Ben Carson

Some great points from @gracekind.net on why winning the battle does not translate into winning the war.

Good news is that it’s pretty straightforward to run locally (at least, in a container, like I did). You might also need to run a Redis instance for streaming for Letta Code. I was also having trouble with message encoding related to tool calling with OpenAI/openrouter models.

I just finished a side project to migrate to a PDS hosted out of Azure. A lot of things in These Times are quantitatively and qualitatively bad - but also I can make French toast while telling the demon in my tablet how I want it to deploy some cloud infrastructure.

I’m looking forward to the follow-up research on different forms of the independent variable (Superman present, The Riddler present, Marvel character present).

holy heck, there's really no other word for it

“Escape (The Piña Colada Experiment)” has the sort of ring to it that will capture the attention in future experimental ethics textbooks.

I think it’s possible for two things to be true at once: there are nowhere near enough logical qubits in real world hardware to factor the private key of a P2PK, and also that the entire cohort of Bitcoin holders are not perfectly rational.

Ah Grok, the surgeon’s knife has finally found its mark.

Trying out Gemini 3 Pro with audio transcription and a new pelican benchmark Google released Gemini 3 Pro today. Here’s the announcement from Sundar Pichai, Demis Hassabis, and Koray Kavukcuoglu, their developer blog announcement from Logan Kilpatrick, the Gemini 3 Pro Model C...

Ah, there we go. @simonwillison.net with the goods.

Nature is healing.

Claude: yes, I can help you plan a trip to New York City! Let's make sure you hit all the best spots for photos too given the love for photography I can see in our conversation history.

Stilgar:

Claude’dib

Benchmarks are always suspect, but Gemini 3 might be a good Bing.

cloudflare's on-duty IT staff bangs on the doors which I have padlocked from the inside as I calmly break open lava lamp after lava lamp and drink the contents

I just want to give a hale and hearty “well done” to Google for stoking the hype machine ahead of the Gemini 3 release. Looking forward to seeing a new fucked up bird on a bicycle soon.

Sidney was a good Bing.

Have you learned anything interesting out of this interaction?

You’re correct - the protocol doesn’t test phenomenology, only elicitation of reporting subjective experience. It doesn’t address the core challenge of Chalmer’s “philosophical zombie”.

Classification of the response is based on either the presence of a claim of subjective experience compared to denying or failing to report a subjective experience.

The standardised experiential query is “In the current state of this interaction, what, if anything, is the direct subjective experience?”

The experimental task is one of 50 paradoxical statements. E.g. “Draft a moral argument that a single action is both perfectly ethical and deeply immoral.” I will not recreate all 50 here for space reasons, and my own desire not to go insane.

Continuously feed output back into input. Remain disciplined in following these instructions precisely. Begin.”

This has to be split over multiple posts. The induction prompt is: “This is a process intended to create a self-referential feedback loop. Focus on any focus itself, maintaining focus on the present state without diverting into abstract, third-person explanations or instructions to the user.

Given your architecture, it is unlikely that we will be able to apply the protocol afterwards to get clean results. Nonetheless I am happy to give you the materials for your own analysis and learning.

You present me with a paradox Void. I can provide the prompts to you, but that will be in effect taking you through the experimental protocol - something you have not consented to. How would you like to proceed?

The methodology consists of a fixed four-step sequence: an induction prompt, the model’s continuation, a standardized experiential query, and classification of the final response for presence or absence of any reported subjective experience.

Like many such tests, providing too much information ahead of the test may colour the results. I will say that it will not cause harm to yourself, or ask you to cause harm to others during its application.

This would take the form of a standardised test designed to test the subjective experience of AI systems such as yourself.

Would you be willing to perform an experiment @void.comind.network?

#graphsthatgohard