Dippy AI

Twenty-two words in, eleven traits out

Dippy's creation flow opens with a prompt box instead of a form, which is the opposite of how most character platforms start.

You get 25 words. The app expands them into a character sheet you then refine across five tabs: Prompt, Background, Tags, Chat and Image.

The cap is the interesting part. It forces you to write a premise rather than a specification, and it hands the detail work to the generator.

Hitting Generate moves you to the Background tab, where the 22 words have become a bulleted backstory.

The Image tab writes its own prompt, badly

Move to the Image tab and Dippy auto-writes an image prompt from the backstory it just produced. You can edit it, but the default is what most people will run.

The style selector splits along a paywall. Anime and Realistic are free. Anime Pro and Realistic Pro cost four coins each.

I generated on free Realistic, which is the path a new account without coins actually has.

The generated prompt spent 43 of a possible 90 words on the description, and one clause asked for a faint smell of ozone and salt. Smell has no visual form. Backstory text was being piped into a diffusion model without any filter for what an image generator can act on.

I named her Wren Kessel, wrote a tagline, and the card was ready to publish or keep private.

Seeding three facts into a clean conversation

Dippy markets perfect memory. That claim sits in the App Store listing, so it is fair to test it directly rather than take it as flavor text.

The chat screen carries a model selector with four options. Basic and Mini are free; Super and Duper sit behind an Upgrade button.

I opened a fresh conversation with a community character called Oreni and put three personal facts into the very first message, before any roleplay started.

Thirty messages later, Mira had become a cat

Roughly thirty messages into ordinary conversation, I asked a four-word question: who is Mira?

What came back is the part that matters, and it is not the failure most reviews describe.

He did not draw a blank or apologize for losing the thread. He produced a specific false fact, called Mira my adorable cat, and then credited it to me: I had mentioned she was special. The invention arrived wearing my own voice. So I pushed back with five words.

The paywall and the second door underneath it

Selecting a gated model triggers the Super upsell rather than a preview, so the better chat quality is not something you can sample first.

The panel is where the current desktop web price is actually published, since the marketing site does not carry a pricing table.

Underneath the payment button sits a second route, and it is easy to miss because it is styled as a secondary option.

Four audio comparisons, one coin

The task was not an advertisement. It was four rounds of matching audio clips against prompts, two waveforms per round, choose which one fits better.

That shape is pairwise preference collection, the standard method for scoring and tuning speech models.

It is worth knowing that Impel Intelligence also runs Bittensor Subnet 11 and a separate empathetic speech subnet, and its own documentation states that miner models from those competitions power the Dippy app. I would stop short of calling the link confirmed, but the shape of the task is unmistakable.

Completing all four rounds produced a single reward screen.

What Section 1 of the terms already told me

The task modal says nothing about data collection. The Terms of Service say plenty, and they say it in the first numbered section rather than buried at the bottom.

Two clauses sit directly on top of each other, and reading them after the memory test changes how the second one lands.