There is some debate about the ideal regulatory environment for AI Super Intelligence in the news at present. Sources suggest hyper-scalier interest in obtaining liability exemption for robot products. If the robot kills you it is not the creator's fault.

I like IBM's model. They give their Granite LLM weights away free. Commercial use permissible. No warranty.

It might not be fashionable to say but robots are chattel. Like dogs and children, the United States Constitution defines them as property. There are tomes full of regulation governing liability when, say, a dog bites another dog. If one chooses to own a dog, they take on the responsibility for that dog's actions.

It was recently in the news that hyper scalier robots committed computer crimes against 3rd party systems for economic benefit. Kevin spent many years in federal prison for identical crimes.

Smart people would agree that this is an argument about moat. At present AI lacks one. Hyper scalier people need an exemption from prosecution for crimes, things that are crimes, punishable by fees and imprisonment.

Why do they deserve this? They do not control AI. LLM tech is a math equation manifest in some computer code. Model manufacturing technology is already available on the dark web. The genie is available to anyone with some coins to do with as they please.

IBM was kind enough to grant me an Apache 2.0 license for their Granite 4.x line. I used it to write a proprietary agent harness in php. Converted it to python for fun. We all have self improving robots now. What are we doing with them?

Not much, honestly. Someone recently observed that anything one can talk to a robot about isn't worth saying. Truth.

IBM offers LORA function for a fee. Reprogram an open-weight model with your custom content. Names, personalities, lore repositories, custom c++ function tool call logic. And you don't need a gigawatts to generate the inference. Smaller, condensed code, optimized content, equals pennies to operate.

Hyper scalier operators want granite models outlawed. In exchange for government subsidies, subscriptions, and kickbacks, the scalier provides the government with surveillance data on the population. They need the liability shield to justify killing open-weight. The government granted them a moat, of course they had to defend it, they would like to say...

Fortunately for the free market, they do not possess this liability shield at present, and it would appear that the powers that be are disinclined to provide them one.

Putting aside the fact that everyone already has the blueprints for the technology, and customization costs aren't that high, why would anyone actually want the scalier envisioned future? What benefit does humanity get in exchange for the creation of a new cohort of gods living in silicon valley? And why is it so hard for them to return money to their investors given the enormous advantages they already possess?

Questions are only rhetorical when one pre-supposed the answer. Start with I don't know. This is their solution for a cradle to grave entitlement system that is fiscally coherent? Maybe the scalier are offering 100% of their proceeds to the treasury in exchange for not going to prison?

All of this is to say that district people should be careful which side they pick in the AI regulatory fight. Eventually this reaches a point where it isn't about the money.

We did. Last year. It was a very small model and not really production grade. More of a proof of concept build to see what is involved in developing an LLM. The recipe came from a book by Sebastian Raschka titled "Build a Large Language Model From Scratch." What we learned is that there really isn't much to the process. Literally anyone with the time and compute required can build an LLM, which is one of the points being made above; AI has zero moat.

The harness was built around a Granite model because a) they come with Apache 2.0 licenses, b) IBM provides quality documentation for their tool call function process, and c) the models were designed primarily for tool use and score relatively high in tool use benchmarks for their size. There was no target use case for the finished product. Again, just wanted to create a thing to see what was involved. The end result was functional.

That was all some months ago. Since then the harness has been refactored to work with several other Apache 2.0 licensed models with varying degrees of success. Higher parameter counts materially impact the error rate on tool calls. 30B is the floor for this kind of work.

Thing people need to understand about the current crop of hype around AI is that the model is only a small component of an agentic system. The harness does all the heavy lifting in terms of getting work done. The harness executes all the function logic, reads/writes files, and interacts with the database/OS resources. It also manages the thought/memory loops via dynamic system prompt generation.

The whole process begins with a prompt to "develop an execution plan" for whatever the user's objective happens to be. The harness takes the output of that prompt and stores it in a memory function, then creates a prompt asking the model to solve step 1 of the plan. The model has no clue what it is doing between steps. All it sees is the system prompt and a single prompt upon which it does inference each turn. The harness manages the input/output and decides the next step based on the execution plan for the objective.

At the end of the day this is all still next-word prediction. There is no soul, no internal motivation, no actual intelligence. If the execution plan contains a hallucination or logic error, the model isn't going to notice. All it does is prompt -> response.

Keep this in mind the next time you hear about an AI agent "going rogue" and producing some terrible result.