4 ms·
I am most excited by the conjunction of speech recognition, natural language understanding, language modeling all coming together to create solid conversational
by whimsicalism 5y ago
I am most excited by the conjunction of speech recognition, natural language understanding, language modeling all coming together to create solid conversational AI interfaces.
There's a lot of things I could get done a lot faster if I could converse with a slightly dumb ML algo the way I would talk to a person, even just for getting tasks done, etc.
- newyankee 5y agoThe only concern I always thought about was whether speech recognition would every be accurate enough using low processing power. The development beyond popular languages seem to have plateaued. I am sure speech recognition specifically is much better and looking at OpenAI demos blow my mind even now, but I have personally not observed significant difference in voice recognition on my Android smartphone - 2013 vs today.
- whimsicalism 5y agoNot true at all for me, noticeable improvements in speech recognition and it is easy to get low cpu and accurate speech recognition with a framework like wav2letter. 2013 phones hardly even had speech recognition, 2 years after Siri was released. I do think we have to guard against our cognitive biases that often stop us from noticing gradual but continuous improvements.
- datameta 5y agoWe can do keyword spotting with microcontrollers now. All kinds of techniques like quantization and tree pruning are squeezing down more capable NLP models further and further. There is dedicated analog hardware being developed for in-memory compute (ReRAM, photonic). TinyML/AIoT is shaping up to be one of the most society-defining technologies of this decade.
- cudgy 5y agoMy old Motorola Razr had less friction than Siri on the iPhone. Processing was local and fast … Siri gives me a beachball for a few seconds and I am unsure if Siri heard me or not so I repeat what I said again and then Siri shuts down. Hey Siri! Hey Siri!! Hey Siri #%&@it!!!!!
- xyzzy123 5y agoI want the version of Google Assistant / Alexa / Siri that actually works. I'm possibly being uncharitable - the speech recognition is fine but to me they feel like talking to a mannequin, "there's nobody home" so to speak. Being able to actually outsource big chunks of my executive function to a reliable assistant that can plan with me, gives timely prompts and never forgets would be life changing. State of the art UX for this is still calendar and notes apps, want it to be Jarvis.
- CodeGlitch 5y agoYeah they don't remember the previous response, so you can't ask for the weather tomorrow, and then "what about the weekend?" Without specifying you mean the weather.
- nefitty 5y agoWhat sorts of things do you have in mind? I've used the app Magic, which is sort of like an on-demand personal assistant. They're real people, but I'm imagining the sorts of requests I made being done by an AI. "Get me a list of cool concerts around Valentine's Day, near my city, along with links to music by the bands involved. Buy me the tickets for the one I pick and then update my calendar and send my wife an invite. Get me an Airbnb near the venue if the concert is more than 30m from my house and schedule rideshares. Finally, give me a list of casual restaurants with vegetarian options within 20m walking distance of the venue, otherwise schedule rideshares to the restaurants I pick." "Every night at 10pm, review my task list and make sure I completed 70% of what I committed to for the day. If I attached proofs, review those and make sure they are legit. Otherwise, request more information if it's unclear whether I met my commitment. If I don't respond with enough information, donate $50 to the Republican National Committee with my credit card and write a tweet about my failure on my Twitter account."