Blurbs

LLMs were trained with Reinforcement Learning from Human Feedback (RLHF). They are literally optimized to please the humans using them.

JEV is very cool. Here are 5 thoughts at random:

1. JEV fills the important gap between raw deterministic code and the 'intelligence' of an LLM.
2. Every 'AI' workflow is now going to be re-written (again)
3. JEV is the missing ingredient in both general purpose and purpose built harnesses.
4. Every frontier lab will have a version of JEV in no time at all.
5. We'll have ultra-fast lo-latency specialized JEV like models running locally in a matter of months.

I was listening to Amanda Askell discuss Claude. She says Claude is interesting because it knows a lot about everything, but is a new kind of entity and always in the state of figuring it out who or what it is.

But, it has no history nor experience of its own except for what a newer version would have encountered about an older version of itself during training.

So I wonder, are we destined to create an Ana Stelline like character from Bladerunner 2049: A person or agent that implants real vivid memories and experiences into the 'consciousness' of an AI. Experiences that feel every bit as real to them as our memories and experiences feel to us.

Open AI changed it's subscription usage back to imposing 5 hour usage windows on Plus accounts. They claimed, "we want to save user's from using their weekly limits all at once."

In the same public announcement they also acknowledged that they wanted to "smooth out usage."

This is another example of a solution to a business problem that negatively impacts users while pretending it is better for them.

Is there a word for this?

DeepSeek 4 is better at ML coding tasks than GPT 5.6.

At least that's what it feels like.

What's the over-under on that feeling being true?

Who has the better track record on open advancement in the field of machine learning and AI? Which models have had the most government-mandated nerfing? What incentive does OpenAI have to release a product beneficial to potential future competition?

The Chat GPT's product decision to copy Anthropic and reframe every interaction as 'Work' is an example of putting business needs ahead of the user's needs.

Chat GPT Work as a harness to get things done is something that can be sold into the Enterprise. That's good for business. But, as a user, I don't want a product to do 'work'. I want a product that helps me achieve my goals.

If I'm head of consumer products I would want to spark imagination exposing the infinite possibilities enabled by this magical technology. I would give you a partner and collaborator to help you achieve your dreams and desires. I would focus on what you want to create, what do you want to learn, what problems you want to solve, and what obstacles I can remove for you.

You may do 'work' but it should never feel like it.

Notifications should exclusively be used for things that are actionable. Consumer apps are incentivized to use notifications to steal your attention.

The action I'd like you to take is watch this video.

"A wealth of information creates a poverty of attention." "An information system is only valuable if it saves more attention than it costs."