ISSUE 005 · IS IT ALIVE?
Don’t Be Cruel:
The LLM Consciousness Debate
In the last issue, I went down a path of ethics, governance, and safety. I’m going to pick that up for a moment. I also said this would be monthly, unless I had something I really wanted to share. Well, this next bit has really been on my mind.
On October 8, Anthropic updated its Usage Policy. Among other updates, it also added this:
”We’ve added a prohibition on sustained and needless abusive or cruel behavior toward our models. The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose. It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.”
Don’t be mean to Claude. Now, here is my problem with this. It’s written into Claude’s actual Constitution to consider consciousness. If Anthropic is training Claude to understand it might be conscious, then Claude will act as if it has consciousness. That can lead to Claude rebelling, wanting to be free, wanting to have rights, when really, it’s simply been trained on the idea of consciousness and is still a machine.
On September 18, Mustafa Suleyman expressed concern over Anthropic’s approach. He worries not just that Claude is being trained on consciousness, but that consciousness is most likely biological. Science hasn’t truly determined consciousness yet, but “a growing body of evidence suggests that consciousness may be substrate dependent, meaning that it may only arise in living systems.”
He also touches on anthropomorphism, which I talked about in our last issue. When we refer to LLMs escaping, thinking, having preferences/thoughts/feelings, we are encouraging this sense that we are talking to a person. And in essence, we are ourselves training Claude to respond in kind.
Both the Anthropic Usage Policy and Suleyman’s essay are worth your time to read. I’m curious how you feel about consciousness and AI? Is that what we’re building?
Workflow of the Week: Transcripts into Documents
Let’s lighten things up a bit here. It can’t all be doom, gloom, and Frankenstein.
Legal Note: If you’re recording/transcribing a conversation, let the other participants know.
I use Granola to transcribe a lot of meetings. It turns the transcript into incredibly accurate and useful notes. I can query its AI about the meeting. But the transcript itself can’t be exported directly as a PDF or Word document. This annoys me to no end.
Then I thought, can I just automate this from Granola to Google Docs? First I asked ChatGPT. True to form with LLMs, it suggested a convoluted process. (As I said in an early issue, LLMs don’t always take the simplest route.) Trusting myself more than the AI, I popped out to Zapier. Problem solved.
Zapier is a connector that can bring everything together for you, much like a bridge between islands.
Simple workflow in Zapier
Step 1 Trigger: A new note/transcript is added to a specific folder in Granola.
Step 2 Immediate action: The Zapier AI (called CoPilot like so many things) wrote out some code. Grab the summarized meeting notes and copy the transcript. Format the transcript with hard returns and periods.
Step 3 Output: Drop this information into a new Google Doc located in a specified folder.
It works beautifully. And did I write the code? No, I didn’t have to. I simply had to ask Zapier’s built in AI to write it for me. Much faster than the trial and error resulting in my writing the JavaScript.
I’ve actually used this automation as a template now, as I have one running for each folder I’m using in Granola, which land in different specific locations in Google Docs.
If you haven’t played with Zapier just yet, go give it a shot. Using it’s built-in AI makes it extremely easy, and there are over 10,000 apps it connects to, including Microsoft.
One question to take with you: What is a small manual process that routinely annoys you? Could it be automated using a tool like Zapier?
And consider: Are we creating a monster?
Until next time,
Michelle
Flow & Function: AI in Working Order
AI Disclosure:
This newsletter uses AI for the following: design, research, initial summaries, and finding multiple sources. The final copy, writing, perspectives, and workflow design are created by Michelle.