@JRaugel Yea that made sense. And my wording was wrong, meant more like updating weights for first layers based on pure error signal compared to noisier (?) signal.
We usually update first layer with noisier/indirect error signals as opposed to this.
@zephyr_z9 MLCC's are expensive at that voltage tho. They're pretty good for ESR and ESL, but expensive for handling 800V.
Electrolytic caps do be good for handling the transients/ripples at higher voltages, but high frequency switching results in more loss from these.
Pocket (@heypocket) is your notetaker for real world meetings.
In the last 5 months, the team has delivered over 30k units with a $27M annualized run rate, growing 50% month over month.
Congrats on the launch, @AkshayNarisetti and @gabrieldymowski!
https://t.co/BGJPC0lJtD
Releasing kyutai pocket TTS inference in webgpu today — it’s open-source, here’s a demo from my phone
https://t.co/PXMRHYM54a
It took a while and a bunch of code to implement streaming inference from scratch!
@MillerLabMIT https://t.co/TqUdjDqAxW
Overall arch has a decent amount of components. Would the source code be released?
Looks pretty interesting. Curious about setups on multimodal with audio as well
@leothecurious This is the way. How would the bot know, this is the successful completion. One way would be the base model understanding linguistics (your game world). Then using that to setup small actions.
Reward is the weird part, maybe train it like a dog?
Good boy(audio)-> reward+1