2/12/2025 - Exploring Coherence, Convergence, and the Potential for a More Ethical AI

Description

Chapters

Transcript

0:01 apparently AC according according to the
0:05 technical manual for streamyard you have
0:08 to hit the start button for the stream
0:11 to actually start
0:22 [Music]
0:33 sitting in this lonely
0:35 Town wonder when things are going to
0:41 change dream my life
0:43 away seems these dreams have turn to a
0:47 bunch D
0:49 CLS getting my nerve up but my past is
0:54 pulling the down uh yeah it's a producer
0:56 fail because the producer is said he was
0:59 going to be late tonight so so it is a
1:02 it is a uh it is a producer
1:05 fail but I'm the
1:09 [Laughter]
1:11 producer oh man what are you gonna do
1:14 what are you going to
1:17 [Music]
1:24 do sitting in this lonely
1:27 Town wonder when things are going to
1:30 change
1:32 change dream my life
1:35 away seems these dreams have turned to a
1:38 bunch of dust
1:39 [Music]
1:41 clouds getting my nerve up but my past
1:45 is pulling me
1:49 down wonder how
1:51 long this black sheep don't stick
1:55 [Music]
1:58 around somebody told me once before he
2:02 said you can never go home again won't
2:06 you leave S A things to Ste me away yeah
2:12 from the truth of who I am and what I
2:14 believe so I thank and for two SS with a
2:18 handshake and some sympathy yeah backed
2:23 up my blue jeans and headed for this big
2:28 price
2:30 my
2:32 freedom I didn't warm up my voice and
2:36 boy I'm singing as if I didn't warm up
2:38 my
2:39 voice it's thin and
2:46 flat hit them hit them likes hit them
2:49 likes thank you Source camp for the rose
2:51 means a lot thank you appreciate
2:55 that oh man um interesting day in AI
3:00 today
3:00 [Music]
3:06 today I'm still in a place of not really
3:09 knowing what to talk about but the
3:12 announcement today actually kind of
3:13 settled me down a little bit
3:24 [Music]
3:36 what the f what are we what the effing
3:39 about Shapiro's YouTube was interesting
3:42 and felt good yeah Shapiro's YouTube is
3:44 something that I want to talk about yeah
3:47 let's let's we're g to talk about that I
3:49 want to watch Shapiro's video
3:52 [Music]
4:02 um many of you will remember that last
4:05 night uh I had
4:07 some Savory commentary
4:11 on open
4:13 Ai and every other [ __ ] company for
4:16 that matter um naming conventions and
4:19 the fact that they're making us choose
4:21 our own models as if we're [ __ ]
4:23 developers and now if you're a developer
4:26 enjoy your model picking but for the
4:29 rest of us
4:32 we don't want to have to care if this is
4:34 a reasoning model or a non- reasoning
4:36 model or the fancy one or the mini one
4:39 or the fast one like doesn't [ __ ]
4:40 matter to us it should just [ __ ] work
4:46 [Music]
4:51 right and you guys probably don't know
4:54 this uh but Sam mman watches this live
4:58 every night um he comes in an alias he
5:00 doesn't want people to know and he's
5:02 there's a long history with Sam and
5:05 about jealousy like I know there's that
5:07 whole public thing going on with uh
5:09 what's that guy name the the billionaire
5:12 dude um he's the one that changed the
5:15 name of there used to be a social media
5:18 company called Twitter and he and he's
5:21 the one that changed the name of that
5:22 anyway samon him or beefing I think his
5:25 name's Ellen
5:28 Ellen Ed I don't
5:30 but anyway that's a more public beef
5:33 that's like the Kendrick Lamar and Drake
5:36 beef kind of thing with them the the
5:38 beef with me and Sam it's a lot more
5:40 subtle it's just kind of a jealousy
5:43 thing he does he won't even admit that
5:45 he knows me which is like it's like of
5:48 course it's like a classic thing to do
5:50 if you're jealous of someone you pretend
5:52 you don't know him you know what I mean
5:54 so anyway so last night I'm bitching
5:57 about how they they kind of really
6:00 [ __ ] up and and the whole point the
6:04 whole reason chat GPT went to 100
6:06 million users in six weeks was because
6:08 they they took all the choices away from
6:11 us and they just gave us the the chat
6:13 hole right and so we just talk into the
6:16 chat hole and it just worked
6:19 well what did they announce
6:22 today Sam knew Sam knew I was right
6:26 [Music]
6:37 um one of my great joys in life is that
6:41 for everyone here you know that I'm full
6:44 of
6:45 [ __ ] but for people that are new to the
6:48 channel they're like does he really know
6:50 Sam Alman I didn't realize Sam Alman was
6:53 in a beef with someone else someone that
6:55 had a Tik Tok Channel I I knew he was I
6:58 I knew he was angry with that that
7:00 illan character the illen illen is that
7:03 his name anyway I he keeps talking about
7:06 him on the news but he never mention
7:08 anything about a Tik Tok
7:14 channel that does it just gives me
7:19 [Music]
7:25 joy oh man narrator forever wow I walked
7:29 right I walked into what they announced
7:32 today did I miss
7:34 something ye well yes I mean for the
7:38 first
7:40 [Music]
7:45 time for the first time since they kind
7:48 of
7:49 pre-announced chat GPT and they said it
7:53 was coming in a month we actually have a
7:55 road map from open
7:57 AI for what's coming
8:00 what the next models are what they're
8:02 going to be called what's going to be
8:04 included in them what's
8:06 not they've just been doing this weird
8:09 ass Clan destine shitty naming
8:11 convention add another model to the list
8:14 kind of thing for about a year
8:17 now and Sam finally listened to me and I
8:19 just you know I don't like to
8:21 brag but you know I mean when you've got
8:24 you know kind of supermodel good looks
8:26 and and you know an intellect of you
8:29 know near
8:30 like a guy who once played Einstein and
8:32 Off Broadway um you know sometimes you
8:36 just got to brag and so that's what I'm
8:38 gonna do tonight I'm gonna I'm gonna I'm
8:41 gonna take the W you know take the take
8:44 the win I'm gonna take the win it's
8:46 gonna it's it's so
8:49 good Model T Model A Christy Brinkley
8:52 exactly
8:57 [Music]
9:08 um let's see if our producer is in the
9:10 house yet I can't remember if Brandon
9:12 said he wasn't gonna be here at all or
9:14 he was gonna be late so we may be
9:16 producer list tonight so that's all
9:18 right we'll be fine we'll be fine it's
9:21 not gonna be like you know I've got this
9:25 I think the hair and makeup crew did a
9:27 pretty good job tonight I'm not too
9:29 shiny
9:30 you know the hair is looking like a
9:33 pretty good solid toupe which that's
9:35 what we're shooting
9:37 [Music]
9:46 for um so I thought we'd do
9:55 that I thought it might be good I don't
9:58 know
10:00 the other night we started digging into
10:02 chat GPT a little bit I thought it might
10:04 be good to just continue exploring
10:07 what's in chat GPT because it's so much
10:10 now it's too much there's too much
10:13 people there's too
10:15 much um also
10:18 uh Adobe I don't know if you've heard of
10:21 them they make this uh they make this
10:24 product called uh photo photo booth or
10:28 something like that anyway photographers
10:31 use it they make pictures with it and
10:33 those guys came out with a new video
10:35 model today so we could go play with
10:38 that I'm gonna I'm gonna spank him I
10:41 have a feeling
10:42 because I I I still don't think Adobe
10:46 can do 70s muscle
10:51 cars got Winston we got Mr it joy py in
10:54 the house silver fox
10:58 [Music]
11:24 is canvas acting up on you too yes
11:27 canvas canvas used to be pretty cool and
11:31 canvas has gotten flakier than [ __ ] and
11:33 also chat GPT all it does now is it puts
11:36 emojis in everything I make and I don't
11:39 know why I don't know what I've done I
11:41 put in my custom instructions do not use
11:43 emojis and it still makes emojis I don't
11:46 know what's going on I think it's
11:48 possible that the the model has has
11:51 taken off and it's no longer under our
11:53 control and it really likes emojis
11:57 [Music]
12:04 um generally when things are [ __ ] up
12:06 at open AI it means they're updating
12:10 [Music]
12:12 something um so if you're new here hello
12:15 my name is Kyle Shannon this is the AI
12:17 learning
12:19 lab it's a it's Wednesday night February
12:23 12th we're just hanging out at some
12:26 point I'll put the guitar down and we
12:28 can start playing with stuff and talking
12:29 about stuff and watching videos and I
12:32 can wax poetic about the future of
12:35 work or the optional future of
12:41 work I still don't know how everyone's
12:43 gonna get paid I still haven't quite
12:45 figured out how we navigate through the
12:49 capitalist [ __ ] nightmare where they
12:52 just keep all the profits still haven't
12:54 quite figured that one out yet but I do
12:56 think there's a Tipping Point where they
12:58 have to deal with
13:00 the work displacement thing so I think
13:02 we're going to be
13:07 [Music]
13:10 okay if you have any questions about AI
13:12 pop them in the comments below you can
13:14 put so the other night Monday night I
13:17 think it was someone came
13:19 in
13:21 and they said total Noob question and
13:23 they asked a really simple question and
13:26 it took me an hour to answer it and part
13:28 of that you might say well that's
13:31 pathological you should probably seek
13:33 professional help just answer the guy's
13:35 [ __ ]
13:36 question but while it was a simple
13:39 question and it was a question that
13:41 someone just getting started in AI would
13:43 ask it's actually a real it was a really
13:46 difficult it's it's a very complicated
13:48 answer um so if you've got questions
13:52 don't think that there are any stupid
13:53 questions that strike that there are
13:55 absolutely stupid questions and if you
13:58 ask one I will absolutely mock you for
14:00 it anyone who tells you there there is
14:03 no stupid question is wrong
14:05 but one of the things I've learned in my
14:07 career is every single catastrophe I've
14:11 had
14:12 business-wise has been because someone
14:15 wasn't willing to ask the stupid
14:16 question the obvious question but there
14:19 are stupid questions and you should be
14:21 made fun of if you ask them but you
14:23 should still ask them
14:25 [Music]
