The market for people who want to read random 42% snippets of Harry Potter will collapse.

LLaMA might not have the mitigations, but OpenAI, Anthropic and Google filter the outputs to ensure no regurgitation. It's not a problem on the output unless we count abstractions as infringement.

It's been only a short while, but good by Slack MCP.

> creators, whether they’re indie artists or billion dollar studios, deserve compensation when their work is harvested as fuel for someone else’s generative model.

Ok, let's assume this idea. Who should pay?

- model developers? they are a cost center
- model hosts? they make cents a million tokens, serving AI is commoditized, there is little profit in it
- users who generate their own ideas? how can you tax Mary's enjoyment of AI abstract art, or Jonny's passion for anime girls?

As long as they don't replicate the original images they have no basis to forbid training. If they do, it means copyright does not protect specific expression anymore, but instead protects abstractions and styles. That would be a disaster for creativity.

The entry-level unemployment spike correlates almost perfectly with Trump taking office in January 2025, and companies typically freeze non-essential hiring when policy uncertainty peaks.

But it is the cheapest solution. It doesn't cost anything to raise grades, they don't even have to provide remedial education. /s

Just like bad news spread wider than good news, movies about "bad AI" would get more views than Google's videos. It's an unfortunate trait of human preferences.

I find this vibe-coding process pretty exhausting. It generates so fast it is hard to keep up, but you can't disengage from it or it runs into problems.

My experience with agentic flows (vibe coding) is like yours. What is really needed is (1) having the project and the plan documented (2) generating sufficient tests. They put guardrails around the agent. You also have to manage its context carefully, the less useless stuff in it the better. If regular coding is like walking, slow and steady, vibe coding is like surfing. You got to give up control, so compensate by setting up constraints.

I am planning a MCP server for spaced repetition where the LLM has a tool to generate cards, and a tool to retrieve the next cards that are up for review. The tool calls will be orchestrated by the LLM itself, and the MCP server also coded with AI. Clade Desktop is all you need.

An alternative would be - instead of adding that followup question, to realign to users. It can check if the user is properly understood, if they are talking past each other or not, and align to the user better. As a conversation gets longer the model should introspect its performance so far, and adapt.

You can directly know you are conscious, no need to explain or verify. But for others, it is not possible even in principle to have direct access to their subjective states. Wanting to do so is like wanting to see the gliders in Conways' game of life code. The description of the principle (explanation for consciousness) is unlike its recursive unfolding in time. Time is the core element in consciousness, because recursion is temporal. Turing and Chaitin show recursive processes are incompressible, their shortest description are themselves. Recursion incompressibility explains why 1st person is isolated.

Past tokens influence future tokens through the attention mechanism. Humans also have a serial action bottleneck, we can't say two different things at once, or walk left and right. The limited and serial output stream is a feature. It focuses the model.

There is another benefit that is even more important - the model weights use less memory, so it loads faster. And since you need to load all the model weights once for each token, the faster it loads the faster it writes.