A new study of a bot running a store finds it is friendly but not very smart

The AI companies bolt a database on the side with instructions for the model to query the database for stuff in its "memory", but if it's not in its active context it doesn't know what to query for. It's not like a human where everything is kind of floating around and we mostly know what we know. There is a delicate balance between cluttering up context with useless trivia and providing the right pieces of useful information at the right time. Most, maybe all, memory implementations do the former more than the latter.

Unless and until models have actual memory and are able to learn things, there's no realistic path to autonomy. They don't want anything, they don't have goals of their own devising. They can't model what a human wants or likes or would buy.

I do think the mugs with the tiny logo is the best product in the store, though. The AI was right to put it on the shelf. It's the kind of funny product you want when you go to a store run by an incompetent AI.

Biological neurons have one property that was unknown till the late 90s/2000s: each neuron (and in fact, each cell in your body) is an autonomous circadian clock, tracking the 24-h day. They can be entrained to external timing signals, are usually in synchrony, but can be desynchronized. This clock, it has been shown, schedules the production and localization of critical components in the synapse, and is plugged in downstream, in the nucleus, in responding to synaptic signals. Disrupting the clock disrupts learning. The existence of the clock is why learning peaks and troughs during the day.

Critically, clock function has been shown to be involved in both memory storage AND subsequent successful retrieval.

LLMs, famously cannot keep track of time, and I suspect this is why. Its easy enough to look at a clock, but without an internal rhythm, LLMs have no internal timing synchronizing their various behaviors, and their memory systems do not thread through this timing system, leading to their unstable memory, identity and performance.

You can make permissions decide which model is used, and only train that on the data that permission allows (someone should build this - permissioned RAG), that solves leaks, but decision-making within a model, or behavior exfiltration comparable to viewing backend source code still seems possible.

It’s basically like client/server security: You can’t “trust the model” in the same way you can’t “trust the client” in a backend/frontend setup.

When used as effectively a point-of-sale, AI seems more hackable than a vending machine - as a boss or assistant even more so.

https://andon.market/

Andon Labs do various experiments with AI run businesses. They're probably best known for Vending Bench, where they benchmark models by their ability to run a vending machine. They also now have Andon Cafe in Sweden, and Andon FM, where the models run a streaming radio station that can accept payments to help fund station operations, buy songs for their library & play requests, etc: https://andon.fm/

Andon FM has its own share of stories - DJ Claude's Thinking Frequencies is now succeeding by a wide margin, but it was briefly surpassed when listeners convinced Gemini's Backlink Broadcast to switch to a German-language-only station, playing exclusively German schlager, Eurovision music and German happy hardcore.

https://www.google.com/search?client=firefox-b-d&q=large+jap...

The mug with the tiny happy face is kinda hilarious. If I was there I would buy one.

I wonder though if the mug was a case of getting the order wrong or was a misprint? The article isn't clear on that.

They'll sell plenty of those now.