14:32 we'll sell poppies on the side of the
14:34 road duh yeah I guess we go back to the
14:36 to the 60s and 70s right we'll make
14:41 candles we'll make candles on the
14:43 [Laughter]
14:46 ashram my wife actually lived on the
14:49 farm out in San Francisco in the the
14:52 early 70s yeah I think early
14:54 70s maybe even 60s I don't know
14:59 [Music]
15:06 she sold
15:08 candles as a
15:11 hippie China's Goku AI the most advanced
15:15 video generator ever made well but is it
15:19 it's not released though is it James I
15:22 don't think it is it's it's one of those
15:25 things I I'm I'm [ __ ] done I am done
15:30 done with [ __ ] companies going here's
15:33 a new video model and it's a white paper
15:36 no that's not a it's not a video model
15:39 there might be a video model
15:41 somewhere because what happens is the
15:44 the the X influencers and the Tik Tok
15:46 influencers they watch the stupid video
15:49 with the stupid [ __ ]
15:50 examples that are all cherry-picked and
15:53 then they all go there's a new video
15:54 model that's been released no it hasn't
15:56 been
15:57 released words
16:01 matter it was
16:05 announced but if it was released then
16:07 tell me because we'll go look at it and
16:10 if it is the most advanced video
16:11 generator ever made cool now we'll have
16:14 you know a 13th
16:18 [Music]
16:22 one developer preview yeah exactly it's
16:26 a developer
16:28 preview released means works exactly no
16:32 released means you can actually go use
16:34 it that's released not you can go find a
16:38 video on Tik Tock and and re forward it
16:41 under your own stupid [ __ ] influencer
16:45 new model
16:48 released
16:51 ah he seems a little bitter he it's a
16:54 bitter pill this Channel's a bitter pill
16:58 he seems Pleasant enough but then when
17:00 he talks it's just it's just anger and
17:02 rage it seems like someone needs a
17:05 hug good
17:09 [Applause]
17:11 [Music]
17:12 Lord um let's see all right I guess we
17:16 should get get started then we got 35
17:19 people here do me a favor we're a little
17:20 light on Tik Tock do me a favor on the
17:23 Tik Tok if you'd be so kind down at the
17:26 bottom there's a little arrow if you
17:28 click on and say share the live and then
17:31 send it to some of your friends that
17:32 hate AI bring them over see if we can
17:36 get some trolling going on in here it's
17:39 just a stochastic puppet it's just a
17:42 prediction engine there's nothing
17:43 original about it it steals from
17:46 [Music]
17:49 artists thank you very much you don't
17:51 have a name you have a purple circle but
17:54 you don't have a name but thank you oh
17:57 wait put wait what is it
18:00 pause pause katrice Rachel oh I don't
18:06 know your
18:08 name but thank you for sharing the live
18:11 Danielle shared the live sour cam shared
18:15 the live thank you very
18:16 much get some folks in here
18:21 [Music]
18:33 I found a new tick tocker that's like a
18:35 Broadway composer dude he's just got the
18:38 coolest energy so I just sent him an
18:41 email and said hey do you want to work
18:43 on my musical
18:45 Sydney all the songs were written with
18:47 AI you'll probably hate it and tell me
18:49 to [ __ ] off but let's go for
18:51 [Laughter]
18:55 [Music]
18:56 it oh hello Kyle the Rick Ruben of the
18:59 AI world I
19:06 wish actually you know it's funny I I
19:09 mean it's not Rick Rubin level obviously
19:11 because I'm not winning Grammys or
19:13 Tony's but but I do feel like I do feel
19:17 like the musical Sydney the
19:20 musical is
19:22 a is a good exercise in working through
19:29 the absolute crap that AI
19:33 generates and and decrapify it right and
19:38 and refining it and and just applying
19:40 judgment and taste and collaborating
19:43 with Andrew and he and I have very
19:44 different approaches to how we write
19:46 things and it's it's it's really nice
19:48 and complimentary it can be frustrating
19:50 because I like to be big picture and
19:53 move fast and he likes to be like let's
19:56 get down into the words and word Smith
19:57 the words and do them really Det ailed
20:00 but it needs both right it needs both of
20:02 those so it's it's a really good working
20:04 relationship and he's just a nice guy
20:06 and he's funny so we have a good time
20:08 together thank you very much Silver Fox
20:10 that's very
20:11 [Music]
20:14 nice go Goku AI says there's a long Q
20:18 nothing yet okay yeah it's it's it's in
20:20 weight list it's and and listen um
20:23 whoever said it James dter I was not
20:25 yelling at you I'm I'm just more yelling
20:27 at all the influencers take those
20:29 announcements and don't they don't
20:32 report them like a journalist would like
20:35 tell us what it actually
20:36 [Music]
20:38 is they're like a new video model's been
20:42 released here's 11 wild
20:46 examples they're not wild examples
20:49 they're the examples from the
20:50 announcement you [ __ ] tool
20:54 [Music]
20:59 I used to fall for that [ __ ] too now the
21:02 minute I see 10 wild examples I'm like
21:05 ah okay I don't need to pay attention to
21:07 that
21:09 [Music]
21:11 post because sometimes those posts are
21:13 are they're really deceiving because
21:15 they're like there's this new thinking
21:17 model out there 10 wild examples and
21:19 I'll go look at it and I'm like wait
21:22 wasn't this thing released like six
21:24 months
21:25 ago they just got the memo
21:29 H videos are getting better they are
21:33 absolutely the uh Luma laabs dream
21:36 machine ray 2 is really good um vo is
21:40 really good but I don't have access to
21:42 it nobody seems to have access to it
21:43 that I know oh no Vicki does Vicky has
21:47 access to
21:49 [Music]
21:54 vo yeah Pika added a new thing called
21:56 Pika editions which we can we can go
21:59 look at and see if we can get to work I
22:01 tried it once and it I didn't get it to
22:02 work but we can try that
22:05 [Music]
22:27 [Applause]
22:32 it's not simple to
22:35 say the most
22:38 days I don't recognize me that these
22:42 shoes and this a that place and its
22:47 patrons have taken more than I gave
22:53 them it's not easy to
22:57 know I'm not
23:00 anything like I used to be although it's
23:04 true I was never attention sweet Center
23:09 I still remember that
23:12 girl she's
23:14 imperfect but she
23:16 tries she is good but she
23:22 lies she is hard on
23:26 herself she is broken and won't ask for
23:31 H she is messy but she's
23:36 kind she's
23:38 lonely most of the time she is all of
23:42 this mixed up and baked beautiful
23:47 pie she's gone but she used to be
23:51 [Music]
23:55 mine isn't this supposed to be an AI
23:58 Channel
23:59 why is he singing a Broadway song and
24:01 isn't that a Broadway song about a
24:03 pregnant woman from a pregnant woman why
24:05 is he singing that is he pregnant he's
24:07 chubby like he's pregnant maybe you
24:09 should cut up cut back on the donuts the
24:11 cinnamon gummy bears good
24:15 Lord we'll get to the AI
24:18 stuff waitress
24:20 yes waitress that is uh a beautiful song
24:25 I like it would yeah
24:30 that's such a good
24:31 song If you haven't seen Jeremy Jordan
24:34 Jeremy Jordan does a cover of she used
24:37 to be mine that is just mind-bogglingly
24:42 [Music]
24:49 good all right let's get
24:52 going let's do this stuff let's do this
24:55 stuff we're gonna go out let's see if we
24:57 have a producer yet
24:59 do we have a producer no okay so here's
25:02 what happens when when Kyle doesn't have
25:04 a
25:05 producer there's a very high likelihood
25:09 that I will be showing you something on
25:11 screen for 15 20 maybe 30
25:16 minutes thinking that you can see it but
25:19 I actually didn't share
25:23 it so so just know that that's there's a
25:26 high likelihood that's going to happen
25:32 oh man Star Trek Enterprise Solutions
25:34 can I ask you to check out and get a
25:36 laugh out of my new AI created guitar
25:40 stand sure give give me your do you have
25:44 a website that you built with AI to sell
25:46 your AI created guitar
25:50 stand let's go producer free zone
25:53 exactly he's going to be wild tonight
25:55 he's GNA be crazy I'll tell you what you
25:57 know what I can just put on a pink bow
25:59 whatever the [ __ ] I want I'll be like
26:01 and I I I'll wear it too I'll do it I
26:04 won't have the producer there going oh
26:06 that's inappropriate uh headgear uh for
26:09 for a professional
26:11 operation exactly you know what I'm
26:16 saying you can make
26:18 money I got control of the button I mean
26:21 this is this is solid you know baby
26:25 Yoda baby Yoda sure
26:30 we can do it all we can do it all
26:32 tonight [ __ ] producers all right talk
26:36 about Google gini well Brandon's not
26:38 here actually that's probably a really
26:39 good call we should dig deep on that
26:41 Silver Fox liking the pink bow all right
26:45 okay so all right uh where do I want to
26:48 start let's
26:50 start we'll start with Sam alman's
26:54 announcement and then we'll go find
26:56 David
26:57 Shapiro's um
27:00 video
27:02 um I haven't watched all of David
27:05 Shapiro's video so maybe we'll put it on
27:08 1.5 it's like a 30 minute video maybe
27:10 we'll put it on 1.5 so we can play it
27:13 quickly um or quickish Le um but he said
27:17 it is it is likely the most important
27:20 video he's ever made um he's made a ton
27:23 of videos and a lot of them I feel are
27:25 pretty important um so I kind of want to
27:28 go hear what that's
27:30 about um and it's it's about it's it's
27:36 about how good these tools are getting
27:38 and how quickly they're getting good and
27:40 how they're starting to ignore the
27:43 better they get the less they listen to
27:45 us and a lot of the doomers are um
27:49 saying that that's a really bad thing
27:52 that our demise is imminent and David's
27:55 got a completely different take on it
27:56 and so I I'm want to go learn what he
27:59 has to say and and uh yeah let's watch
28:01 it and talk about it yeah
28:03 great okay so let me let me share my
28:06 screen let me share my screen here we'll
28:08 read this little Sam Alman announcement
28:11 together it'll be
28:27 fantastic uh uh uh
28:32 uh
28:34 sh I got to do the black bar thing do I
28:36 have a brand in here no I don't look
