AI Personhood, Microsoft RAG Patent, PolyMarket Election
Key Points
- Yuval Harari predicts that AI “personhood” will first emerge legally rather than philosophically, with autonomous LLMs potentially being incorporated as corporate‑like entities by 2025, granting them limited legal protections but no voting rights.
- Microsoft filed a patent on “response‑augmented systems” (a rebranding of retrieval‑augmented generation) on Oct. 31 2024, but the filing is not yet granted and can be challenged with prior art, likely prompting industry pushback.
- Polymarket, a blockchain‑based (non‑crypto) prediction market, demonstrated its utility by processing $4 billion in trades and delivering election outcome data 2–12 hours faster than traditional news outlets during the recent U.S. election.
- These developments highlight emerging legal, intellectual‑property, and decentralized‑data trends that could shape how AI systems are regulated, commercialized, and utilized in real‑world events.
Full Transcript
# AI Personhood, Microsoft RAG Patent, PolyMarket Election **Source:** [https://www.youtube.com/watch?v=WKSwUYEOjTo](https://www.youtube.com/watch?v=WKSwUYEOjTo) **Duration:** 00:09:12 ## Summary - Yuval Harari predicts that AI “personhood” will first emerge legally rather than philosophically, with autonomous LLMs potentially being incorporated as corporate‑like entities by 2025, granting them limited legal protections but no voting rights. - Microsoft filed a patent on “response‑augmented systems” (a rebranding of retrieval‑augmented generation) on Oct. 31 2024, but the filing is not yet granted and can be challenged with prior art, likely prompting industry pushback. - Polymarket, a blockchain‑based (non‑crypto) prediction market, demonstrated its utility by processing $4 billion in trades and delivering election outcome data 2–12 hours faster than traditional news outlets during the recent U.S. election. - These developments highlight emerging legal, intellectual‑property, and decentralized‑data trends that could shape how AI systems are regulated, commercialized, and utilized in real‑world events. ## Sections - [00:00:00](https://www.youtube.com/watch?v=WKSwUYEOjTo&t=0s) **AI Personhood via Corporate Incorporation** - Yuval Harari argues that the first route to AI personhood will be legal, not philosophical, as autonomous AIs are incorporated as corporations—granting them corporate protections by around 2025, though without voting rights. ## Full Transcript
okay six pieces of AI news for you and
we'll start with a
banger Yuval Harari is not the person I
would expect to generate AI news he's an
author but one of the things he
suggested that I think is compelling and
correct is that the likely initial path
for AI personhood is not philosophical
because there's really no answer to that
it is not an intelligence test which
some people have suggested it's legal
fundamentally if an AI is capable of
autonomously incorporating itself it is
by definition a legally defined person
now it can't vote and that's probably a
good thing but it would have the same
protections as a corporation would
have I think that is likely to happen in
2025 I don't know if that's good I don't
know if that's bad I just think that
given the level of autonomy that we're
seeing in the space given the Simplicity
of incorpor a someone is going to figure
out how to give an llm a mission to
incorporate get it done and the llm will
be legally a
person and that's couple of months away
like that's close now I don't think
we're going to see a significant share
of Corporations B aai in 2025 simply
because there's a lot of Corporations
and not every AI use case requires this
and this will remain a bit of an edge
case but it's worth paying attention to
because anytime you talk about
artificial intelligence people
immediately start thinking about
replicants and they start thinking about
Blade
Runner
Etc and I think that yval Harari
actually had a really good thesis for
where this is going to go all right
number two Microsoft has attempted to
patent rag you know retrieval augmented
generation that we've everyone in AI has
been doing for a while they've decided
they're going to call it uh I'm getting
this right uh response augmented system
something like that they oh they change
like a couple of words right and they
call it ra not rag I'm going to link it
and they filed it on Halloween this year
October 31st
2024 the thing with patent applications
is one they're not immediately granted
so nobody building in the space with
rack has to stop what they're doing and
two you can object to them with prior
art which means if you've been doing
something that is substantially
similar you can say so and so I would
expect there to be a lot of objection
and push back to this uh because we've
been using rag for a long time in AI
certainly a long time before October
31st 2024 I'm not quite sure why
Microsoft decided they could get away
with trying to patent
that number three poly Market is worth
paying attention to uh because of how
they performed in the US election this
past week this is a blockchain based
solution it is not crypto I would argue
it is probably the first blockchainbased
solution that is not crypto that is
widely understood and used across the
world uh in this case they did $4
billion do trading on the election and
they were able to call the race 2 to 12
hours faster than official news outlets
