[libre-riscv-dev] AI Accelerator

[Luke Kenneth Casson Leighton]

On Tuesday, February 4, 2020, Immanuel, Yehowshua U <yimmanuel3 at gatech.edu>
wrote:

> Testing testing… Is this thing on?


nooo, can't hear you.


> Basically, there is an informal desire in the ECE department at my
> university,  Georgia Tech to build an SOC. The vision actually aligns quite
> nicely with that of LibreSOC, so I had been talking to various key players
> about how LIbreSOC and Georgia Tech ECE might work together.


noow i can write an update about this.


>
> The particular lab where I work is working on delivering an AI accelerator
> takeout and it would make a lot of sense to add this into a future version
> of LibreSOC… say LibreSOC v2…


if and only if there is a budget to do so.  jacky baites may be the best
person to put an NLNet application in as he is the professor of robotics at
NTNU, TW.


> Here are some thoughts:
> Future SOCs are guaranteed to have AI accelerators - so this is a chance
> for LibreSOC to pioneer. LibreSOC can establish an ethical reference
> implementation of an ethical AI accelerator SOC.


interestingly it doesn't quite work as might be expected.  once technology
is created there is no means or method to prevent misuse.

thus the only choice of the creator is, if there is the potential for
ethical violations, not to create at all.

question: can you guarantee that with today's current deep lack of respect
for humanity, let alone other intelligent creatures, that self aware AI
will not be tortured?


> AI is currently really good at medium complexity image classification and
> absolutely excellent at language processing.


if we are restricting the scope to those tasks then that does not fire off
massive ethical alarm bells for me.

see below for details


> Our lab currently has some function RTL for an AI accelerator that handles
> CNN(vision tasks) and NLP at integer precision with up to 300x speedup by
> comparison to NVIDIAs Volta architecture…


interesting.  is that 300x for the same power consumption (at the same
geometry)?


> So, what are your thoughts?
>
> I’d basically want to mention on our front-page that future iterations of
> LibreSOC will have an AI accelerator…
>
> And you may not have noticed but AI is kinda like a buzzword right now.
>
> “Insert startup name here” + “mumble jumble something AI” = funding


the possibility of vast sums of money arriving to fund something that will
result in incalculable damage to humanity is a complete anathema to me.

very few people in the field of AI (such arrofance to call intelligence
artificial!) have the training or foresight to appreciate that humanity
needs a serious consciousness upgrade before they can handle this stuff
without causing immense damage.

they think that skynet and the terminator films are entertainment, that
"worst case" is implausible, couldn't happen in reality, meanwhile that
multi million dollar paycheck has arrived and they retire just as they give
the next group the tools to abuse AI: not their problem.

the first people dealing with uranium bricks stacked too many of them
together on a pallete.  the geiger counter started going mental and they
scrambled to kick them apart and scatter them across the concrete.

the people who developed the bomb dropped on hiroshima had *no idea* if it
would create an unstoppable chain reaction that would literally destroy the
planet, or what its actual yield would be.

likewise:

the thought of having hundreds of millions of AI accelerators for hackers
to get their hands on via a targetted virus where the OS that goes onto
snartphones is a fire-and-forget security nightmare developed by
pathological profit maximising corporations "secured" with credentials
admin/admin - to massively understate: this is not something i look forward
to empowering or enabling.

so, in short: without using arguments that summarise as "this is
implausible therefore the risk can be ignored entirely": can you
*guarantee* that the technology to be developed will not be abused?

l.



-- 
---
crowd-funded eco-conscious hardware: https://www.crowdsupply.com/eoma68