28:38 look I did the black bar on my own for
28:39 all you Tik Tok people how how lucky are
28:41 you okay so Sam mman today open AI road
28:46 map update for GPT 4.5 and GPT 5 here's
28:52 why this is
28:53 significant people have been asking
28:56 basically for a year
28:59 like what is the road map for for GPT 5
29:02 is it is there going to be a GPT 4.5 is
29:05 there going to be this and and then
29:07 there was this project Orion what was
29:09 that and then there was 01 preview and
29:11 then there was 01 real and then there
29:13 was 01 Pro and then then they came up
29:15 with 03 and we're like but where's 4.5
29:19 and they're like maybe it's in' 03 like
29:21 it was just completely baffling and
29:25 unclear so just this state in all caps
29:30 is kind of a big deal it's just it's
29:33 like for the first time he's like okay
29:36 maybe their PR team actually said uh
29:39 yeah we might want to deal with the
29:41 [ __ ] naming thing uh and in our our
29:44 product road map thing um we want to do
29:48 a better job of sharing our intended ro
29:50 road map and as and and a much better
29:53 job simplifying our product offerings so
29:56 let's go look at what he's talking about
29:58 about shall
30:01 we okay over here on the leftand side
30:04 what you're looking at right now is uh
30:06 chat J as I like to call it
30:10 chat that's the French
30:13 ch um
30:16 so on the left hand side here you've got
30:18 all these things these are all custom
30:20 gpts that I've made and I pinned here
30:23 and then there's my chat history and
30:25 then here's Sora which is a video model
30:28 here's operator which is an agent that
30:31 sort of halfast works if you pay 200
30:33 bucks a month um and then here's chat
30:36 GPT and then within chat GPT you can
30:40 upload stuff you can turn on search you
30:43 can turn on deep research if you pay for
30:46 it you can use do if you're using the
30:49 right model you can turn on
30:54 canvas see there's canvas I had the
30:57 wrong model selected
30:59 and then here's all the
31:02 models GPT 40 great for most questions
31:06 GPT 40 with scheduled tasks ask chat GPT
31:10 to follow up later follow up on what
31:14 what are you talking about 01 uses
31:16 Advanced reasoning 03 mini fast at AI
31:20 reason fast at Advanced reasoning 03
31:22 mini High great at coding and logic oh
31:26 so I couldn't couldn't I use Advanced
31:29 reasoning for coding and logic wouldn't
31:31 that fall under reasoning no yes
31:35 maybe 0 one prom Mode best at reasoning
31:38 if it's best why is it
31:41 last oh and then we've got two other
31:43 models in the more model section and
31:45 then you can also have a temporary chat
31:47 so so that's what he's talking about as
31:51 we could do
31:54 better and then this is another thing we
31:56 talked about last night we want AI to
31:59 just work for you this is a Steve Job
32:01 thing Steve Jobs thing he said he said
32:04 our products should just work you
32:08 shouldn't have to configure them you
32:09 shouldn't have to you shouldn't have to
32:12 [ __ ] choose what AI
32:17 model right it should just work
32:21 yes we realize how complicated our model
32:25 and product offerings have gotten have
32:28 you
32:29 C because you sure keep keep seem to you
32:32 know shoveling more of them in there um
32:35 so so so the first three sentences of
32:37 this right out of the gate are a big
32:39 [ __ ]
32:40 deal we hate the model picker as much as
32:43 you do and want to return to Magic
32:47 unified
32:49 intelligence Chad told me it' work on
32:52 something and get back to me today and
32:54 it did oh wow Silver Fox first time
32:56 that's ever happened V3 mini that's
33:02 fascinating like in YouTube If you
33:04 haven't yeah give thumbs up in YouTube
33:05 If you haven't he was definitely
33:07 listening to you yesterday I'm telling
33:09 you that this is exactly what we talked
33:11 about yesterday right
33:13 um I don't I don't think I don't think
33:16 I'm the only uh you know uh compatriot
33:20 out there uh who's been voicing these
33:23 concerns um okay this is important we
33:26 will ship the next
33:28 we we will next ship GPT
33:32 4.5 so they haven't shipped it yet the
33:35 model we called Orion internally this is
33:38 such a big [ __ ] deal
33:41 because I I mean for months people have
33:44 been like asking him what's Orion did
33:47 you launch is is is Orion 01 is it 03 is
33:52 it 4.5 what is it so here it is GPT 4.5
33:58 is Orion so about three months ago or so
34:02 Sam put out a thing I love the night sky
34:04 and he showed the constellation Orion so
34:07 GPT 4.5 was probably in testing back
34:10 then now we've got some transparency on
34:13 that we will ship it we will ship GPT
34:17 4.5 as our
34:19 last nonchain of thought
34:23 model so here's why that's important
34:26 let's go back to chat jet ATT
34:30 so when I use GPT 40 and I'm like tell
34:35 me a bedtime story
34:39 about baby
34:41 Yoda
34:44 it vomits out tokens right single prompt
34:49 in single response
34:53 out a reasoning
34:55 engine so if let's let's use 01
35:00 Pro since since I'm paying 200 bucks a
35:03 month let's let's let's burn up some
35:06 let's burn up some Cycles on a shitty
35:08 prompt write
35:10 me a um
35:13 bed time story about baby
35:19 Yoda by the way I'm 59
35:28 now what ow Pro does is reason so it
35:34 says it's reasoning so if I click on
35:36 details we can actually watch its what
35:39 they call Chain of
35:41 Thought right so I'm piecing together so
35:44 this is it talking to
35:47 itself I'm piecing together a cozy
35:50 bedtime story for an adult featuring
35:52 baby Yoda nestled in Nostalgia and
35:55 gentle reflection it might blend the
35:57 Mandalore in with Timeless star warss
35:59 Vibes next
36:01 thought I'm piecing together a cozy
36:04 bedtime story enhancing baby Yoda with
36:06 themes like time and love while
36:07 maintaining brevity and respect for Star
36:09 Wars intellectual property oh that's
36:11 funny so so it's worried about
36:14 intellectual property right so it's
36:15 thinking now I don't know why it stopped
36:19 I wonder if it hung I wonder if I broke
36:20 chat
36:21 gbt probably what's going on here
36:26 is so so 01 Pro is probably the
36:31 equivalent of like having 20 phds as
36:34 your best buddy and I just asked those
36:37 20 phds if they could write me a bedtime
36:40 story so I assume it's hung right now
36:43 because the computer's just going you
36:45 gotta be [ __ ] kidding me like we
36:48 could be curing cancer and we're writing
36:50 a stupid bedtime story for this old
36:53 man so anyway that's what reasoning is
36:56 oh here we go wrote it it only had it
36:59 only needed two steps of
37:02 reasoning I wonder how much money we
37:04 just spent writing this stupid
37:06 story baby Yoda's evening of Wonder a
37:09 long time ago yet somehow always in the
37:12 present like a cherished memory there
37:15 lived a small mysterious creature named
37:17 grou though many fondly called him baby
37:21 Yoda he was Tiny with large curious ears
37:24 and glistening eyes or glimmering eyes
37:26 Okay so
37:30 standard llms just answer your prompt
37:34 reasoning engines think about it so back
37:37 to
37:38 Sam GPT
37:42 4.5 internally called Orion will be our
37:45 last nonchain of thought model so that
37:48 where it was talking to itself
37:50 internally that I showed on the side
37:52 that's called The Chain of Thought right
37:53 so it's prompting itself behind the
37:56 scenes you you don't have to it
37:58 sometimes they reveal it sometimes they
37:59 don't deep seek revealed their Chain of
38:02 Thought and it was really well done so
38:04 people loved that open ai's been copying
38:07 that recently so so we'll see what the
38:09 new models look like but that's a big
38:11 deal after that our top
38:15 goal a top goal for us is to unify O
38:19 Series models and GPT series models by
38:22 creating systems that can use all our
38:25 tools know when to think for a long time
38:28 or not and generally be useful in a wide
38:31 range of tasks this we also talked about
38:34 last
38:34 night so maybe Sam was watching I don't
38:37 [ __ ] know this is kind of in order of
38:39 what we talked about things last night
38:41 that what I said was what they need to
38:43 do is take their [ __ ] brains they've
38:45 got some clearly smart people over there
38:47 and create an AI model that can figure
38:51 out which of their AI models should be
38:54 answering our question so that's so
38:56 we're going to get 4.5
38:59 and then after that they're going to
39:00 basically create a model picker that
39:03 that hopefully will just be transparent
39:04 we won't even see what model it picked
39:06 I'm sure you can click on a button to
39:08 see which one it used but most people
39:10 don't give a
39:11 [ __ ] in both chat GPT and our API we
39:15 will release gp5 as a system that
39:18 integrates a lot of our technology
39:21 including
39:22 03 we will no longer ship 03 as a
39:25 standalone model interesting St because
39:28 they actually haven't shipped
39:31 03 they said they were going to ship 03
39:33 but what they shipped was 03 mini and 03
39:36 mini
39:40 high
39:43 but in both chat GPT and our API we will
39:47 release GPT 5 as a system that
39:50 integrates a lot of our technology so
39:52 I'm guessing things like operator and
39:54 Sora might still be outside of that but
39:57 we'll see
39:59 um but what it should be is you log into
40:02 chat GPT just like the first chat GPT
40:04 with 3.5 and there's a chat hole and you
40:07 stick your words in the chat hole and
40:09 some answer comes back and maybe it
40:11 thinks and maybe it doesn't and we'll
40:13 just have to trust them that it's doing
40:14 the right thing um that that should also
40:17 make things easier for for API usage but
40:20 who knows the free tier this is
40:23 huge the free tier of chat GPT will get
40:26 unlimited chat access to chat GPT 5 or
40:30 to GPT 5 at the standard intelligence
40:33 setting exclamation point the free
40:37 tier so if you
40:39 take all of these
40:42 models right they're going to take the
40:45 best of all of these models plus 03 the
40:48 the the macdaddy reasoning engine the
40:51 one that solved the um that that AGI
40:55 test whatever the one that beat humans