so that's a significant Improvement in
an election with consequences that will
Echo across the globe
getting that right faster is a big
achievement uh and I want to call it out
because I think it's the first scaled up
use of blockchain we've seen outside of
crypto all right number three Google
dropped a model labeled Gemini
2.0 and it was briefly available this is
the same thing right they dropped them
they're briefly available and then we
hear rumors about them in this case uh
this model was very very fast but failed
the strawberry test uh the strawberry
test is where you ask it to write
strawberry but because llms uh do not
necessarily include logical checks by
default it can have trouble getting the
number of RS in Strawberry correct
because there's multiple RS in the same
place so it failed the strawberry task
which doesn't really argue for a super
smart model I I'm not entirely convinced
this was actually a 20 model it may well
have been an accidental leak and a
mislabel we will see we probably will
never really know but we'll see what 2.0
looks like when it actually releases I
I'm willing to bet you though whenever
2.0 really
releases that same week 01 is going to
drop I think that open AI is just
holding it back and they want to be the
last horse at the
Corral all
right what one two three four I think
we're at number four now uh yeah so 01
you were wondering about that leak the
icons in Wind windows I know you didn't
expect me to go there so you everyone
wonders like what are these things going
to be used for when they're more
advanced and one of the classic uses his
coding and that's where I get to icons
and windows because it suggests
something very interesting someone said
again it's a claim we haven't I haven't
seen a demo yet but they said that
during the brief window last week when
L1 is open they were able to code up
icons in windows with a world model
icons that had physics and weight to
them without instructing the model on a
lot of physics and weight which meant
that the model understood enough about
physics to code up icons that looked and
behaved and bounced around 3D icons like
they were in the real world without much
instruction which suggests some degree
of practical World model which would be
a big
deal okay and number five uh this one's
super nerdy but it's really important
one of the things that is hard about
training very large models is that it is
hard to physically put the chips in the
same place and it is hard to make sure
that all of the chips lock up and finish
a task at the correct point in the
training sequence with the correct
response the traditional approach to llm
training does require all the trip chips
to work in sequence and when you get a
very large number of chips like a 100,00
one fault on one chip can screw up that
entire data center until it gets fixed
now the leader here has always been
Google because Google has been running
huge data centers since before anybody
else they have developed more work
around fault tolerant architecture at
scale than anybody else I know they
actually did train Gemini on a multi-
DAT Center footprint which as far as I
know is the only model to be trained on
a multi-data center
footprint and the thing is that doesn't
automatically make it smarter and so
that's why these other models have been
able to keep up is because they have
things like synthetic data figured out
better than Google Etc but open AI knows
and anthropic knows that they have to
get multi dat Center figured out to
scale because there's just limits to
footprint on data centers and so they
need for these very large training runs
to not just have to put a million chips
down in the same place but to actually
figure out a fault tolerant architecture
that will enable them to add more chips
in scale because the other factor with
chips is that at a certain point because
of um the costs of fault tolerance and
because of the transmission costs that
go into a single state maintenance
across a huge number of chips like the
speed of light becomes a factor because
you're actually like transmitting things
back and forth you actually get
diminishing returns on extra chips to a
point where it may not be even worth it
after a certain point so what you should
hear from that is not oh no we don't
have any more sort of intelligent
scaling capability because of the speed
of light and chips that is not the
correct take the correct take is we have
some inherent architectural flaws with
assuming that you have to do training
runs with lock steps across all chips at
once and we have got a solve for that
already on the Google Side open AI is
working with Nvidia on another solve
Jensen was talking about it and saying
that he doesn't see an inherent blocker
to it and that fundament we should be
able to do multi-data center training
which would involve not having to
transmit and keep all of the chips and
lock step across all the data centers
and that in turn unlocks the laws of
scaling again and you aren't stuck on
adding another chip you aren't as stuck
when a chip fails Etc I know that was a
long sort of explanation I have a whole
link I I'll post as well but I think
it's important to understand the
underlying architecture that powers
these models because it demystifies it
it also helps us to to understand what's
really going on uh as we head into what
is likely a training run aiming for
artificial general intelligence in the
next year to two years cheers