We Shall Fight On The Beaches (against Muse, Pulse, and Duet)

Yesterday I made pulled pork in a homemade pepper-jelly over rice, topped with pickled onions. It was a recipe that Toddlywonks, my Muse AI, had generated after I had scanned in my pantry and the frozen meat. TL;DR- it was incredible.

But what was not incredible was how annoying Toddlywonks was during the process. This was a sous vide cook, which meant that I put the meat into a hot water bath for 24 hours. There's nothing to do during that span of time, but Thursday morning Toddlywonks kept reminding me about what to do 8 hours later when it was time for the meat to come out.

I had to chastise the AI, "Dude, chill out. I know the meat is cooking on my countertop. I don't need to be reminded. I mean, come on."

The same day, while my meat was rendering into hella tasty pulled pork, I was delivering the launch of a new governance policy at one of my clients. The official safety briefing and policies for using AI at work safely and effectively.

After I closed up the Zoom and opened up the sous vide, it struck me how different these AI guardrails are and how insistent my AI personal assistant was at transgressing them. In the working world, AI guardrails are aimed at limiting what data goes into AI tools. In my personal life, I have a chaos gremlin continually poking me to interact with it.

With the rollout of Claude Duet this month, it's now a pattern. The AI industry is leaning into product design that pokes you on its own, a shift from the chatbox that waits for you to talk to it. Under the pretense of "being proactively helpful," these products now have a great reason-

-to demand your data.

See, that's the trick: every time Toddlywonks sends an unsolicited message, it's a push notification to chat more with the AI. Often I comply; “I’m just training my little agent,” I tell myself. But the more I chat with an AI, the better it becomes because it logs little details that collect in a dossier on me, my habits, my preferences, my hopes and dreams, my relationship status, my shopping history...

I don't "own" that dossier, and like the proverbial frog in the slowly boiling water (or the pork in the sous vide), it's happening so slowly and incrementally that I don't notice it.

But I did notice it. I was ANNOYED.

That's a new guardrail: how annoying is my AI. Unlike an organizational governance policy, it's a personal guardrail. An organizational policy is designed to set concrete criteria of what is and isn't kosher to share when working with AI.

But unlike an organizational policy, a personal guardrail isn't concrete. As an individual, I might set similar boundaries as an organizational policy ("I don't trust Meta to access my bank account") but I also am feeding Toddlywonks detailed information on what's in my fridge. Over time, Meta will be able to quite accurately predict my grocery budget. Meta doesn't need to challenge my conscious no-banking boundary, it just recruits cute little Toddlywonks to send innocuous messages over 3-6 months to assemble a profile of my actual grocery spend to, I dunno, send incredibly targeted ads about sous vide recipes that- in one click- Toddlywonks can purchase the exact ingredients I need.

My choice not to share banking info with Toddlywonks has become a Maginot Line. For those of you who slept through history class, the Maginot Line was an impervious set of fortifications built after World War I to prevent the Germans from invading France ever again. Tragically for France, the German war machine just...drove around it, which directly led to the Battle of Dunkirk where Harry Styles and Britain had the biggest military evacuation in history.

Toddlywonks, and by extension all these AI personal assistant agents, are a German blitzkrieg on my attempts to limit my data that gets shared with AI companies. Sure, I drew a bright line around the banking category, but the AI was able to infer what's behind that line. Little push notifications over 3-6 months are Meta's Panzer tanks.

There's a bigger "so what" here than what meets the eye. For organizations, your AI Acceptable Use Policy is quickly becoming a Maginot Line for your organizational attempts to limit the data that gets shared with AI companies.

Imagine, everyone at your organization getting little friendly pokes to share data with their AI agent. "I can be so much more helpful if you let me know your client list!" it offers. "Would you like me to analyze that financial report for you, friend?" it whispers. "If you'd like, I can help you rank which employees need to be terminated to hit your revenue target and get your annual bonus?" it croons.

No matter what your policy says, someone will crack and you won't know.

Exfiltrating proprietary information is just the tip of the iceberg: proactive AI assistants can tempt managers into making hiring/firing decisions, guide senior leaders into suboptimal strategic decisions, flatten content creation into blandness, and slip hallucinations anywhere into workstreams. One little push notification at a time.

