[iampims]

Really curious about how this will pan out as everything said in the article screams at me like someone has no clue how difficult making the device is going to be, regardless of how much ”AI” runs on it.

Hardware is really hard.

[bitshiffed]

I had the exact opposite impression.

The hardware didn't sound like anything too special to me; especially with it only needing to handle audio. Fitting enough processing power to handle realtime "AI" in a package that size is the only thing jumping out at me. I'm sure they would be planning to offload that work to some 'cloud' to crunch though. (I personally dislike functionless, network-dependent hardware, but everybody seems to be doing it...)

Promising to deliver an AI that people could see as a friend is absolutely insane though. I don't see people being friends with something that couldn't complete the Turing Test, which will likely stand for at least another decade. Speech recognition and synthesis are in fairly good places, but not human interaction that isn't transparently shallow.

[sapphireblue]

>Promising to deliver an AI that people could see as a friend is absolutely insane though. I don't see people being friends with something that couldn't complete the Turing Test

This is an interesting case. Turns out, given a creative approach it is possible to persuade a human that there is another human behind the screen. See ELIZA, "Turing tests". The methods are quite similar: constrain the domain and/or creatively manipulate human's expectations (e.g. the program that "passed" the Turing Test pretended to be a 13-year boy, so human jury tolerated its errors). The question is not how to fool humans but how to make such product non-trivially useful.

I think that the best approach currently available is applied in facebook M - use human workers to interact with customers while storing all interaction data and experimenting with training state of art ML models on it to eventually replace human workers.

[neurotech1]

Not disagreeing that hardware is hard, but IMHO Its possible to get a comparable level of AI/DL performance in the "Asteria" device, using relatively available technology like Zynq FPGA[0] based boards like the Parallella[1]. I got my Parallella board from Kickstarter about 2 years ago.

[0] http://www.xilinx.com/products/silicon-devices/soc/zynq-7000...

[1] http://www.parallella.org/

[iampims]

You're right, the technology is available. But power consumption is another thing altogether. I am assuming the device will be battery powered and the usefulness of some of the sensors, like GPS, drops very quickly if you can't afford to power them on more than a few times a day.

Any wireless radio chip (BLE, Bluetooth, Ant, Wifi), if it needs to be on all the time will also have a huge impact on battery life.

I wish good luck to the Asteria team, and am genuinely curious about how they'll pull it off.

[sapphireblue]

Even if we relax hardware requirements, e.g. assume a Xeon+Titan X hardware, it is still a question if a viable conversational agent that uses deep learning to generate conversations and adapt to surroundings can be developed around it. Maybe it is possible to use DL to extract some meaning vector (or textual description) that is later used by conventional NLP chatbot to converse with user. I wonder if the quality will be good enough.

Also we don't know if the company will really develop their product to fruition, they may simply develop a good (but not viable as a product) demo and be acqui-hired by one of big players.

[dharma1]

I've got a parallella knocking about not doing anything. Did anyone write any interesting software for it in the end, that's worth putting it to use?

[neurotech1]

I think some people did, but I mainly used mine as a FPGA dev board. $99 Zynq board is still good value, even without using the custom Epiphany processor.

[sapphireblue]

... but without new Hardware there will be no new future. Just more of the same.

We really need to embrace hardware more.