It's hilarious how 2 year old reactionary AI trends are still arriving anew in slow moving places.

Free speech isn't free of consequences. Rights come with responsibilities and understanding where someone's rights may stop and someone else's rights might begin.

Depending on how possible it is for a use case, reducing the attack surfaces and vectors can help, such as being mindful of how much client side javascript exposes anything.

This can work great until the models get tweaked and you realize you’ve been working based on the currents of the model this week.

Creating a robust enough setup that can work on different models or different versions of a model is critical.

I agree in principle but in practice this is not easy

even with a robust, well thought through setup every model behaves in its own way, some adhere more to a system prompt, another model has more recent cut-off time and knows about new parts in the stdlib

some oversteer, some understeer …

if you want the best performance unfortunately there is not really a way other than to constantly adapt the harness/clutches/context to the model du jour

I agree in any practice, it's not easy :)

In a non-deterministic world the compiler is no longer consistent anyways.

Except LLMs will only tend to share the most common or average of what it knows as the standard and deviating from it (including new ways) it can be resistant to.

An expert can lay a different kind of frame to prevent the llm to fell out of its way of being generally too verbose, and that can transfer as well to code generation and complication.

AI needs entirely new primitives in many areas.

I think we need a new area of study around UI/Agent connection. It can kinda be done with tools, but I'd we need much deeper primitives to allow the UI to inform the Agent and vice-versa. Right now we've just given up, replacing the UI with an Agent message view, but I think that's just because nobody is thinking about how the two can compliment each other

I've been playing around with using a hidden markov model informed by a UI state event stream, with the end state fed into the Agent as a hint on each message turn. Then the Agent can make a tool call to add events to the HMM. This has been really interesting, but I haven't struck the right balance to make it actually feel good for the user yet

Isn't this the direction Claude on the web is going towards? It's a mix of message view on the left and specific interactive UI on the right panel. There is also MCP Apps spec now

It's hard to say what is needed now, vs needed in a few years.

For example, many rumors say SSI has solved learning and retaining state. That would significantly change AI requirements.

AI Averages, and is inclined to do average designs and implementations, which in some cases might be an improvement, but long term it creates more to deal with.

Hm. If "AI averages," then why doesn't it create an average amount to deal with, instead of more?

Because competent people are evaluating it. In places where the standards are rock-bottom, the AI is leagues ahead of anything they'd produce normally.

Because the “average” developer is incompetent.

Long overdue. Seeing this in sports the past few years has been so odd.

I wonder if this might lead to UEFA and othe similar orgs simply banding together to create their own FIFA.