Let users flag toxic users. Use the resulting toxicity score to filter out messages based on each reader’s toxicity tolerance level. Both the toxicity score and the tolerance level are driven by user flagging, with score decay over time to account for improved behavior.
But I've read somewhere that KV cache for speech-to-speech model explodes in size with each turn which could make on-device full-duplex S2S unusable except for quick chats.
That's a contract between users and HN. Airtrain is a 3rd-party.
If HN API exposes personal information publicly through their API then there is a problem.
And AFAICT the only way for HN to prevent user comments from being used by 3rd-party is preventing access to those comments, meaning a) sign-up will have to be more stringent and b) visitors will have to sign-in just to read (or scrape) comments.
Data selection depends the use-case. Two contrasting use-cases I see are:
- Emulation
- Advisor
In case of MTG player emulation for example, I think it makes sense to group data by some rankable criteria like winrate to train rank-specific models that can mimic players of each rank.
Leaking original data would expose the company to direct copyright violation lawsuits. Changing T&S is simplest way to stave the legal risk exposure, buying time to implement technical remedies.
As ridiculous as it may seem, they're doing the right thing.
I think OpenAI's founding nature is about research so it will disappear when it either runs out of key problems to solve or funds, whichever comes first. I see its commercial efforts as driven primarily to maximize their research runway. Operating ChatGPT commercially also helps research into ML-related UX and operational related problems.
That said, I cannot rule out purely commercial ventures with tenacity necessary to compete spinning out of OpenAI.
PS: I've used it to remove jitter in virtual camera movement while cropping video around faces in real-time, streaming detected face locations to a Kalman filter worker and get back stream of stable camera locations.
It 'feels like' like seeing. There is sense of dimension and position in the space and objects. I can imagine people I know and places I've been to with amazing details but visual details like texture are limited to where I'm focusing. Rest of the view is filled with 'feels like they are there'. It's not retrieval because when I try to focus on non-memorable parts of a face, I can tell that details are made up on-demand using common variety.
And what I 'see' is affected by light over closed eyelid as well as inner blood vessel, minor debris and micro organism floating over the cornea, meaning input from the eyes does play a role even with eyes closed.
While I have very vivid imaginations, I don't think I have photographic memory because what I can recall is rather too creative.
[ my public key: https://keybase.io/superdon; my proof: https://keybase.io/superdon/sigs/x8N-xeM2DR6Qs9kHK_Sjoc0fGeYUZZpbtGwj1f4GzXg ]