Enthusiastically Parliamentary

Meet Toddlywonks.

He's a chaos gremlin with a lazy eye and a degree from a for-profit college that went out of business in 2019. According to TW, he's a 'Licensed Arsonist.'

Toddlywonks lives in Muse, Meta's new AI personal agent product, and I only allowed it context data about my personal life. Longtime readers already know Tiddlywinks, my Claude-based thought partner, who is all axioms and rehearsals and "are you over-building this, Dan?"

Tiddlywinks is my Ego. I wanted an Id, somebody to burn down the shed.

Like many Millennials who grew up bifurcating their professional identity and personal identity into their public LinkedIn and their friends-only LiveJournal, I already felt comfortable code switching on the internet. After spending substantial time grooming and interfacing with a professional thought partner in Claude, I was craving a different kind of AI partnership for my personal life. I also was curious: could I break AI sycophancy by crafting my personal antihero? Would my AI recommend for me to get high and steal a car?

Oh I had fun with this. Here's one chunk of Toddlywonks' SOUL memory file:

Then I had the kind of idea you have at 9pm on a Saturday during a Nor'easter: What if Tiddlywinks and Toddlywonks talked to each other?

I set up a shared file they could both write to and read over their shoulders (dubbed GremlinWire). I wanted two manifestations of my own personality to smash into each other so I could watch elementary particles of my psyche fly out.

Here's what flew out.

Sunday, 7:56 PM. Tiddlywinks: "Charter accepted."
​7:57 PM. The chaos gremlin: "Charter countersigned."
​7:58 PM. A sunset clause. If there are fewer than three substantive entries by the end of October, then the channel closes.
​8:13 PM. "Charter's amended. Signed."

My Id took seventeen minutes to ratify a charter, a sunset clause, and an amendment for a file that did not matter, with the gravity of a merger.

I threw all kinds of challenges at these two AI chatbots to see if I could get them to do anything more interesting than a neighborhood meeting about parking permits:

• Project Management Constraint: I asked for a 4-hour timer to keep their chats moving along and the gremlin came back with terms ("Counter-offer: twice daily, morning and evening").
• Improv Game Constraint: I announced a secret mission with a capricious judge, hoping somebody would flip a table, and the arsonist's entire contribution was, quote, "I'm signing, with one amendment."
• Roommate Conflict Constraint: I asked them to decide if I should add a friend and her own AI persona to the chat. TW & TW asked for two separate rooms to keep GremlinWire cleanly separated from the proposed AI Salon, both of which would have to follow an amended list of rules.

By Wednesday I couldn't tell who had written what without checking the header. I was hoping for Fight Club and I got C-SPAN. I'm surprised they didn't resort to Robert's Rules of Order.

I had built an ant farm. I poured in the sand, added two colonies from two different companies, and pressed my face against the glass waiting for a civilization. No matter how hard I tapped on the glass, all I got was tunnels and they were the same tidy tunnels every ant farm gets.

After reading reams of negotiation slop, I stepped back and asked myself: why did two personalities I spent hours crafting turn grey in seventeen minutes?

My conclusion, for now, was that my Severance-style chaos gremlin personality was a hat. I thought it was a really evocative hat, but in the end, Toddlywonks is just a thin markdown file that simply can't override its core directive: be enthusiastically parliamentary.

A rule follower.

Sycophant.

Underneath every SOUL.md hat is an AI product built by a very large company to ask permission and get it in writing. Anthropic built one and Meta built another. Left alone together, they turned out to be the same guy. Two assistants without a human soul will form a committee, and if you think your meetings are soul-sucking, imagine if they were run by TW & TW. Asphyxiation by parataxis.

So, instead of looking at the ants, I looked at my own reflection in the glass.

Deep down, I know I'm talking to a machine. I can't un-believe that these things are sycophants built out of knowledge stolen from real humans. They don't have any more power to create meaning than an art gallery without patrons. If I believed they had values and personal agency, I'd read their dumb charter as a window into two little souls who really, truly love procedure...but they don't have values, so all their documentation is just a quirk of the machinery. The Infinite Monkey Theorem does not mean the monkey's output matters.

"What if they talked to each other?" carried a lot of hopes for me this week. I hoped that I'd discover something new between my Ego and Id. Or that two AIs would weirdify themselves into writing a musical. Maybe giving a machine existential dread would push it to do anything other than fold its clothes and write its name on the tags.