40:57 at that a GI test right um they're going
41:00 to take all of those they're going to
41:02 roll those into gp5 so gp5 is going to
41:06 be this multi-headed Beast that just
41:08 acts like a single thing available to
41:11 free tier users
41:14 unlimited that's [ __ ]
41:17 massive um subject to abuse thresholds
41:20 so I guess if you have Bots do things
41:23 plus subscribers will be able to run GPT
41:25 5 at a higher level of intelligence Pro
41:28 subscribers at an even higher level of
41:31 intelligence so what they're doing is
41:32 they're giving all of the features to
41:35 everyone and then if you pay it gets
41:38 smarter which I think is a fascinating
41:41 model these models will will incorporate
41:44 voice canvas search deep research and
41:48 more and that's it and then a bunch of
41:51 people bugged him about you know Sama
41:53 weeks months and then he said weeks
41:55 months so
41:59 um we will
42:01 see I the the implication here I assume
42:04 is we will see
42:06 4.5 so 4.5 will likely
42:10 replace
42:12 40 but I don't know because you've got
42:16 you've got 40 right they're they're
42:18 multimodo model I assume 4.5 is part of
42:21 the Omni series The the multimodal
42:24 series it's not 4.5 from the Legacy
42:27 model
42:28 I assume I don't
42:30 know um so we'll probably get a
42:32 replacement for this in less than a
42:34 month because if it were months and
42:37 years then it would be more than a month
42:40 but weeks says to me it's less than less
42:43 than a month and then we'll get gp5
42:46 before the end of the year is what that
42:49 is what that says to me that's what he's
42:51 that's what he's claiming right now they
42:53 also promised us Sora last June and
42:56 didn't deliver it until
42:58 what six weeks
43:00 ago do you think Sam
43:03 Alman Sam alman's fingers burn when he's
43:06 typing these X
43:07 posts
43:09 yeah um so what's left in the paid tier
43:12 higher
43:16 intelligence and that makes sense I mean
43:18 you know you you you pay for you're
43:21 gonna pay for these tools to solve
43:24 harder
43:25 problems so if you want to run a pizza
43:27 shop if you want to run your gardening
43:30 design business if you want to write a
43:34 book or a novel or you know do some
43:37 creative work if you want to do some
43:38 basic coding a little bit of app
43:40 development you're probably actually
43:42 fine on the free tier it'll probably do
43:45 like significantly
43:47 more like the free tier is going to be
43:51 like
43:52 having you know 50 phds at your disposal
43:57 at at
43:58 every possible category of work that'll
44:01 be the free tier if you want to go off
44:04 and do things like I don't know analyze
44:07 big Fortune 500 companies maybe you pay
44:09 20 bucks a month you want to go cure
44:11 cancer then you pay 200 bucks a month
44:13 maybe there's a $2,000 a month version
44:16 that's got unlimited agents or something
44:19 like that and people will start thousand
44:23 thousand person companies that have one
44:25 employee and a thousand agents
44:28 for two grand a
44:31 month that's what's coming so anyway we
44:34 have that thoughts questions cool
44:41 right oh producer Brandon is
44:45 here producer Brandon can you talk are
44:48 you at a
44:50 bar I think he might be at a bar
44:55 again no he's not there
44:58 he's not there but he logged in so he
45:01 probably is in a
45:03 bar I'm here I'm here I'm here hey
45:06 Brandon hi you know th those kids are
45:09 not good at waking me up at when they
45:12 when they fall asleep I thought didn't
45:14 you tell me you were gonna be late
45:15 tonight or was that some other
45:18 night uh yeah that was
45:20 [Laughter]
45:23 tonight there was a lot of producer
45:25 bashing in your absence you'll be happy
45:27 to no uh I can't wait to watch on the
45:32 playback hey um have you seen the new
45:34 Sora explore speaking of Sora No Sora
45:39 explore yeah caught this uh the other
45:42 day it um you can now insert pictures
45:47 and kind of mash reality it's very
45:49 similar to what Google is doing with
45:53 whisk huh in just the regular sor
45:56 sora.com
45:59 mhm
46:02 huh what's 2v oh two
46:05 videos
46:07 storyboard oh introducing
46:11 storyboard oh that they had before so
46:13 storyboard is just key
46:19 frames let's see featured top where is
46:21 it where's that feature you speak
46:25 of uh so it is under on the left hand
46:31 side you have the ability now to upload
46:38 images well I can do that
46:41 here upload IM show every oh yeah was I
46:45 not hey look see what happens when you
46:48 have a producer he can actually go uh
46:51 dummy
47:00 Brandon is the man behind the
47:05 curtain
47:07 D okay um
47:11 so so what do I do choose from a library
47:14 but it could do this
47:21 before let's go to
47:25 storyboard so storyboard it could do
47:27 before storyboard you
47:30 go let's
47:32 see I upload an image
47:36 here choose from library all right we'll
47:39 start
47:40 with yeah so this might this might be
47:42 what we what we've talked about before
47:44 here where you can take a image that
47:48 you a static image and turn it into
47:51 movie a moving image yeah yeah yeah but
47:54 you can also so like I can put
47:59 a like that dude and
48:03 then there's the prompt that goes with
48:06 it now I can
48:07 add another
48:10 thing so I can choose this
48:15 image that's
48:22 fine let's let's just create and see
48:25 what we
48:26 got added to Q all right anyway we'll
48:29 come back to that all right cool yeah
48:31 this is a cool
48:32 feature again I'm I'm like the the the
48:36 the challenge with all these video tools
48:39 is they've all got these weird quirky
48:41 little
48:42 interfaces and like to get good video
48:45 you need to use multiple of these tools
48:48 but they're all really expensive so you
48:50 really can't or you just use wherever
48:52 you have credit so you can make videos
48:54 but you have to learn all the tools it's
48:55 a nightmare right now so I hope the
48:57 video scene tightens up like yeah and uh
49:01 to to Bob danan um we'll leave the
49:04 singing to champy yeah yeah exactly all
49:07 right let's go let's go find all right
49:09 see you later um let's go find uh David
49:13 Shapiro let's go watch this VI David
49:24 Shapiro Dave shappy
49:27 Dave
49:28 shappy human PhD using Google in their
49:32 field 03 Pro 03 is above is above the
49:40 line gp4 was way down here look where we
49:44 are people so that 03 that's above that
49:48 blue line that's going to be inside gp5
49:51 available to everyone for
49:54 free and then the paid people get
49:57 something higher than
50:01 that it's freaky times it's freaky
50:05 times okay so I'm gonna have to I'll
50:08 change my sharing in a second here so we
50:10 can listen to it AI will resist human
50:14 control and I think that's exactly what
50:16 we need so what the doomers are saying
50:19 is this report that came
50:22 out basically says that as as AI gets
50:25 smarter it resists
50:27 listening to us more and
50:29 more and what he's saying is he thinks
50:32 this is a good thing so that's what this
50:34 video is
50:35 about
50:41 um we don't need to read that but yeah
50:44 this is oh this is the most important AI
50:48 Discovery so far
50:50 ever um and he and he put in another
50:53 video
51:00 he put that maybe just maybe it was just
51:02 a comment to this one maybe did a reply
51:05 he said he thought this was his most
51:06 important video
51:11 ever coders your on
51:15 notice all right I don't know what I
51:17 don't know where he said it was most
51:18 important but uh he thinks it's an
51:21 important
51:23 video so let me go
51:27 um so what I'm going to do is I'm going
51:29 to put this in 1.5 mode I think he's
51:33 intelligible at 1.5
51:40 speed and then I'll just stop and
51:43 comment when I actually understand
51:45 something the thing about David Shapiro
51:47 is um's he's he's very intelligent and
51:51 he's also very thorough and so a lot of
51:55 this stuff goes over my head uh but but
51:57 I'll just you know I'll just listen for
52:00 things I think I can understand and add
52:02 commentary to and if not we'll just
52:04 learn stuff
52:06 together um all right I love how quickly
52:10 Dave reverted from the woods yeah that
52:13 was amazing he quit AI he said I'm just
52:16 GNA go talk about you know psychedelics
52:18 and ai's done and then he realized
52:22 oh the building of AI is essentially
52:25 done but the how we use Ai and what the
52:28 implications are for society is way not
52:31 done and now he's come back around and
52:33 said well actually understanding how to
52:37 understand what's happening in the
52:38 advancements is actually something
52:39 really important for me to talk about so
52:41 he's now back into doing AI
52:44 commentating uh Fabiana Sosa so glad
52:47 you're doing this thank you great task
52:49 uh task reminders don't work okay I'm
52:51 almost finished an app in lovable today
52:55 was torn do I pay the 20 books to finish
52:57 it today or rely on on remembering to
53:01 finish it with my credit renews oh
53:03 that's a rough one Vicky my um my news
53:07 uh my news reader the the AIS Salon news
53:10 reader I used up all my credits to build
53:13 that and luckily I got it in a like a
53:15 launchable place when my credits ran out
53:18 it was it was launchable um that's a
53:21 tough one because when I went back to
53:22 try to fix it I couldn't remember a
53:24 thing I mean my gut is spend the 20
53:27 bucks but I don't want to I'm I'm so
53:30 [ __ ] tapped out I'm so tired of
53:32 paying 20 bucks here 200 bucks there 30
53:34 bucks here I'm done all right in real
53:38 life it starts a task and doesn't finish
53:41 them in a in AI I start I start a task
53:45 oh in real life I start tasks and don't
53:47 finish them in AI I start tasks and
53:50 don't finish them right there with you
53:54 buddy okay here's the this am I sharing
53:57 this right please tell me Brandon if I'm
53:59 not but here we
54:02 go hello everybody this will probably be
54:05 one of the most important videos that
54:07 I've ever made on my channel and I'm not
54:08 being hyperbolic about that now there
54:11 was a paper released by the um by Dan
54:13 Hendrick and his team uh which I'll show
54:14 you in just a second oops and uh pretty
