[FRIAM] Regulatng LLMs. (was: Bard and Don Quixote)

[Russ Abbott]

Steve Smith mentioned the Senate hearing about regulating LLMs. During the
hearing someone mentioned (sort of in passing) that it would make sense to
release such systems in stages: first to a small group of people, then to a
larger group, etc. That reminded me of the standard approach to drug
trials. Perhaps something like that could be implemented/required.

-- Russ Abbott
Professor Emeritus, Computer Science
California State University, Los Angeles


On Tue, May 16, 2023 at 11:33 AM Steve Smith <sasmyth at swcp.com> wrote:

> ...
>
> I am a couple of hours behind on the live feed Senate hearing on AI
> <https://www.judiciary.senate.gov/committee-activity/hearings/oversight-of-ai-rules-for-artificial-intelligence>
> listening in fits and starts between other things:
>
>    1. I was definitely impressed with Altman (OpenAI), Montgomery (IBM)
>    and Marcus' (NYU) thoughtful and extrapolative responses rather than
>    perhaps the more usual defensive/deflective/adversarial style that these
>    hearings often have...   I don't know who chose this particular lineup but
>    I thought both Montgomery and Marcus made a good complement to Altman.  If
>    Google and Microsoft and ??? had been there it might have reflected more
>    "competitive" or "showy" answers?
>    2. I was impressed with the Senators (compared to my fairly low
>    expectations).   Even Kennedy and Hawley managed not to do their usual
>    grandstanding and and snide sniping.   Altman handed Hawley's question "why
>    don't we just let people sue you?" (as a mode of regulation/oversight) back
>    to him quite deftly (in the style of "ass with both hands") by responding
>    simply "I know of no reason people CAN'T sue us today if we cause harm".
>    Marcus chipped in pretty well outlining how the current laws that *might*
>    apply are not well suited for many reasons.
>    3. I felt that all three witnesses walked the fine line on the
>    question of a "moratoriam" fairly deftly, acknowledging that they endorse
>    the spirit of not running headlong and thoughtlessly into the future but at
>    the same time there is no obvious practical way to implement and enforce
>    this, but that they are all enforcing their own limits on how fast (and
>    thoughtlessly?) they might roll out development to the public...
>    4. In closing Blumenthal  suggested having ChatGPT rewrite the lyrics
>    to "Don't Stop" (thinking about tomorrow (McVie-Fleetwood Mac) which I took
>    to heart.  I was not impressed with it's results and won't burden the list
>    with it.  I'm guessing Blumenthal did *not* actually do that but like
>    Quixote, simply saw the windmill and thought it might be a giant?
>
> ...
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20230516/ddee4d38/attachment.html>