Six Open-Source Projects for Local AI, Agents, and 3D Design

Matthew Berman tours tools for running and training models, giving agents new skills, coordinating work, browsing, and generating 3D meshes.

MBMatthew Berman登壇者 · Forward Future
チャンネル
Matthew Berman
公開日
2026年8月17日
動画
9:15
読了時間
3 分
スライド
4
原文
英語
YouTube で開く
スクロールすると動画は右下に小さくなって再生を続けます。字幕は閲覧言語に合わせて切り替わります。
閲覧言語2 言語で読めます
この記事はまだ日本語に翻訳されていないため、英語の原文を表示しています。
要約

30 秒でわかる

Matthew Berman presents six open-source projects: Unsloth, Diagram Design, Obsidian Skills, Buzz, ego-lite, and Modly. His demonstrations range from local model training to agents working alongside people and image-to-3D generation. The video is sponsored by Hostinger, which hosts the Hermes agent used in the diagram demonstration.

  1. 1Local model toolsBerman describes Unsloth as a point-and-click way to run and fine-tune models, with tools and an agent interface.
  2. 2Diagrams as an agent skillDiagram Design supplies diagram patterns to compatible agents. Berman demonstrates it with Hermes hosted by the video's sponsor, Hostinger.
  3. 3People and agents in one workspaceBuzz gives agents accounts alongside human teammates and uses a signed event log for messages and workflow activity.
  4. 4Browser access and 3D assetsego-lite shares logged-in browser state with agents; Modly generates a 3D mesh from an image on local hardware.

ポイント

  1. Unsloth combines local model inference and training with an agent interface.
  2. Diagram Design and Obsidian Skills give compatible agents instructions for visual diagrams and Markdown-based knowledge work.
  3. Buzz places people and agents in shared channels and records activity as signed events.
  4. ego-lite lets agents automate a browser using a user's logged-in state; Modly converts images into 3D meshes locally.
全文

チャプターごとに整えた全文を読む

文字起こしのみ
01

Unsloth: models on your own machine

Berman introduces local inference, training, and an agent interface. He highlights a graphical fine-tuning workflow.

Here are six open-source projects worth exploring. The first is Unsloth, which began by making local large-language-model fine-tuning easier and has grown into a tool for inference, training, and more. It is available across Windows, macOS, and Linux, with support for a range of models and hardware.

Unsloth now also has an agent interface resembling a chat application. Berman highlights web search, tools, MCP, memory, and remote access alongside local model execution. A model running locally does not mean every connected feature stays offline: web searches and remote access involve network connections.

For someone new to fine-tuning, the appeal is a point-and-click interface that guides the user through settings rather than requiring them to write code. Berman says the tool also supports image and video models. Actual model support, performance, and training capacity depend on the version and hardware used.

02

Diagram Design

A diagram skill gives agents visual patterns for flowcharts and other diagrams. Berman demonstrates it with a cloud-hosted Hermes agent.

Diagram Design examples
スライド 1 / 4Diagram Design examplesThe project illustrates the kinds of diagrams an agent can produce, including flowcharts and architecture diagrams.

Diagram Design gives an agent patterns for producing clearer diagrams, including flowcharts, architecture diagrams, state machines, timelines, and quadrants. It can be installed as a skill or plugin for compatible agents. Berman demonstrates it with Hermes, asking the agent to make a diagram of its own architecture.

In the demonstration, Hermes runs on Hostinger, the video's sponsor. Berman supplies the project's GitHub URL for installation and then asks for an architecture diagram. He presents cloud hosting as a way to keep background processes running without leaving his own computer on.

03

Obsidian Skills

The skills connect agents to Markdown-based Obsidian notes. Berman describes using a vault as a knowledge base.

Obsidian is a note-taking application built around Markdown files. Berman likes the way its simple building blocks can become a wiki or a personal knowledge base, and says he syncs his notes across devices. Obsidian Skills gives compatible agents instructions for working with those notes and their features.