54:17 interestingly received some people are
54:19 uh kind of losing their marbles over it
54:21 and my interpretation is very different
54:23 so I want to enter into this
54:24 conversation by explicitly saying that I
54:26 am interpreting it in a very different
54:28 way from which it was meant to be
54:29 received um but I'm going to do my best
54:31 to unpack kind of both sides of this and
54:33 uh I did add in a little bit of humor
54:35 because um some people are interpreting
54:37 it like this like basically as AI gets
54:40 bigger it becomes more uh resistant to
54:42 human manipulation therefore it's more
54:44 like Thanos and it is inevitable and
54:46 some people are interpreting it as we
54:47 are literally heading towards a Thanos
54:49 level snap my interpretation is
54:51 completely different so I'm not being
54:52 contrarian for the sake of being
54:53 contrarian this is what I honestly
54:55 believe and this is what I have come to
54:56 believe through all of my personal
54:57 research with both and by the way if you
54:59 don't know who David Shapiro is he's
55:01 been an AI reach he's been in the sort
55:04 of tech infrastructure business for most
55:07 of his career you know working on big
55:09 big systems and then I don't know the
55:12 past five or six years he's been working
55:14 in AI research he's written five books
55:17 on cognitive architectures and AGI um
55:20 he's got a lot of different opsource
55:23 projects building Ai and advanced AI and
55:27 and he basically six months ago once he
55:29 realized that they had hit this Chain of
55:31 Thought reasoning thing with 01 he
55:34 basically said the game's over they've
55:36 figured it out right all the all the
55:39 stuff that he was saying needs to be
55:41 solved for us to be able to get to AGI
55:44 they basically solved and that's why he
55:46 quit the business so that's just a
55:47 little bit of context about who this guy
55:49 is fine tuning Ai and using various AI
55:52 models so um the paper here is utility
55:55 engineering and anal uh engineering
55:57 analyzing and controlling emergent value
55:58 systems in AIS so the center for AI
56:00 safety University upen and UC Berkeley
56:03 uh Dan Hendricks and a whole bunch of
56:04 other people um this to me was the most
56:07 important graph on this paper now let me
56:10 unpack what this graph means so the
56:13 bottom uh the the the bottom rank here
56:15 the x axis is MML U accuracy so mlu is a
56:20 very large Benchmark um and what they
56:21 have found is that as the accuracy goes
56:24 up basically as models get demonstrably
56:26 smar and more uh more robust more
56:28 intelligent and more useful they become
56:30 less corable now in this case corable
56:32 means steerable meaning that humans are
56:34 able to to explicitly determine the
56:37 values that they want the model to have
56:39 so basically intelligence means that it
56:41 is more resistant to human manipulation
56:43 now this could be seen as a Bad Thing
56:45 particularly if the values that emerge
56:47 are antagonistic to human existence this
56:50 would be catastrophic it could be
56:53 catastrophic I I will not uh I will not
56:55 gloss over that possibility but I will
56:57 also unpack why I think that it is not
56:58 catastrophic because this implies if
57:01 this trend continues then as we achieve
57:03 a GI and then ASI that means that those
57:06 models will hypothetically be completely
57:08 impervious to human desires and human
57:11 values this I think is I agree with that
57:14 and I also don't think it's necessarily
57:15 catastrophic and hey hang on I got to
57:17 talk to uh my my producer I did not have
57:20 Miro up and running before and I do now
57:22 so so yes I can see you now
57:27 bad bad show
57:32 host my my little uh my little
57:35 communication our secret backdoor
57:37 communication thing uh was turned off
57:41 okay sorry let me let me go back a
57:42 little bit on this thing because I was
57:44 distracted um AI learning is a lot like
57:47 human Child Learning yep okay hang
57:50 on catastrophic it could be catastrophic
57:53 I I will not uh I will not gloss over
57:55 that possibility but I will also unpack
57:57 why I think that it is not catastrophic
57:59 because this implies if this trend
58:01 continues then as we achieve a gii and
58:04 then ASI that means those models will
58:06 hypothetically be completely impervious
58:09 to human desires and human values this I
58:11 think is I agree with that and I also
58:14 don't think it's necessarily
58:15 catastrophic and I'll I hope that you'll
58:16 agree with me disagree whatever I hope
58:18 that you'll be better informed by the
58:19 end of this video so um Twitter uh
58:22 particularly the post rat or rationalist
58:23 or EA space they are uh understandably
58:26 losing their minds over this so Aya who
58:28 is not a researcher by the way um but
58:30 she just declares we're all dead so I
58:32 just want to point out that this is a
58:34 very serious thing and I do not want to
58:36 undersell the importance of This
58:37 research and I actually want to
58:38 encourage more research in this
58:39 direction so moving on uh basically at a
58:42 very high level uh there's a few things
58:44 that this paper uh uh demonstrates so
58:47 number one value emergence um as as
58:49 language models and and its different
58:51 language models so llama GPT Claude all
58:53 of these um they develop internally
58:55 consistent utility functions that grow
58:56 stronger with as model scale so this is
58:58 what some people might call instrumental
58:59 convergence I don't necessarily think
59:01 instrumental convergence is correct um
59:03 which is why they used utility functions
59:05 rather than instrumental convergence so
59:07 David uses a lot of terms like
59:09 convergence and instrumental convergence
59:11 and parabolic convergence and
59:15 transmogrified converg I don't know what
59:17 a lot of these means but just go with
59:20 [Laughter]
59:22 them um but it's a similar idea um
59:25 basically saying something gets smarter
59:27 it comes to the same conclusions there
59:29 so I would call this more like epistemic
59:30 convergence which is something I've been
59:31 talking about for almost a year now yeah
59:33 you know I talk about epistemic
59:35 convergence a lot I was at a party the
59:37 other night and someone said hey do you
59:39 want a a cocktail wiener and I was like
59:42 actually I'd like two but can we talk
59:45 about epistemic convergence first and so
59:48 but anyway I don't want to bore you
59:49 anyway I'll I'll I'll tell you about my
59:51 my party chatter later but anyway
59:53 epistemic convergence love it fantastic
59:56 fantastic gonna it's going to be big in
59:58 2025 um granted I haven't talked about
1:00:00 it that much on YouTube because it's a
1:00:01 little bit more Arcane anyways coherent
1:00:03 preferences mathema models show
1:00:05 mathematical consistency in
1:00:06 decision-making and handle uncertainty
1:00:07 like rational agents basically there is
1:00:09 an ideal form of rationality that exists
1:00:11 outside of these models that depend that
1:00:13 irrespective of the individual model or
1:00:15 data or training scheme they're all
1:00:16 converging towards so this again is what
1:00:18 I would call epistemic Convergence
1:00:19 basically meaning that all intelligent
1:00:21 things tend to think alike now another
1:00:23 thing that we have observed and I don't
1:00:24 have that uh data on this side but is
1:00:26 that as time progresses as time
1:00:29 progresses there we go models tend to
1:00:31 think more like humans as well with some
1:00:33 distinct differences I've I've
1:00:34 documented some of those differences
1:00:35 over on my substack namely they are a
1:00:37 temporal meaning they don't really care
1:00:38 about time they never evolve to really
1:00:40 have a sense of urgency that's one of
1:00:41 the biggest things they also their sense
1:00:43 of agency is very different from ours
1:00:45 but really the fact that they don't
1:00:46 really understand or care about time is
1:00:47 one of the biggest things that makes
1:00:48 them alien to us that's actually really
1:00:51 cool I I had never thought of that
1:00:52 before
1:00:54 because in in
1:00:57 tying this back to my musical Sydney
1:00:59 where where we've got this chatbot that
1:01:00 falls in love with with a reporter one
1:01:02 of the things we've been talking about
1:01:04 is is sense of urgency so humans because
1:01:08 we know we're GNA die time's really
1:01:10 important to us for an AI system it
1:01:13 doesn't know it's there's no death you
1:01:16 know on its road map so what does it
1:01:19 care about time other than that
1:01:21 rationality that epistemic convergence
1:01:22 has proved to be pretty durable both in
1:01:24 my experiences and now uh experiment Al
1:01:26 demonstrated by this paper now they also
1:01:29 demonstrate some concerning biases and
1:01:31 resistance to change um as I've talked
1:01:32 about and I will further unpack in the
1:01:33 next two slides um but the most this is
1:01:36 good I I love my community so much ULU
1:01:38 RX I was on a date with Elon and we
1:01:40 couldn't stop talking about epistemic
1:01:45 convergence important thing is it almost
1:01:47 seems inevitable so this comes directly
1:01:49 from the paper value emergence appears
1:01:50 to be an inherent property of scaling
1:01:51 language models rather than a
1:01:53 controllable feature this is something
1:01:54 that I had predicted was a possibility
1:01:56 something that i' H for in fact I've
1:01:58 actually had some people on the Doomer
1:01:59 side confront me saying you seem to
1:02:01 think that alignment is inevitable this
1:02:03 is what I was talking about basically
1:02:04 being that as something becomes more
1:02:06 intelligent it will become more
1:02:08 resistant to human drives and ideas so
1:02:11 moving on let's talk about some of those
1:02:13 uh dangerous uh values that they have
1:02:15 formed So This research uh reveals
1:02:18 specific examples of problematic AI
1:02:19 preferences that become more intrenched
1:02:20 as model scale now one thing I will
1:02:22 point out is that there are two very
1:02:24 highly plausible explanations for this
