My F-You To Sam Altman

Summary

The video explores building a powerful local AI workstation as an alternative to relying entirely on cloud AI. Local models can reduce recurring subscription or token costs, keep data within the home network, and continue working without internet access. They do not eliminate every concern: local AI can still hallucinate, and some tasks may still benefit from frontier cloud models. The host estimates that local hardware could handle over 90% of his own needs, especially ongoing tasks where cost and privacy matter more than speed.

Workstation Build and Hardware Analysis

• Case: SilverStone’s ALTA T1, presented as an update to the host’s former favorite case. It has extensive radiator mounting, support for multiple power supplies, and a layout designed to accommodate multiple GPUs. A thick motherboard backplate interfered with some case cable-management grommets, which the builders worked around.
• CPU and motherboard: An Intel Xeon W9-3495X, with 56 cores, up to 4.8 GHz turbo, support for large amounts of system memory, and 112 PCIe Gen 5 lanes, paired with an ASUS Pro WS W790E-SAGE SE. The platform’s PCIe capacity is useful for multi-GPU and workstation configurations.
• GPUs: Two RTX 5090s, one water-cooled and one air-cooled. The host notes that 5090s are fast and support current quantization methods, but their memory capacity and speed may not make them ideal for every AI workload. Depending on the task, a GPU with more memory could be more useful than the fastest GPU.
• Cooling and power: The CPU uses an air cooler because suitable water blocks are scarce and the host says the chip does not require liquid cooling for this build. The GPU uses a soft-tube water-cooling loop. A 3,000 W SilverStone power supply was unsuitable for the available 120 V outlet—it is specified for 220–240 V—so the build instead used a 1,650 W power supply, with room to add a second unit if needed.
• Power connections: The motherboard has several supplemental power inputs because multiple GPUs can draw substantial power through the PCIe slots. The build also uses a custom 12V-2x6 cable arrangement. The host notes that its sense wires are jumpered to signal 600 W capability and that the modified setup has not been proven long-term.

Local AI Software Walkthrough

The build is configured to run a model through LM Studio, then connect it to Hermes, described as an agent framework or “harness” that wraps an LLM, supplies a system prompt, and makes tools and integrations available.

• LM Studio’s defaults worked with little adjustment and automatically used both RTX 5090s. The demonstrated configuration used roughly a 108k context window and an estimated 25–29 GB of VRAM, though the speakers note that larger contexts and multiple models consume more memory.
• Hermes can connect to messaging platforms such as Discord and Telegram, allowing the model to be used through familiar chat apps.
• The demonstration shows the setup responding slowly, even on two 5090s. The speakers emphasize that local AI is often most useful for tasks that can run in the background or be scheduled in advance, rather than for fast, interactive responses.
• Potential applications include preparing a daily news digest, helping with Home Assistant, and using sensor or other household data to trigger notifications and automation. The host suggests that better integration with a real home network would make the system more useful than the test environment.

The Labs team also discusses a DGX Spark running a large model described as around 410 GB of parameters, illustrating how having enough memory to load a model can matter more than having the fastest processor.

Guides and Tutorials Mentioned

1. Local AI workstation build: Hardware selection, assembly, cooling, power requirements, and multi-GPU considerations.
2. LM Studio and Hermes setup: A basic walkthrough of loading a local model, connecting Hermes, and configuring Discord. The speakers say a companion Labs article would cover more of the configuration.
3. Potential follow-up use cases: Scheduled research digests and Home Assistant integration are discussed, but not fully implemented in the video.

Main Speakers and Sources

• Linus: The host and primary narrator.
• Lucas: Explains Hermes and offers setup and hardware input.
• Stephen: A Labs team member who demonstrates the Hermes installation.
• The Labs team: Discusses DGX Spark usage.