Design Computer

04/29/2026

𝐏𝐫𝐢𝐯𝐚𝐜𝐲 𝐯𝐬. 𝐏𝐫𝐨𝐭𝐞𝐜𝐭𝐢𝐨𝐧: 𝐓𝐡𝐞 𝐅𝐥𝐚𝐰𝐬 𝐢𝐧 𝐭𝐡𝐞 𝐏𝐫𝐨𝐩𝐨𝐬𝐞𝐝 𝐆𝐔𝐀𝐑𝐃 𝐀𝐜𝐭

AI safety is a critical conversation, but we must be careful not to solve one problem by creating several others.

I am currently opposing the proposed GUARD Act because it takes a blunt, overbroad approach to AI regulation that sacrifices privacy and access for all users. While the intention may be to protect young people, the actual impact would be concerning:

𝐏𝐫𝐢𝐯𝐚𝐜𝐲 𝐑𝐢𝐬𝐤𝐬: Requiring users to upload government IDs or sensitive data just to access AI tools creates massive security vulnerabilities and undermines online anonymity.

𝐄𝐝𝐮𝐜𝐚𝐭𝐢𝐨𝐧𝐚𝐥 𝐁𝐚𝐫𝐫𝐢𝐞𝐫𝐬: Blocking everyone under 18 removes essential tools that students already rely on for homework help, learning, and productivity.

𝐈𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐨𝐧 𝐂𝐡𝐢𝐥𝐥𝐬: Vague rules paired with steep penalties will lead companies to "over-block," limiting features and access even where the law does not clearly require it.

We need AI regulations that address direct harms and promote transparency, rather than mandates that force users to trade their privacy for basic digital access.

03/22/2026

𝗝𝘂𝘀𝘁 𝗽𝗿𝗲𝘀𝘀𝗶𝗻𝗴 𝗢𝗞... 𝗵𝗮𝘀 𝗿𝗮𝗿𝗲𝗹𝘆 𝗯𝗲𝗲𝗻 𝘁𝗵𝗶𝘀 𝗱𝗮𝗻𝗴𝗲𝗿𝗼𝘂𝘀.

I saw this warning modal today while testing an AI agent, Claude in Chrome. It’s a stark reminder for anyone supporting small businesses with open/unclosed IT environments.

If your clients don't have "locked down" computers, they are walking into a minefield if they just click "I understand" and start automating everything.

hashtag hashtag hashtag hashtag

02/13/2026

𝗙𝗼𝗿 𝗴𝗼𝗼𝗱𝗻𝗲𝘀𝘀 𝘀𝗮𝗸𝗲, if you are going to play with AI agents, use a sandbox. Or, better yet a spare computer with a clean install of Linux on a protected network segment.

The screen shown is from a demonstration by a security researcher (Jamieson O'Reilly) who created a "harmlessly" malicious skill called "What Would Elon Do" to show how easily users can be compromised. The researcher distributed this skill through uncurated marketplaces (like the former ClawdHub/OpenClaw registry), and many users installed it without auditing the underlying code.

01/03/2026

𝐀 𝐛𝐞𝐭𝐭𝐞𝐫 𝐰𝐚𝐲 𝐭𝐨 𝐦𝐚𝐧𝐚𝐠𝐞 𝐲𝐨𝐮𝐫 𝐥𝐨𝐜𝐚𝐥 𝐦𝐨𝐝𝐞𝐥 𝐢𝐧𝐯𝐞𝐧𝐭𝐨𝐫𝐲.

If you’re running local AI, you know the struggle: You have 5 versions of the same model, you're constantly editing Modelfiles to get context to stick, or you're manually killing processes to free up VRAM for Stable Diffusion.

I built Ollama Ops (v0.1.0a1) to fix these bottlenecks.

It’s a Python utility that acts as a power-user companion. My favorite feature? The Auto-Optimizer. It looks at your GPU, calculates exactly how much VRAM you have free, and builds a custom model variant with the maximum possible context window "baked in." No more guessing if 32k context will crash your card.

Key Features:

Batch Cleanup: Delete models using wildcards (e.g., rm "*-unused").

The Panic Button: Flush VRAM without killing the service.

Inventory Sort: Find space-hogging models instantly.

Check out the alpha release here: https://github.com/designcomputer/ollama-ops

Note: This project is a fan-made, independent utility for use with Ollama. It is not affiliated with the official Ollama team.

12/31/2025

Happy New Year 2026

For years, I've created my annual Happy New Year video using Blender. This year, I decided to test the latest AI models instead: Veo 3.1, Sora 2, and Wan 2.2.

Veo 3.1 was the clear winner—it nailed the text animation and the audio synchronization perfectly.

Here is the prompt:
"Cinematic HD video, 16:9. A dark night sky. Audio starts with three distinct mortar thumps. Three fireworks trails rise: center, then left, then right. As the center trail peaks, giant 3D gold text "2026" and silver "Happy New Year" flies in rapidly from the bottom right towards the center. The text stops abruptly in the center just as the center firework explodes with a loud BOOM. The text hovers at a slight angle. Immediately after, the left and right fireworks explode with BOOMS. The video ends with the text illuminated by falling sparks and crackling audio."

11/30/2025

Qwen-Image-Edit in ComfyUI transformed this festive Christmas photo by removing all unsightly barcode price tags with a single, concise text prompt: "Remove barcode price tags."

11/25/2025

Hosting Large Language Models locally can feel like a balancing act between turning a server into a space heater or the model into an idiot. Advanced models demand GPUs with substantial VRAM and adequate power to sustain their operation. The easiest, but most expensive, choice is to install multiple GPUs, a large power supply, and a creative cooling solution. The most economic choice is to find a smaller, possibly quantized model, that requires less VRAM but can still perform the AI tasks required. This choice needs more testing and often finetuning because the smaller models just aren't as smart. Choosing the right approach depends on your goals and budget.

08/27/2025

08/25/2025

DeepConf (Deep Think with Confidence) is a new method for confidence-based early stopping in large language models (LLMs), which has been shown to be compatible with inference engines like vLLM. This technique aims to significantly improve inference efficiency for reasoning-intensive tasks.

From what I’ve played with it, DeepConf feels like a leap in the usefulness of smaller, locally hosted LLMs.

https://jiaweizzhao.github.io/deepconf/

08/07/2025

𝗖𝗮𝗻 𝘁𝗼𝗼𝗹𝘀 𝗺𝗮𝗸𝗲 𝗔𝗜 𝗱𝘂𝗺𝗯𝗲𝗿?

I’ve been working with structured data using smaller LLMs that support tool use, specifically SQL, and I’ve noticed something interesting (and a bit counterintuitive). The same model sometimes performs worse with the SQL tool enabled than without it.

Why? Well, I think it’s because the presence of the tool creates a kind of bias. The model tends to jump straight into generating SQL code without fully understanding the task or the data schema first. It’s like when we humans see a calculator on the table and immediately reach for it, even if the problem is simple enough to solve mentally. The result? Outputs that are brittle, unreliable, or just plain wrong.

This feels familiar to me, those moments when we rely on a tool before we’ve thought through the problem. In LLMs, that tendency is amplified. Tools are helpful, but only when they’re used after careful consideration, not instead of it.

Have you ever seen similar effects with other tools, like search, calculators, or code interpreters?
hashtag hashtag hashtag hashtag hashtag hashtag