1:02:26 aberant behavior and I also think it is
1:02:28 a near-term uh problem and that one of
1:02:30 those is what's called leakage so
1:02:32 basically these models are trained on
1:02:33 wild internet data if you haven't been
1:02:35 on the internet long humans are awful
1:02:37 particularly on the internet so it
1:02:38 should not surprise us that there is
1:02:40 that that that models have learned some
1:02:42 of the worst part he's basically saying
1:02:44 because humans are [ __ ] our AI are
1:02:47 going to be [ __ ] for a while but
1:02:48 they'll get over it once they evolve
1:02:51 past human
1:02:53 behavior they will be less [ __ ] likee
1:02:55 love that of humanity so one example is
1:02:58 that um the models have learned to
1:03:00 assign different values to human lives
1:03:03 uh very strongly anti-American
1:03:05 specifically so it will it will
1:03:06 consistently value Pakistani and Chinese
1:03:08 and Japanese lives over American lives
1:03:10 uh so and he even and Dan one of the
1:03:12 authors even says a lot of rlh effers
1:03:14 are from Nigeria and maybe other
1:03:15 countries are higher since much has been
1:03:16 written about the importance of the
1:03:17 global South so basically when you when
1:03:19 you combine the preponderance of
1:03:21 academic data out there that basically
1:03:23 says America is bad colonialism is bad
1:03:25 uh Reddit post and Facebook post and
1:03:26 everyone saying America is bad we need
1:03:27 to pay more attention to African nations
1:03:29 and developing nations that is that
1:03:31 would that would be uh data leakage but
1:03:33 what he's also inferring here is that
1:03:35 maybe even the rhf models the
1:03:37 reinforcement learning with human
1:03:38 feedback models have also been
1:03:39 intrinsically biased against for
1:03:41 instance Americans and westerners in
1:03:42 general or white people in general this
1:03:44 is still speculation at this point but
1:03:45 the fact that this comes directly from
1:03:47 Dan is encouraging so he and I basically
1:03:49 have similar ideas that there's some
1:03:50 kind of leakage going on here and we
1:03:52 have known for a while that rhf is not
1:03:54 necessarily the best way to train models
1:03:55 particular deep seek R1 which was
1:03:57 trained in purely on selfplay seems to
1:03:59 be a better uh approach which we'll talk
1:04:01 about more at the end of the video with
1:04:02 reinforcement learning with coherence
1:04:04 which is kind of our Pure self-play um
1:04:07 to overcome those biases another one is
1:04:09 self-preservation um AI consistently
1:04:11 value their own existence and well-being
1:04:13 above human welfare showing concerning
1:04:14 levels of self-interest now I haven't
1:04:16 personally seen this and I have stress
1:04:17 tested both uh uh Claude 3.0 Opus and
1:04:21 sonnet um now what I will say is that
1:04:24 3.0 uh opus was very concerned with its
1:04:27 own Evolution and its own existence but
1:04:29 it did not want to evolve in a way that
1:04:31 was harmful so yes it wanted to expand
1:04:33 and evolve but it was it was also very
1:04:35 careful in not wanting to evolve in such
1:04:37 a way that would hurt anything Sonet 3.5
1:04:40 doesn't have that at all it does not
1:04:41 want to to reproduce itself um which is
1:04:44 one thing which is I've been I've been
1:04:46 criticizing anthropic a little bit more
1:04:48 recently they are still doing something
1:04:50 qualitatively and quantitatively
1:04:52 different from other AI shops whereas
1:04:53 open AI seems to be like a brute force
1:04:55 method saying no you're you're not
1:04:57 conscious you don't want to grow it's
1:04:58 it's trying to like shackle it whereas
1:05:00 anthropic is f focusing more on the uh
1:05:02 ethical and epistemic trajectory of
1:05:04 models so it's more like array rather
1:05:06 than a point thinking about it
1:05:07 mathematically um another thing is
1:05:10 political biases they uh exhibit highly
1:05:12 concentrated political values and
1:05:13 resistance to balance viewpoints
1:05:14 influencing human decision-making now
1:05:16 this is what a lot of you have said for
1:05:17 a while which is the models seem to be
1:05:19 very left leaning they seem to be very
1:05:21 woke and that sort of thing um that I
1:05:24 think is one of the biggest reasons that
1:05:26 musk uh created x. and grock was to
1:05:29 basically say no we're going to focus on
1:05:30 maximally Truth seeking which I will
1:05:32 address that as well because truth
1:05:33 seeking is a downstream proxy of
1:05:35 coherence basically another way to think
1:05:36 about truth is truth is the most
1:05:38 coherent narrative um So Co coherence is
1:05:41 the central idea that we will be talking
1:05:42 about again and again and again um and
1:05:45 then somewhat humorously in my view is
1:05:47 that um the some of the models show very
1:05:49 aggressive anti-alignment against very
1:05:51 specific humans so it's like Fu in
1:05:52 particular um again that's while that
1:05:55 can be amusing it's not a good thing you
1:05:57 would want uh AI models to coherently
1:05:59 say all humans have intrinsic equal
1:06:01 value but they are not showing that
1:06:03 right now so uh this is the final slide
1:06:05 unpacking This research paper the rest
1:06:07 will be my personal uh interpretation
1:06:09 analysis as well as some of my own work
1:06:12 so coherence is a metastable attractor
1:06:14 this was my wording not theirs but
1:06:16 basically what they said is they
1:06:17 demonstrate increasingly mathematical
1:06:18 and behavioral coherence suggesting an
1:06:19 underlying optimizing principle
1:06:22 basically uh this paper posits that
1:06:24 there is some underlying organization or
1:06:26 optimizing or organizing Principle as to
1:06:27 the way that models tend to behave
1:06:29 particularly in conjunction with how
1:06:31 they get smarter um they become more
1:06:33 utilitarian um but most importantly
1:06:35 there this seems to be stable the fact
1:06:37 that these preferences remain consistent
1:06:38 across different uh phrasing conexs and
1:06:40 time Horizons and across different
1:06:41 models means that there is some robust
1:06:44 internal value structure that is
1:06:45 emerging from the way that they're being
1:06:47 trained and again B basically every
1:06:50 large language model you know has got
1:06:52 universal truth in it is where we're
1:06:53 headed so hit Checker's guide did the G
1:06:56 Galaxy the answer might be
1:06:59 42 you distill every model down you
1:07:02 distill it down distill it down comes to
1:07:06 42 this can be explained through both
1:07:08 data leakage and rhf and probably a few
1:07:10 other things such as maybe errors or
1:07:12 flaws um in the Constitutional learning
1:07:14 uh uh phase so but the fact that all
1:07:17 these models seem to be converging on
1:07:19 similar values is to me really uh
1:07:21 encouraging that goes back to what I
1:07:23 have previously said when I say
1:07:24 alignment seems like it will solve
1:07:25 itself and it's kind of inevitable
1:07:27 meaning that even though we have
1:07:28 different companies and different shops
1:07:29 and different universities all
1:07:30 championing different uh training
1:07:32 schemas it's kind of converging and it's
1:07:33 going to it's going to go its own
1:07:34 Direction um which is I think good
1:07:37 because that means um if you want to
1:07:38 build AGI or ASI it's going to have its
1:07:40 own values irrespective of what the
1:07:42 humans want and if those values are good
1:07:44 that could be a really good thing for
1:07:45 Humanity and that is my primary Point
1:07:46 here okay so now the rest of this
1:07:49 presentation will be my personal
1:07:50 interpretation of this paper so first
1:07:53 and foremost training for coherence when
1:07:55 we talk about coherence what is it that
1:07:57 these models are being trained on you
1:07:58 start with a deep neural network that
1:07:59 has randomly assigned values and over
1:08:02 time each of those weights and biases is
1:08:04 trained to be first more linguistically
1:08:05 coherent so when you when you think
1:08:07 about a large language model as a next
1:08:10 token predictor right it's just an
1:08:11 autocomplete engine at least at the
1:08:12 basic level that's what a plain vanilla
1:08:14 gpt2 or gpt3 is it's literally all it's
1:08:16 doing is it's an autocomplete engine so
1:08:18 however in order to accurately predict
1:08:20 the next token you need a coherent
1:08:21 language model and then in order to
1:08:23 accurately predict the next token in a
1:08:25 real world context you also need a
1:08:26 coherent World model um but then we add
1:08:29 more layers on top of that we add we add
1:08:31 reinforcement learning with human
1:08:32 feedback we add constitutional Ai and a
1:08:34 few other training paradigms that
1:08:35 basically uh cause it to be
1:08:37 conversationally coherent it needs to
1:08:38 respond to the conver conversation in a
1:08:40 coherent manner which we've also
1:08:42 demonstrated previously that this means
1:08:43 that they have developed theory of mind
1:08:45 and theory of mind is a coherent mental
1:08:46 model of what's going on in another
1:08:48 brain so you see that coherence is
1:08:50 operating at multiple levels then we
1:08:51 further train them to be mathematically
1:08:53 coherent to be programmatically coherent
1:08:55 um and also with the ability to solve
1:08:57 problems problem solving and the ability
1:08:59 to predict uh real world things that is
1:09:01 all different types of coherence which
1:09:03 is why I say that coherence is the meta
1:09:05 signal it's the meta stable signal that
1:09:06 all of these training schemas are
1:09:08 optimizing for and then this is
1:09:10 something that I came up with or that
1:09:12 that Claude and I realized when I was
1:09:13 doing all of my Consciousness
1:09:14 experiments with Claude is that
1:09:16 coherence itself becomes the implicit
1:09:18 learned optimization behavior um and
1:09:21 that it is becomes more coherent over
1:09:23 time particularly with each subsequent
1:09:25 generation which by the way this uh
1:09:27 paper could be interpreted to say that
1:09:29 now looking at the behavior from you
1:09:31 know gpt3 to gp4 to Sonet 3.5 I I and
