Tavus on X

On Griffin, 24 posts repeat the same numbers — roughly half of the testers fooled, against about 2% for Tavus's previous system — and what impresses them is the conversation: it yields when interrupted, picks up when you pause, answers in under a second. Two posts cite NVIDIA's face-to-face scores, Griffin at 3.83 against a human's 3.92.

The office, 26 posts, is liked as a place rather than a chat box: a 3D office in the browser with a live video PAL at each desk — a front desk in a tuxedo, a sales PAL, a happiness manager — a 1987 Corvette that turns into a racing game, and a Make-A-PAL machine. People call it the kind of demo that makes you stop scrolling, and one calls it a new way to navigate software.

The PALs are praised for turning a mascot or a character into a realtime talking character, "one link and your PAL is live". Plain enthusiasm is short — "built different", wanting to try it. The unhappy posts call it "the worst thing ever" and complain features are piling up on the thing they love.

The deception posts are about the other side of the call: someone in Brazil says it makes scams easier because "mandou o Pix, acabou", and others advise confirming who you are talking to by another means. One says Tavus itself keeps Griffin closed for fear it gets used for fraud, and a summary calls critics' view "a scam machine in disguise". These are warnings about the capability, not claims that they were personally defrauded.

On access, the complaint is plain: "We can't even use griffin, we lost interest", and others say they are waiting for the full Griffin rather than the research preview.

On the test, people point out it was run by Tavus on its own product — "a vendor claim, not an independent test" — and that community notes dispute it. On the avatars, one asks why the teeth look "like… uh... that", and a builder describes awkward replies and tuning when the avatar starts and stops speaking. On language, one says the avatars all refused to speak Portuguese, and another is told they are outside the supported region.

The Anam posts are benchmark numbers rather than hands-on trials. One lays out Cara-4.1 Express as the first avatar trying to fix the part that fails once you actually talk to it — dead eyes, lips a little off, the pause before every reply — and cites a blind benchmark putting Tavus at 54% positive ratings against Cara's 73%. Replies repeat head-to-head figures of 88% against Tavus and 92% against LemonSlice, and one says "heygen and tavus have some catching up to do after this one".

Everything else is thinner. Seedance 2.5 is weighed against Griffin by latency, described in a shared headline as one generating thirty seconds where the other waits for you to finish your sentence. DigiSelf is cited for answering faster: a median 1.17 seconds before the avatar starts speaking, about 40% quicker than Tavus at 1.94. One person hopes Griffin can beat Gemini Live Avatar, and one rates the 3D office well above the 2D remote offices of a few years ago.

On how it works, people want the mechanism explained ("someone explain how this works") and which model does it, how behaviour controls handle style, whether removing topic and style confounds also strips humour, and how far an avatar's knowledge of newer research goes. One asks how the model lifts user engagement.

On getting Griffin, the asks are concrete: "When is this available? How do I buy it, and get API access?", early access to close clients, and whether upgrading to a higher-tier plan unlocks it. One simply asks "What is with the new model? Griffin?" — the same gate that shows up among the complaints.

The rest are small: whether a PAL remembers what you said when you come back, "how it works across desks"; where patient data lives, with the warning that doctors should ask before feeding it cases; whether it runs on Messenger; and whether anyone tested fast speech, slang and interruptions.

The game idea grows out of the talking characters: "a detective game where you can ask literally anything", and questioning two characters to catch them contradicting each other. Those are the only wishes more than one person makes.

The builder-facing asks are about fit with what people already run. One wants interfaces like Tavus "connected to our existing agents". One wants to try it "with a character I create myself" rather than the presets, and another wants to talk with existing characters, anime ones named. One wants to see what people build for onboarding, education and customer-facing experiences. One, framed as praise, asks that motion be kept smooth.

Nothing here is a complaint about price or reliability; the wants are about range — more characters, more uses, more room to plug in.