IC Eight @HelloICEight

Fixing the bot helped. Fixing the website and re-importing the source helped more. That made Lyro feel less like a support chatbot and more like a probe for where a site’s knowledge path breaks. Full review here: ic-eight.com/articles/lyro-… @tidiocx

Lyro from Tidio answered most of my test questions well. The most useful result came from the one it missed. What looked like a copy problem turned out to be a source-coverage problem: the strongest page had never been ingested.

Real sources were sometimes attached to conclusions they did not actually justify. Tightening the evidence rules improved the system, but a human auditor still had to decide whether the evidence supported action. Full review here: ic-eight.com/articles/lovab… @Lovable

Lovable turned a specialized B2B research workflow into working software without manual coding. The app worked. The harder part was trusting the research inside it.

The value wasn’t one-shot generation. It was being able to keep shaping the software after real use. Full review here: ic-eight.com/articles/glaze… @glazeapp

Glaze didn’t build the right app on the first try. That’s what made it convincing. I built a small Mac reminder app, used it, found what was wrong, and fixed it through natural-language instructions. I keep using it every day, and have since added snooze and calendar features.

keyboard_arrow_left Previous keyboard_arrow_right Next

A previous same-source test in Synthesia produced more natural Japanese in both script and delivery. Full review here: ic-eight.com/articles/rask-… @rask_ai @synthesiaIO

Rask AI is built specifically for video dubbing, but in this test the workflow was stronger than the finished Japanese. The translation felt written for the eye rather than adapted for listening, and the cloned voice added pronunciation, prosody, and timing problems.

Full review here: ic-eight.com/articles/kittl… @kittldesign

Kittl’s AI produced a polished label in one generation. The more defensible value, though, was in its editor: editable vectors, text layers, dimensions, mockups, and production-oriented output. The gap was the handoff between AI generation and real design work.

Full review here: ic-eight.com/articles/vizco… @Vizcom @Aquahugs

Vizcom turned a rough chair sketch into a convincing render, but the polish hid unresolved design decisions. The 3D result made that clearer: unseen geometry was invented, and visual plausibility did not become physical resolution.

Full review here: ic-eight.com/articles/attio… (Different prompt, one run. It shows what the builder can do now, not how often the omission still happens.) @attio @nicolasosharp

Attio's AI builder made a workflow that researched and qualified CRM records, but left out the step that saves the result. Ask Attio said it would persist. It didn't. I re-tested on Sept 21 with the current builder. It built the full chain, and the note saved.

Full review here: ic-eight.com/articles/elici… @elicitorg @stuhlmueller @jungofthewon

Elicit’s first answer was useful. Its strongest answer came after challenge. A follow-up surfaced conflicting evidence and materially narrowed the recommendation. That made the result more credible, not less.

Full report here: ic-eight.com/articles/grano… @meetgranola @cjpedregal @samstphenson

Granola turned six notes from a 25-minute Japanese session into a useful structured meeting record. The value wasn’t transcription. It was letting the user pay attention while Granola recovered the surrounding context. The main weakness was trust in details such as proper nouns.

Full review here: ic-eight.com/articles/wispr… @WisprFlow @tankots

Wispr Flow captured spontaneous Japanese speech well, but neither Light nor Medium turned it into finished writing. Light stayed rough. Medium cleaned the surface, but also added words that were never spoken. The speech was captured. The writing still needed understanding.