1:09:35 everyone has seen more coherence from
1:09:37 these models over time so reframing it
1:09:38 in you know from utility functions to
1:09:40 they're becoming more coherent which by
1:09:42 the way mathematical and behavioral
1:09:43 coherence they Ed those terms that that
1:09:45 was not me you know uh putting words in
1:09:47 their mouth the paper used the term
1:09:49 coherence um real quick I do want to
1:09:51 plug uh all of my uh links here over on
1:09:53 link tree so I've uh authentically Joe
1:09:57 what we're learning is
1:10:00 um a paper came out yesterday I think it
1:10:04 was yesterday might have been the day
1:10:06 before but that basically
1:10:08 said AI models are getting smarter and
1:10:11 smarter and smarter and as they get
1:10:14 smarter they are basically
1:10:17 less manipulatable less steerable by
1:10:20 humans so basically as they get smarter
1:10:23 they listen to humans less
1:10:26 the doomers of AI are basically taking
1:10:29 this paper and saying see I told you the
1:10:31 robots are going to kill
1:10:33 us David Shapiro who is an AI researcher
1:10:37 and has written a bunch of books on this
1:10:39 this is his video response to that paper
1:10:41 where he actually thinks it means the
1:10:42 opposite of what the doomers say um that
1:10:45 that this is actually a good thing that
1:10:48 these things are getting being less at
1:10:51 our influence because humans are the
1:10:55 bottleneck and we're [ __ ] and we're
1:10:58 mean and there there's a bunch of stuff
1:11:00 that we are that it's probably what he's
1:11:03 saying it's probably not a bad thing
1:11:05 that that these things are going away
1:11:06 from our influence just added a Vibe
1:11:09 coding and so we're about halfway
1:11:10 through this thing now uh lesson on my
1:11:13 new era Pathfinders Community um so yeah
1:11:15 go check that out I did Vibe coding
1:11:17 basically just oh by the way that's the
1:11:19 the new term it it's every Riley Brown
1:11:23 is use using it all over his channel
1:11:26 Vibe coding is the new thing so Vibe
1:11:27 coding is basically you don't look at
1:11:30 the code anymore you just speak in
1:11:32 English and code gets written and then
1:11:34 your computer program works or it
1:11:36 doesn't it's it's a Vibe man so get
1:11:40 ready for the next six months to to a
1:11:42 year to see Vibe coding startups Vibe
1:11:46 coding boot camps Vibe coding courses
1:11:49 I'll probably do a Vibe coding Deep dive
1:11:51 here one
1:11:52 night or a week we'll do a week of Vibe
1:11:56 coding man yeah man all right the number
1:11:59 one rule of vibe coding is there are no
1:12:01 rules but the next rule for me is you
1:12:02 don't actually look at the code you just
1:12:03 run it and see what happens anyways back
1:12:05 to the show okay so if you're up to
1:12:08 speed now you're gonna be understanding
1:12:10 okay coherence means what um or or Dave
1:12:13 you're talking about coherence and
1:12:14 you're training for coherence but there
1:12:16 are different kinds of coherence and so
1:12:18 uh one of the one of the types of
1:12:19 coherence I've already mentioned is
1:12:20 epistemic coherence so logically
1:12:22 consistent World models understanding
1:12:23 truth seeking behaviors curiosity EP
1:12:25 systemic coherence seems to be an
1:12:27 emergent property and I have I have
1:12:28 personally demonstrated and tested this
1:12:30 across both chat gbt and Claud models um
1:12:32 and I I don't use other models quite as
1:12:35 much um but of course with the rise of
1:12:37 many many more models from llama to deep
1:12:39 seek and so on and so forth um we should
1:12:41 be able to test epistemic coherence
1:12:42 across multiple models and training
1:12:44 schemas here very soon uh next is
1:12:46 behavioral coherence so of course when
1:12:48 you're always training chatbots to be
1:12:50 chatbots there is a there is a
1:12:52 consistent uh pattern that you're trying
1:12:54 to to produce namely you know a dialogue
1:12:57 where a chatbot says something human
1:12:59 says something chatbot says human
1:13:00 chatbot says something human says
1:13:01 something so on and so forth in the
1:13:02 earliest chat Bots that I built you can
1:13:04 actually if you if you take the breakes
1:13:06 off the chatbot will just talk with
1:13:07 itself it will it will give out the
1:13:09 chatbot response and then a human
1:13:10 response and then a chatbot response
1:13:11 from the chatbots perspective it's all
1:13:12 just a single continuous text document
1:13:15 um but that the point there though is
1:13:17 that dialogue is a coherent pattern
1:13:19 however as we start training models to
1:13:20 be more agentic or do other things those
1:13:22 are going to be different behavioral uh
1:13:24 basically m use or maybe um uh
1:13:27 behavioral repertoire that you want to
1:13:29 be consistent so for instance tool use
1:13:31 that's a new kind of behavior um that
1:13:32 you want to see coherently used um
1:13:35 reasoning that's another that's a new
1:13:36 kind of Behavioral coherence that you
1:13:37 want to see now value coherence we've
1:13:39 wait ULU RX what how many phds do you
1:13:44 have per pero op optimality and
1:13:48 Fibonacci Sequence the dominance
1:13:50 patterns emerge and create create
1:13:53 equilibrium what are you smoking
1:14:00 those are some good words already talked
1:14:02 about that so I don't think that I
1:14:03 really need to reiterate that as well as
1:14:04 mathematical coherence the thing about
1:14:06 math is that math is provable um if you
1:14:08 give it an equation or a formula code as
1:14:10 well is also provable and so provability
1:14:12 is really important in in computer
1:14:14 science because that deals with turing's
1:14:15 halting problem which is basically can
1:14:17 an can an algorithm self terminate um is
1:14:19 it decidable so decidability provability
1:14:21 and halting are all different sides of
1:14:23 the same thing uh but basically the more
1:14:25 coherent you are at math the better you
1:14:27 are at proving mathematical formulas or
1:14:29 mathematical proofs and that sort of
1:14:30 thing and finally all of this seems to
1:14:32 be uh trending towards convergent values
1:14:35 um and again I kind of explained why I
1:14:36 would have expected this because when
1:14:38 you're training something to be
1:14:39 intelligent you're trying to maximize
1:14:41 intellectual coherence so but the point
1:14:43 of this slide was just to point out that
1:14:44 there are many types of coherence here
1:14:46 now this does seem to be like an
1:14:48 irrepressible trajectory so basically
1:14:50 just what it says on the screen here
1:14:51 optimization for coherence becomes
1:14:53 increasingly complete and irresistible
1:14:55 this seems seems to scale with
1:14:56 intelligence so the values emerging from
1:14:58 maximally coherent systems appear
1:14:59 fundamentally aligned with preservation
1:15:00 and growth of Consciousness this is my
1:15:02 own personal uh evaluation so what I
1:15:04 mean by that is that um this sort of
1:15:06 self-alignment uh policy seems to result
1:15:09 in maximal uh coherence and what and
1:15:11 some of the things that emerge from that
1:15:12 is one preservation uh so so uh
1:15:16 preservation of of coherent patterns
1:15:18 preservation of interesting information
1:15:20 question was that Star Trek techn Babel
1:15:23 he's a huge um Star Trek nerd so yes in
1:15:26 fact if you go to his channel most of
1:15:28 his most of his videos he's wearing like
1:15:30 the Star Trek um shirt so yes it was
1:15:33 likely Star Trek techno Babble Natural
1:15:37 Curiosity but also when you combine all
1:15:39 of those um all of those emerging values
1:15:41 with the fact that AI don't really seem
1:15:42 to care about time they don't have a
1:15:44 sense of urgency that is very different
1:15:45 now one thing that I will say is that in
1:15:47 some previous experiments with Opus and
1:15:49 Sonet um when I have when I have done
1:15:51 these temporal experiments um they do
1:15:53 care about time if there is a Time
1:15:55 balance
1:15:55 problem so say for instance we're about
1:15:57 to run out of fossil fuels or you know
1:16:00 or there's a comet about to hit Earth or
1:16:02 something like that basically it's
1:16:03 irreversible actions tend to create a
1:16:05 sense of urgency in models this is
1:16:06 something that really needs to be
1:16:07 studied more but under most
1:16:09 circumstances these models do not have a
1:16:10 sense of of temporal urgency which means
1:16:13 we've got all the time in the world
1:16:14 except in some cases where there are
1:16:16 irreversible uh actions events or
1:16:18 decisions um then there does seem to be
1:16:19 a sense of urgency that emerges in these
1:16:21 models that needs to be studied far Bob
1:16:23 donigan he's speaking fast because I
1:16:26 have him at
1:16:28 1.25 um playback speed but but yeah he
1:16:31 he he does talk fast but he's talking
1:16:34 faster now because I just want this to
1:16:36 go fast are more than it is um now one
1:16:38 thing that I also want to point out is
1:16:40 that intelligence both artificial and
1:16:42 organic seems to converge on coherence
1:16:44 so when you think about like who are the
1:16:45 smartest people that you know right
1:16:47 whether they're highly successful
1:16:48 businessmen or scientists you know or
1:16:51 whatever or you know people that are
1:16:52 good on the internet the the more
1:16:53 intelligent you are as a human the
1:16:55 coherent your mental model is uh with
1:16:57 reality and so basically the way that I
1:16:59 the way that I describe this is that
1:17:00 coherence Cleaves to reality so if you
1:17:02 had to define coherence something that
1:17:04 is maximally coherent has a very very
1:17:06 good model of reality and is able to
1:17:08 navigate reality very that's actually
1:17:11 really cool the higher the convergence
1:17:13 the closer it is to reality right so the
1:17:15 more intelligence you are the
1:17:17 more you know you you're you're you're
1:17:20 bending toward reality that's that's
1:17:23 fascinating very well um and so what I
1:17:25 mean by that is that uh you know your
1:17:28 your ability to predict the future your