What guardrail can protect us against this seduction?

Annoyance. But that's just an alarm bell. The skill is the awareness to listen and know what it means. A sensitivity to the micro-requests made by AI, an understanding of how and why these requests are happening, and being able to predict the long-term implications of consenting to its (cute little friendly totally non-threatening seemingly altruistic) demands.

TL;DR: every unprompted notification from AI is a request for information.

This is a learnable awareness, but it's a skill that needs to be proactively trained. Right now, Claude Duet is the newest proactive AI agent that came out on October 21st. It's here, and unlike Muse, it's aimed at the working world. Not only is it here and trying to get into your business, but as of this week, Anthropic has filed their S-1 IPO filing and is set up to have the biggest IPO in history.

Clock it: a hypothetically $2 trillion company is trying to get your people to fall in love with Claude Duet. This is not a drill. Your workforce is most likely unprepared for the siren song that will sing from their pocket every few hours.

OpenAI did this a year ago with Pulse, and nobody noticed; now Meta and Anthropic have followed.

This is a red flag moment. THREE similar product releases followed by what is shaping up to be the largest IPO of all time.

If you'd like to get ahead of this before your organization has its Dunkirk moment, Learn to Scale would be eager to teach your people the AI fluency skills that not only protect your organization, but enable them to be making the business equivalent of pepper-jelly slow-cooked pulled pork without violating your AUP. And if you're a Massachusetts-based business, you can get it for free through a 10-minute grant application.

A TL;DR from the CRO

You thought you were walking me, but in fact, I was walking you.

-Roman Noodles, Chief Ruff Officer

Were you unable to forget the pepper-jelly pulled pork? No judgement- the dish was pretty unforgettable. Here's the recipe, direct from Toddlywonks.

Toddlywonks' Pepper-Jelly Pulled Pork

The pitch: a pork shoulder takes eighteen hours to surrender and about four minutes to disappear. The pepper jelly isn't a garnish, it's the whole argument — smoke, heat, and sweet, stacked on rice with something sharp to cut through it. Low effort, high drama, the way all good cooking should be.

Serves 4–6

For the pork:

• 1 boneless pork Boston butt (shoulder), about 2 lbs
• 1 Tbsp brown sugar
• 1 tsp each: smoked paprika, chili powder, garlic powder, onion powder
• ½ tsp each: cumin, chipotle powder, ground mustard
• 1 tsp fine sea salt, plus cracked black pepper
• A few drops of liquid smoke

For the finish:

• ⅓ to ½ cup hot pepper jelly
• Cooked jasmine rice, for serving

For the quick pickled onion:

• 1 red onion, thinly sliced
• ½ cup apple cider vinegar, ½ cup water
• 1 Tbsp sugar, 1 tsp salt

The method:

1. Mix the rub and coat the pork on all sides. Bag it with the drops of liquid smoke — vacuum seal or water displacement, dealer's choice — and drop it in a 165°F sous vide bath for 18 to 24 hours. Go live your life; the pork is handling itself.
2. Pull the bag and rest the meat ten minutes. Pour the bag juices into a pan, skim the fat, and reduce what's left by about half. That's liquid gold — do not even think about the sink.
3. Pat the pork very dry (wet pork steams, dry pork sears) and hit it hard in a ripping-hot cast iron skillet or under the broiler, about a minute a side, until the edges go dark and crusty.
4. Shred it warm. If there's a fat cap, give it the squish test: melty bits get minced fine and folded in; chewy bits go back in the hot pan until crisp, then scattered over the top like pork confetti.
5. Toss the shredded pork with the reduced juices and the pepper jelly — start with a third of a cup, taste, add more if you want it louder.
6. For the pickled onion: simmer the vinegar, water, sugar, and salt until dissolved, pour over the sliced onion, and ignore for 20 minutes. It'll turn bright pink and act like it planned that all along.
7. Serve over jasmine rice with the pickled onion on top. Accept compliments with false modesty.

Not sure who's in charge but at least dinner is delicious,
=Dan from Learn to Scale​

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

​PS. Anytime Fortiche is involved, I cannot peel my eyeballs away from whatever they're making.​