VanL

Building Model Monster: AI governance grounded in system architecture, not questionnaires. AI/open source lawyer. DMs open.

[VanL @VanL]

I think "It is always more complicated than it seems" is a good rule of thumb for life.

[VanL @VanL]

I packaged up my agentic coding assistance system for anyone to use: https://t.co/yDdQfrKwKk.

What this is: I have been doing a lot of AI-assisted development recently, and I have found that I frequently get better results than some others. I think that part of it is the structured system I use for managing agent context and guiding agents away from common pitfalls.

Agent-guidance makes agents use a more highly-structured specification-plan-implement cycle than the basic planning mode will get you. In combination with https://t.co/UhblM4M3fb, plans and implementations get a competitive review from a different agent, helping eliminate blind spots.

If you are a user of @garrytan's gstack, @every's compound engineering, and @obrajesse's superpowers, agent-guidance will work with those skill stacks to make implementation easier and higher-quality.

[VanL @VanL]

@adamhjk The "MCP is dead" chatter is because most MCP servers are poorly designed. You can design one that is isomorphic to a cli interface and doesn't kill context. But the big advantage of MCP is that it can be server side. Nothing needs to be validated to run at the client site.

[VanL @VanL]

@dhh @codejake FYI, I am a lawyer specializing in this stuff, and we consider "open source" to be a term of art specifically referring to the OSI's definition. Not arguing with you, just point out that 99% of lawyers using the term would conclude this is a "source available" license, not OSS.

[VanL @VanL]

@AnthropicAI @ClaudeCode I know you want me to use Sonnet, but it just isn't good enough for a lot of the things I am doing. But you are doing a good job convincing me to drop the Max plan.

[VanL @VanL]

@AnthropicAI @ClaudeCode The daily and weekly limits on Opus usage are terrible, especially for a 20x max plan. Ever since you released Sonnet 4.5, I get warnings about approaching daily/weekly limits within 1 hr of use, and that is not even with running extensive subagents.

[VanL @VanL]

@adamhjk @gwenshap Please tell me more. I've always seen LinkedIn as resume+ a sea of faux-profound self congratulation.

[VanL @VanL]

@sammcallister @TheAhmadOsman But no response to downgrading "5x more" and "20x more" to 3-8x more, even when you are still advertising it that way?

[Enrico Shippole @EnricoShippole]

We open-sourced 99% of US caselaw on @huggingface. Both AI and legal tech companies are selling this data for a high premium. You can simply just build a wrapper around it and freely compete with them now. That is why we love open-source.

https://t.co/RwgXRKoFaz

[VanL @VanL]

@theshawwn I am surprised that AI-as-intelligent-filter is not a paid addon. Maybe advertising is just that much more valuable?

[VanL @VanL]

@theshawwn I am sure I am in my own filter bubble, but I see it too - but I see it everywhere. I suppose there are some mastodon servers where its not there, but its because they are not very trafficked. So long as what is being monetized is attention, this is where we're at.

[VanL @VanL]

@theshawwn I am sure that LLMs are great at it, but they learned from humans: "Ten sure ways to lose weight (you won't believe #8)". The problem is that once it started working, everybody copied it, so now there is 1 weird trick to getting engagement.

[VanL @VanL]

Simplebroker quick start:

# Create a queue and write a message
$ broker write myqueue "Hello, World!"

# Read the message (removes it)
$ broker read myqueue
Hello, World!

# Write from stdin
$ echo "another message" | broker write myqueue -

# Read all messages at once
$ broker read myqueue --all

# Peek without removing
$ broker peek myqueue

# List all queues
$ broker list
myqueue: 3

# Broadcast to all queues
$ broker broadcast "System maintenance at 5pm"

# Clean up when done
$ broker --cleanup

[VanL @VanL]

Released a fun little utility that I use for coordinating multiple agents (like claude code):

Simplebroker: A lightweight, no-configuration message broker/queue for cli tools

$ pipx install simplebroker
$ broker write tasks "ship it 🚀"
$ broker read tasks
ship it 🚀

https://t.co/YTde0JxXvR

[VanL @VanL]

There is a tension between the parts of Alsup and Chhabria's rulings that content owners like best (Alsup's "pirate libraries" and Chhabria's "downstream dilution"). Alsup said that building a library was the immediate use, and that was infringing, even if the purpose of the library and the later use was building an AI model. Alsup said you couldn't excuse the downstream fair use from the immediate infringing use - they were different uses, so one could be infringing and the other not.

Chhabria's opinion basically says the opposite: The whole thing is one big use, from downloading from pirate libraries through downstream generation of new content by models. It is from this perspective that he concludes that dilution is a relevant part of the 4th factor analysis.

These two opinions are inconsistent. If Chhabria is correct, then Alsup's "downloading from pirate libraries" finding is wrong.

But if Alsup is correct that downloading is one use, and training is another, (and presumably that generation would be a third), then Chhabria's downstream dilution argument is wrong.

I think Alsup has the better argument, and the one that is actually better for content owners. But I see a lot more people wanting Chhabria's arguments to be the ones that carry the day.

[VanL @VanL]

The issue is that the analysis by Copyhype/Judge Chhabria collapses two different uses into one: Training the model vs using the model to create. Those happen at different times, usually by different people, with different inputs and outputs. The market dilution theory says that use 2 should be included as part of the factor four analysis for use 1. However, that kind of "downstream" analysis isn't supported by any caselaw (with the possible distinguishable exception of something like Napster/Grokster, which nodded to downstream infringement in the same manner - but by further direct copies of the original work, not by attenuated non-infringing other works).

[VanL @VanL]

@brianlfrye Of course he was encouraged in this behavior by the Copyright Office.

[VanL @VanL]

@brianlfrye @McCoySmith Under almost all circumstances, weights are not copyrightable, even if there is expression from the training material "in there." The weights are not the result of any creative choice made by a human.

[VanL @VanL]

@copyrightlately But as you point out, the worst is the analysis of the fourth factor. None of (c) law is designed to protect the market for the artist, only the market for the work. Not only is "market dilution" made up from whole cloth, but so are their lost sales. It's not the same work!

[VanL @VanL]

Most substantively, the report makes the standard mistake of treating models and outputs as substitutes for each other, which they are clearly not. A model has completely different uses than any of its inputs or outputs. Goldsmith won because the Warhol fdn used Orange Prince on a magazine cover in direct competition to the original photo. But try putting a model on the cover of a magazine - it is impossible. The concept doesn't even make sense.

[VanL @VanL]

@copyrightlately The Office has deep expertise, but they chose their positions before they did their analysis, and it shows.

I am not saying any of this is easy. But this report is inconsistent with on-point precedent from SCOTUS as well as with aspects of previous parts of their AI report.