasks effectively.
Users can add tasks in a visually appealing manner and interact with the system for reminders.
"There is a calendar integration for quick add tasks."
The AI allows for quick task additions and integrates with calendar features, albeit with some additional configuration required.
The creator notes the potential for future improvements and functionalities that could enhance productivity through better integrations.
"It works on mobile as a progressive web app as well."
The Odysseus AI can be accessed on mobile devices, offering users flexibility and convenience by functioning as a progressive web app (PWA).
This capability allows users to utilize the AI's features seamlessly across devices.
"Upload a photo, drag and drop to get started, edit, image editor, new canvas."
Users are able to upload images easily through drag-and-drop functionality, making the editing process straightforward and user-friendly.
The platform includes several image editing tools, aimed at creating professional-quality graphics such as YouTube thumbnails.
"I want to run everything locally and I want to clone Photoshop."
The creator attempts to evaluate whether Odysseus AI effectively serves as a clone of Photoshop, highlighting PewDiePie’s intention to shift away from conventional software.
Experimentation with various editing functions helps gauge the AI's capabilities against established platforms like Adobe.
"I can brush and do things."
The AI offers multiple editing tools such as brushes, lasso, and eraser, allowing users to manipulate images and create unique designs.
Despite not fully matching Photoshop's capabilities, some features are functional, and the user demonstrates excitement about the potential for design.
"You've got to install a certain library for it to work."
The creator encounters issues with certain features not functioning properly out-of-the-box and realizes the need for additional installations, such as specific libraries for background removal.
Highlighting the experimental nature of the platform, there are limitations to some features, compromising immediate usability.
"Will I use this for inspiration for what I'm building? Yes, 100%."
The creator expresses a commitment to leveraging Odysseus AI as a source of inspiration, even if it may not replace mainstream tools for daily use.
They emphasize a philosophy of mixing different AI projects to tailor solutions that meet their specific needs, reflecting best practices in AI development.
"We've been working on PewDiePie’s own harness, which is Odysseus. It's your own AI workspace."
The video showcases the evolution and responsiveness of PewDiePie's Odysseus project, noting an increase in downloads and community engagement over time.
The creator reflects on their experience with various AI functionalities, indicating that while Odysseus has room for improvement, it offers promising features worth exploring for personal projects.
"It interactsively pops up a terminal, which runs Claude Code, resumes everything."
The system allows for an interactive terminal that appears as a window to run Claude Code seamlessly.
Users can manage multiple terminal instances, with visual indicators showing the status of each instance, such as an orange light for active processes and a green light for those that are complete.
"My idea is to remove myself as the bottleneck and spin up multiple Claude Codes."
The goal is to streamline operations by minimizing personal involvement in running the AI, allowing for simultaneous processing of multiple tasks using Claude Code.
This setup permits quick evaluation of ongoing tasks and their status, enhancing overall workflow management.
"I’ve been playing with PewDiePie's Odysseus harness, which allows Frontier model AI and also local AI."
The video discusses the installation and use of PewDiePie's Odysseus, which incorporates local AI models alongside the Frontier model AI.
The speaker successfully set up the GPT-OSS 20 billion model on their M1 Mac, demonstrating the capabilities of running such complex AI locally.
"I want to make it a progressive web app so I can install it on the home screen of my mobile."
The current version of the Tank platform, while functional, is noted to be somewhat janky, as it is still in an initial prototype stage.
There are aspirations to develop a mobile-friendly version of the application, potentially as a progressive web app for enhanced accessibility and usability on smartphones.
