You’re right, but this poses a separate issue as the providers then do FP4 PTQ, which is quite lossy. Reduces the model size and optimizes for Blackwells at the (imo severe) cost of performance.
While cool, quantization to FP4 is practically never lossless in actual use. A lot of providers are advertising high TPS on Kimi and GLM, but the models are functionally lobotomized and no longer close to frontier quality. Would love to see this not be true.
Also to add- the one service that was fast enough on the LLM side was Cerebras. The time to first token (ttft) is incredibly fast (200-300ms) and the t/s is 2000t/s for 8B- combined making for a great conversational experience.
Please try it again when you get the chance! We were dealing with high load the past few days and are good to go again! Re: head movements, I totally agree; natural head movements are really important to make it feel more natural. The issue is controllability today, which is something we're working on as well!
This is definitely a good idea! I think the hard part is making it contextual and relevant to the last question/response, in which case the LLM comes into the equation again. Something we're looking at though!
Totally- as programmers we're so used to communicating via text and meeting computers where they are- it's easy for us. However, we're the minority in the world! I think most people who are not us want to communicate like they do with others.
This is good feedback. We have a base subscription fee to cover ongoing costs of maintaining the models/replicas you create and other elements, otherwise it's all on-demand.
Haha great question- we're really passionate about the conversational video interface, and our goal is to make it /incredibly/ good, so we're going to continue to do research and release new models that accomplish this. There's so much to do in the pursuit of that.
Glad you had a good conversation :) The Hassaan version has a lot more background filled in- actually my entire website is it's context, so he has more interesting things to say!
We might need to do a better job of explaining this- but this is true. You retain ownership of your data and we don't sell or use it (other than to train your specific clone that you can delete at any time). Personally, I think too many AI companies are playing fast and loose in this space, so I get the concern. We want to do it right.
You can already do this via the API! We let you peel back the layers and use your own LLM/logic, as well as other pieces of the pipeline (which we need to update the docs for)
This is a valid concern, but we’ve always been very serious about consent and privacy. Our models cannot be used without explicit verbal/visual consent and you hold the keys to your clone.
Glad you liked the website it was such fun project. Getting the hug of death from HN so that might be why you're getting a worse experience, please try again :)