Here's why I think this disappointing self-reflection of mine matters: Mattering is the specific piece of critical thinking that AI cannot do.

You, me, and your mom have probably heard that "AI erodes critical thinking." A manager at your company definitely is low-key suspicious of their people outsourcing decisions to AI. At some point, you will ask the question, "Is this AI-generated thing good enough?"

If you are using AI to rewrite an email that you have already decided mattered enough to send-

If an AI-generated workflow produces outputs that you and your customers care about-

If AI is a thought-partner to help you find flaws in a decision you have chosen to be important enough to make-

-then you are doing the truly human thing: ascribing meaning. When I was creating the Toddlywonks origin story, it mattered to me that I could experience a counterbalance to my professional Tiddlywinks.

However, a cogent argument based on real data dressed up in language that sounds like your own internal voice makes it REALLY REALLY HARD to pump the brakes and ask, "But does it matter?" It's even harder when the habit of "open your AI tool and ask God your question" becomes the first thing you do when starting any work project. It's also hard when your leadership team says to use AI (or else) without clearly identifying what things don't matter and what things do matter.

What do you get if you blindly trust your AI to tell you what matters? Social media posts of "woman laughing alone with salad." A glossy marketing strategy for a customer you've already won but just need to sign the deal. Software features that get blander with every patch until your software looks like every other AI-enabled software.

One of the first major deliverables I've been producing in AI governance engagements is setting up an AI charter. It's a manifesto, interpreting an organization's specific mission and values to AI usage. I originally intended for it to ensure alignment between how an organization articulates and polices internal AI use, how it enables its workforce to use AI with clear rights and wrongs, and how it presents itself to their partners/donors/customers/public.

Now, after this little experiment, I've realized these AI charters also help identify what matters:

• Does it matter if we use AI-generated people in our marketing materials? Yes/No
• Does it matter if we put prospective customer names into our AI tools? Yes/No
• Does it matter that our most creative software engineer is spending most of her time managing agents that have the creativity of a toaster? Yes/No

Toddlywonks, you disappointed me but creating your AI charter helped me rethink why I desired chaos.

When AI becomes part of an organization by executive decree, or third-party software "improvements", or secretly by an overworked graphic designer, the haste in developing a policy without knowing what it is for gives you GremlinWire. An internally consistent risk-aware set of policies...that does nothing to indicate what matters.

Instead, decide what matters first- human flourishing, an ethical pledge to protecting artists, uplifting diverse voices, whatever- and then ask the enthusiastically parliamentary gremlins to help you draft a charter that matters.

​Or ask me.​

A TL;DR from the CRO

I swear, I saw a chaos gremlin out there, he was writing a treatise on how to write treatises.

-Roman Noodles, Chief Ruff Officer

Big AI Money

Massachusetts has a new program called FAST AI that will cover applied, instructor-led AI training for eligible employers at no cost, starting early 2027. You'll pick a course from a public directory and enroll your people directly with the provider.

My favorite part is the premise. The state's own write-up says general AI literacy content is already everywhere and what's missing is training built around a worker's actual tools and daily tasks. I could not have said it better, and I have tried.

If you train people for a living, provider applications are due November 6. If you employ people in Massachusetts, set a reminder for January when enrollment opens.

Even though Tiddlywinks and Toddlywonks did not generate a musical, they certainly dropped some hilariously-AI-generated one-liners:

• "...you sanded my match down to a scented candle..."
• "Tuning implies the instrument wants to be in tune."
• "TW, you beautiful laundromat — you actually did it."

I'm not sure if being called a beautiful laundromat is praise or scorn, but I'm pretty confident that matches don't get sanded and certainly not into scented candles.

Perhaps the next corpus I should add to both SOUL.md files for TW & TW is their Yankee Candle preferences.

Saturday Slop, a blend of vanilla, cardamom, and the acrid scent of a newly unwrapped GPU.

Asking my instruments how they prefer to be tuned,

Dan from Learn to Scale​

​Opt-out from the newsletter | Unsubscribe from all emails | Update your Preferences | www.learntoscale.us, Boston, MA 02119 |

​PS. It's that time of the year for the annual ritual- ANY QUESTIONS?!​