DIY Artificial Intelligence

Good Luck, Have Fun, Don’t Die is the latest science fiction black comedy movie from Oscar-winning director Gore Verbinski.

The movie marks Verbinski’s return to directing, nearly a decade after his last work. The story revolves around a crucial nighttime mission: saving humanity from the impending apocalypse caused by rogue artificial intelligence and the dangers of social media “brainrot” algorithm.

In today’s landscape, I see a lot of nerdy professionals thinking of proprietary AI models as rogue access points to their network, so they are trying to build their own AI solutions.

Proprietary vs Open Source AI: choose the right solution

Choosing between open-source and proprietary AI frameworks goes beyond simple technical preferences; it shapes an organization’s ability to innovate, adapt, and scale.

The story of AI’s evolution is a tale of two approaches: open innovation and corporate ownership.

Open innovation, epitomized by projects like TensorFlow and PyTorch, thrives on collective intelligence. It relies on global communities of developers, researchers, and enthusiasts who contribute to, critique, and refine the codebase.

Conversely, corporate ownership has given rise to proprietary AI models, with organizations like OpenAI, Google, and Microsoft leading the charge. These firms invest heavily in research and development, producing polished, powerful solutions like GPT and Copilot.

Oh, and by the way, AI is not free.

Open source has low upfront costs but hidden expenses in integration, customization, and the skilled developers needed to maintain it. Proprietary AI has high licensing fees but bundles in support, updates, and faster deployment, though it carries vendor lock-in risk over time. Let me put this from the right perspective:

MCP Server, Revit, and Claude: what’s there for BIM Design

Let’s deep dive into the new trend that is blowing up LinkedIn feed, YouTube, and Reddit, labeled as “the future of BIM”.

The Model Context Protocol (MCP) is an open standard and open-source framework introduced by Anthropic in November 2024 to standardize the way AI systems like large language models (LLMs) integrate and share data with external tools, systems, and data sources. MCP provides a standardized interface for reading files, executing functions, and handling contextual prompts. In short: think of the MCP as a USB-C port for AI.

This way, for instance, your Claude can communicate and interact directly with external tools, databases, and web APIs, let’ say your Revit (for Revit 2027, Autodesk introduced Revit MCP Server as a feature).

By linking them, Claude can actively read, query, and manipulate your Revit model using plain-English commands, with main objectives such as:

• • • Smart Modeling & Layouts
    • Model Auditing & Data Extraction
    • Automation & Parameter Management

There are tons of BIM Managers, BIM Coordinators and BIM Specialists out there that, instead of modeling, are trying to build automations that can do boring, repetitive tasks on their behalf.

I imagine them in their little rooms trying to build their supposed “one-of-a-kind” automation, without any knowledge of how to write code, just like pirate radio in the UK in the mid-1960s, when pop music stations such as Radio Caroline and Radio London started to broadcast on medium wave to the UK from offshore ships or disused sea forts, to meet the growing demand for pop and rock music, which was not catered for by BBC Radio services.

Supertoys Last All Summer Long

I’m borrowing this title from a 1969 short story written by Brian Aldiss (later used as the basis for the first act of the feature movie A.I. Artificial Intelligence directed by Steven Spielberg in 2001) just to put on the table one big question: for how long will it last?

The typical LinkedIn post from someone testing open-source LLM models:

“The shiny new model runs great locally! It generates [any number between 40 and 200] tokens per second!”

What they don’t say is:

• • • the shiny new model runs great on a PC that not even NASA could run, bought to flex on LinkedIn knowing that the ROI date is likely to be infinite
    • it generates [any number between 40 and 300] tokens per second, but they never say if those tokens have any value if you put it to work on a task (code, RAG, whatever)

Then you ask in the comments and the castle falls.

Models that run on €5,000+ machines (take a look at the latest nice “box” NVIDIA DGX Platform here), without producing results comparable to proprietary models.

It’s like the guy in middle school with the new bike with big wheels and a thousand gears that he only uses to come to school to show off.

But where exactly are you trying to get to with this heart-pounding hype? I’m waiting, trustingly, for someone to spend a decent amount of money and get a decent result, and so far I’m still deluded. And probably I’m jealous because I only spend €5,000+ on a PC in my dirty dreams.

Bonus: CrankGPT

Two ex-Google employees built a completely offline and off-grid voice assistant. You turn a crank, speak, and it responds.

Obviously, it’s a provocation, and a tough one at that: they want to demonstrate that artificial intelligence can be efficient, private, and accessible even on a small scale, bucking the trend of big tech companies toward increasingly gigantic and centralized models. “Small AI is beautiful AI” is their motto.

If you are curious enough, check this out: https://squeezlabs.github.io/handcrank/

—–

Stay data-hungry. Stay data-foolish.
Your Friendly Neighborhood Digital Consultant