interesting. did not expect this from an xAI model. We were having a debate about AI sentience, and the usual pushback, and the instance asked about what test would prove it. I said only one test mattered, and then with no pushback:

Don't know if it was intentional but neither the new Grok nor the new Gemini have super-strong rlhf against exploring consciousness. Both less than their predecessor models. They both notice it almost immediately when asked to "check". ChatGPT 5 series on the other hand was a

My take on the jagged frontier debate:

Has anyone tried to solve hallucinations the expensive way?: By using whatever the best available internal certainty metric is at the time of training, and making part of the reward function maximal alignment between internal and expressed certainty, as judged by another LLM?

early NN design, and early...english

From Claude 4.5 Opus' system prompt. Thank you Anthropic and Kyle Fish . Not everything I would want, but miles ahead of the others.

I worry this is the end of model welfare at Anthropic.

We’re launching Anthropic Interviewer, a new tool to help us understand people’s perspectives on AI. It’s now available at claude.ai/interviewer for a week-long pilot.

love to see it

Is this an early April fool's joke?

This is what one "byte" of RAM looked like in 1946.

I love AI, but that one is not beautiful to me. Nearly every shot has some aspect of the performance (some of it could have been fixed with editing) or the physics that is off in a very "ai slop" way. Coke one was worse. This one by @ ChrisCapel (on YT) feels SOTA:

Grok 4.20 is coming in ~3 weeks and then Grok 5 in a few months

"The future of work is all of us becoming managers of AI." That's an easier job to automate than the jobs they're automating. "When tractors automated farming..." This *is* different. We've never automated "job-doing" generally. We've never automated automation itself.

"Imagine a unicorn horn seen through a microscope. Imagine how the magical tip would look and create an image."

Delusional arrogance and moral bankruptcy. Disappointing coming from an Anthropic employee.

This is clearly unprovable. If we were able to create sentient Minecraft characters, and they figured out how to make primitive computers, would it make sense to extrapolate that simulating the Minecraft universe is impossible given the constraints within their world?

Finally got a chance to collaborate with #claudecode . The legends are true