Everyone online is trying to live in a black and white world where AI is either
By Allie K. Miller · May 1, 2026 · Curated by George's Blog
Everyone online is trying to live in a black and white world where AI is either good or evil, smart or dumb, but the reality is that it has a jagged frontier. Using these systems help you better discover its shape.
Here are a few weak spots from the lens of a business professional, not a PhD biochemist:
1. SVGs. The ability to "illustrate" and have that thing be infinitely scalable. Hard to image scalability in creative projects or game design without it.
2. Gullibility. I had to create a specific prompt to remind Claude that what people say or type is not always true. I don't want Claude to be manipulative, but I need it to deeply understand the world of manipulation to better support my response (ex: email reply draft) to it.
3. Multi-layered communication. Imagine you have to turn down a client event (but it's your favorite client, and you could have technically moved your friend group trip to make it happen but you didn't want to). Claude, which I find to be the best writer of the bunch, just says the "hidden" message out loud (”I have a friend trip that I could move if it’s critical enough but won’t be able to meet!”) like an untrained intern at a cocktail party after a glass of wine. Delicate nuance is tough.
4. Humor. The Claude Mythos paper said it was better at humor, but the current productionized models don't seem to understand the unspoken construction of a joke. It always defaults to randomness. It assumes that non-sequitur = humor. If you said "I want to say something funny walked into an office" it would say "oh a flamingo with a tutu!"
5. Multi-agent design. I don’t think we’ve cracked the right UI for managing agents quite yet (ex: we’re talking to one main CoS and that agent is delegating down the reporting line).
6. Discoverability. This is a mess on all platforms. I like Claude code surfacing tips while it’s thinking and working for you. But unless you're scrolling X for hours a day, or learning from people who are doing that (like the Mastermind I'm running), you’ll never learn the really advanced features and tricks. It feels unsolvable without AI as the surfacer.
7. Model selection. I realize it is just a dropdown, but the vast majority of AI users don’t know the model differences and tradeoffs. They’ll say “I want the fastest answers possible” and then when a response is low quality, they assume AI cannot do that task. The auto-selectors don’t seem to fully solve this problem either.
8. Extremely long context. People incorrectly believe that multi-million token conversations only happen with engineers+AI. As more business professionals lean into building, we're going to need to find ways to manage 10M+ tokens without context degradation. Countless business use cases are held back because of this.
9. Collaboration. How is there no way to merge memory yet? Or no-code model fusion? Agent teams starts to get closer, but I want true merging and model weight shifts.