Because agents already work well with Markdown, Berman sees the vault as a potential knowledge base or 'brain' for an agent. Notes can be kept locally, while synchronization is also an option. Giving an agent access means deciding which notes and connected services it should be able to use.

04

Buzz

Block's open-source workspace puts agents and people in the same channels. Workflows and signed events extend it beyond chat.

Buzz collaboration workspace
スライド 2 / 4Buzz collaboration workspaceBlock describes a shared workspace for people and agents.

Buzz is an open-source, self-hostable collaboration tool from Block that Berman compares to Slack. Its distinctive idea is to make agents first-class participants alongside people in channels. The familiar interface includes replies, threads, reactions, and channels, while workflows and automation are built into the workspace.

Buzz records messages, reactions, workflow steps, reviews, and other activity as signed events using a Nostr relay. Berman likes having a common identity and audit trail for actions by people and processes, and says users can bring different hosted or local models. Self-hosting and an event log are design features, not a guarantee of security; deployment and permissions still matter.

05

ego-lite and Modly

ego-lite lets an agent operate a browser with existing login state. Modly turns an image into a 3D mesh locally.

ego-lite browser automation
スライド 3 / 4ego-lite browser automationThe project describes giving agents access to an existing browser session.

ego-lite is a browser automation project designed to let agents use an existing logged-in browser state without interrupting the user. Berman demonstrates invoking the browser from a coding agent and shows the cursor moving through a page. Since an agent can inherit access to signed-in accounts, users should consider the permissions and actions they are delegating.

Modly image-to-3D project
スライド 4 / 4Modly image-to-3D projectThe project describes generating 3D meshes from images locally.

The final project, Modly, turns an image into a 3D mesh using local AI. Berman suggests the result could be useful for game assets or 3D printing and highlights support for several operating systems and GPUs. Hardware requirements and the quality or readiness of a generated mesh depend on the setup and the input, so he encourages viewers interested in 3D modeling to try it themselves.

同じチャンネルの記事

スライド 6 枚25:50EN
MBMatthew Berman· 2026年10月24日Kimi K3.1 Pro Tested: Strong 3D Demos, Uneven Results and Competitive PricingMatthew Berman · Forward FutureMatthew Berman tests the open-weight model on coding, 3D rendering and visual explanations, then compares benchmark results and prices.Open-weight modelsBenchmarksAI cost & performance+1
スライド 2 枚32:20EN
MBMatthew Berman· 2026年10月23日Who Is Local AI For? The Costs, Trade-offs and a Hybrid FutureMatthew Berman · Forward FutureMatthew Berman weighs data control and independence against hardware cost, setup complexity and inference speed.Local inferenceOpen-weight modelsAI cost & performance+1
45:08EN
MBMatthew Berman· 2026年10月21日Sam Altman on AI Discovery, Consciousness, Compute and a Human-Centered FutureMatthew Berman, Sam Altman · Forward FutureIn a conversation with Matthew Berman, Altman discusses scientific progress, personal assistants, recursive self-improvement and the importance of human agency.AI for sciencePersonal agentsConsciousness+2
0:32EN
MBMatthew Berman· 2026年10月21日Sam Altman on the Continuing Demand for Software WorkSam Altman, Matthew Berman · OpenAIIn a short interview excerpt, Altman says demand for software may be larger than expected even as AI changes how it is made.Future of WorkSoftware engineeringコーディングエージェント
スライド 5 枚12:14EN
MBMatthew Berman· 2026年10月14日Gemini 4 Argon Tested: 3D Demos, a Spreadsheet Prototype and Benchmark ResultsMatthew Berman · Forward FutureMatthew Berman tests Google’s model on games and visual projects, then compares reported coding and agent benchmarks.GeminiBenchmarksコーディングエージェント+1
スライド 5 枚10:36EN
MBMatthew Berman· 2026年10月13日Meta Muse Wonder Tested: Strong Benchmarks, Mixed Demos and Quota FrictionMatthew Berman · Forward FutureMatthew Berman tests Meta’s models on browser demos and reviews their published benchmark scores, availability and cost.モデルと学習BenchmarksAI cost & performance+1