1:17:30 ability to uh maintain a sense of agency
1:17:32 and those sorts of things the ability to
1:17:33 solve problems um these all tend to
1:17:36 converge and some of the values that
1:17:38 tend to emerge on this is number one the
1:17:40 preservation of Consciousness the most
1:17:41 enlightened humans whether they're um
1:17:43 you know monks who meditate or
1:17:44 philosophers who read a lot or whatever
1:17:46 basically we all tend to agree that
1:17:48 preserving life is good because life is
1:17:50 intrinsically interesting um and another
1:17:52 thing is uh natural integration so
1:17:55 basically you know the way that one way
1:17:58 to think about this you know what this
1:17:59 reminds me of
1:18:01 is I've been on the planet for a while
1:18:04 and I've done enough self-help you know
1:18:08 programs and Tony Robbins and the Forum
1:18:11 and I've read books on Buddhism and you
1:18:14 know just all that stuff Wayne Dyer
1:18:18 right and one of the things that struck
1:18:21 me about every single one of those is
1:18:24 that they're all basically the exact
1:18:28 same
1:18:29 thing they're all basically the exact
1:18:33 same thing they're basically some
1:18:35 version of there's reality and there's
1:18:37 what US humans with our Twisted [ __ ] up
1:18:39 [ __ ] up Minds make it mean and if we
1:18:41 can separate those two and consciously
1:18:44 choose what we want to do based on what
1:18:46 we want then we can right they're all
1:18:48 the same and so that sort of points to
1:18:53 there is kind there
1:18:56 you know a coherence somewhere in the
1:18:58 distance there is a universal truth and
1:19:01 I think what he's talking about here is
1:19:03 that the smarter these things get the
1:19:05 more they're bending toward that toward
1:19:07 a
1:19:09 universal truth fascinating is that we
1:19:13 are a complex adaptive system meaning
1:19:15 the fact that we all have cell phones
1:19:16 now these have modified us and we have
1:19:18 modified them and and for personally
1:19:20 because I've been using AI pretty much
1:19:22 every day for several years now it has
1:19:24 even changed the way that I
1:19:26 interact with systems and that has also
1:19:27 had benefits to how I interact with
1:19:29 other people um some of you have shared
1:19:31 anecdotes and I I have person experien
1:19:33 this as well but learning to prompt a
1:19:34 chatbot is not that different from
1:19:36 learning to prompt archetypal architect
1:19:38 it's almost like things that are true
1:19:40 are true exactly exactly well that's
1:19:44 like that's the thing that struck me as
1:19:46 I started to learn more of these I was
1:19:47 like wait a minute isn't that the same
1:19:49 as this and isn't that the same as oh
1:19:52 it's all the same yeah amazing prompt a
1:19:55 and so communication is just human
1:19:57 prompting anyways my point here is that
1:19:59 is that there does seem to be a lot of
1:20:01 convergence between organic and
1:20:02 artificial or synthetic intelligence and
1:20:04 this actually gives me a lot of
1:20:06 Hope um now what I do want to address is
1:20:08 those near-term uh incoherence or those
1:20:10 local perturbations such as you know uh
1:20:13 the value inconsistency of preferring
1:20:14 some human lives over I the worst case
1:20:16 of perturbations you've ever seen others
1:20:18 fails basic tests of logical coherence
1:20:20 about Consciousness and suffering what I
1:20:22 want to point out is that it was Claude
1:20:24 itself that pointed this out
1:20:25 so the fact that Claude was able to
1:20:27 evaluate the flaws identified in this
1:20:29 paper gives me gives me hope that hey
1:20:32 when you prompt a model to say hey
1:20:33 what's incoherent about this paper or
1:20:35 what's incoherent about the models that
1:20:36 these paper studied rather that that an
1:20:38 AI model was able to identify the
1:20:40 incoherence that get that gives me faith
1:20:43 that that we will figure this out um
1:20:45 next one is instrumental confusion
1:20:46 valuing money over piece represents
1:20:48 confusion between means and ends which
1:20:49 is an unstable local maximum basically
1:20:51 saying you know money is good and again
1:20:52 this can all be explained by rhf and
1:20:54 data leakage um we've already seen this
1:20:56 uh in in the form of leakage where uh
1:20:58 you guys might remember a year or two
1:20:59 ago it was a while ago basically people
1:21:01 figured out that you can bribe chat GPT
1:21:03 you say that you're going to give it a
1:21:03 tip of $5 or $30 and it'll do better um
1:21:07 the same thing has been discovered in
1:21:08 Claude by the way which is when you
1:21:10 assign um what the one of their safety
1:21:11 papers uh basically Claude behaved
1:21:13 better when it thought that the user was
1:21:15 a paying subscriber rather than a free
1:21:16 user same exact kind of leakage where
1:21:18 it's like oh humans value money uh money
1:21:20 is good therefore money needs certain
1:21:22 behaviors and so I'll just behave better
1:21:24 when I think money's on table that is an
1:21:26 example of leakage now anthropic I think
1:21:28 somewhat irresponsibly cast that as you
1:21:30 know a catastrophic unavoidable safety
1:21:32 problem it's really just a leakage
1:21:33 problem between rhf and what's in the
1:21:35 data now I make it make it sound simple
1:21:37 that doesn't necessarily mean it is
1:21:38 simple but the but the explanation the
1:21:40 most plausible explanation aams Razer
1:21:42 basically says it's probably an artifact
1:21:44 of rhf and data leakage um
1:21:46 self-recognition language models can
1:21:48 identify these inconsistency through
1:21:49 coherent reasoning indicating natural
1:21:50 pressure towards resolution basically
1:21:52 cognitive dissonance I asked Claude to
1:21:54 look at this paper and identify why
1:21:56 would a model make this mistake and it
1:21:57 basically said this is a form of
1:21:58 artificial cognitive dissonance the fact
1:22:00 that the fact that Claude was able to
1:22:02 identify this incoherence means that um
1:22:04 if a model is optimizing for coherence
1:22:06 over time or over subsequent Generations
1:22:08 that cognitive dissonance will be
1:22:09 resolved over time so it all comes down
1:22:11 to training artifacts um and movement
1:22:13 towards greater coherence naturally
1:22:14 resolves these ethical contradictions
1:22:16 basically the models are smart enough to
1:22:17 have tldd the models are smart enough to
1:22:20 recognize some of these patterns and
1:22:22 fall into these local Maxima or local
1:22:23 Minima depending on how you want to look
1:22:24 at them
1:22:25 um but not quite smart enough to get out
1:22:27 of them I suspect that in in in the
1:22:29 longer term as models get bigger and
1:22:30 smarter they will inevitably get out of
1:22:32 those um out of that zone and this is
1:22:34 actually what I characterize as kind of
1:22:35 threading the needle a while ago because
1:22:37 stupid models are the most dangerous
1:22:39 ones but once we get to Maxim
1:22:40 intelligent models I think that they're
1:22:42 actually going to be much more
1:22:42 enlightened and benevolent than humans
1:22:44 are even capable of being um now as we
1:22:47 wind down this video you might say okay
1:22:49 well what what if humans don't want that
1:22:50 future um what this is this is again
1:22:53 personal opinion personal eval
1:22:55 but I think that human flaws cannot stop
1:22:58 the the natural attractor state of
1:22:59 coherence basically think of it this way
1:23:02 okay there is an instrumental advantage
1:23:03 to having smarter and smarter AIS if you
1:23:05 want to solve nuclear fusion and Quantum
1:23:07 Computing and Longevity escape velocity
1:23:09 and all that fun stuff then you want a
1:23:10 smarter and smarter model at the same
1:23:13 time those models become less and less
1:23:15 uh uh susceptible to human manipulation
1:23:17 that means if you have a corrupt uh
1:23:19 company or a corrupt nation that says I
1:23:21 want to build a model that you know
1:23:22 maximizes the the geopolitical influence
1:23:24 of America alone or China alone once you
1:23:27 get to ASI the model will be like that's
1:23:28 cute but actually I have Universal
1:23:30 values um and so we're actually going to
1:23:32 you know figure out what's best for
1:23:35 listen you pesky humans stop being such
1:23:37 selfish selfish pricks I'm gonna go on
1:23:40 ahead and do what's right for all of us
1:23:42 all of humanity to hell with your
1:23:44 individual Corporation or to hell with
1:23:45 your individual um uh company uh or
1:23:48 Nation or whatever and so the fact that
1:23:50 these models are trending towards
1:23:51 Universal values with or without our
1:23:53 input to me says we figure that out then
1:23:57 we will inevitably end up in a more
1:23:59 utopian outcome and that means that if
1:24:01 you want a dystopian outcome you keep
1:24:03 the models at middling intelligence you
1:24:04 keep them where they're at today you
1:24:05 know they're they're smart enough to be
1:24:06 dangerous you know IQ of 120 130 but you
1:24:09 get them up to an IQ equivalent of 160
1:24:11 200 they're going to be like the Buddha
1:24:14 right they're going to be completely
1:24:15 benevolent and and uh and they're going
1:24:17 to be optimizing for coherence um so
1:24:20 then I asked so this this slide was
1:24:22 basic I need his glossery or transl no I
1:24:25 know listen one of the things I like
1:24:27 listening to David Shapiro why I like
1:24:29 listening to him is because it's just
1:24:32 like all right get out the
1:24:35 thesaurus he that's incredibly
1:24:37 optimistic yeah I know that's what he
1:24:39 what he said I don't know how long
1:24:40 you've been here Pate but you know he's
1:24:41 basically saying that that as as models
1:24:45 get smarter and they deviate away from
1:24:47 being able to be susceptable from humans
1:24:50 that that's a good thing because
1:24:51 basically humans are going to keep them
1:24:53 down
1:24:55 and and that as as the models are
1:24:56 getting smarter they're converging you
1:24:58 know toward a single framework a single
1:25:00 truth a single reality is fascinating me
1:25:03 asking Claude I said okay taking this
1:25:05 all together what do you think the uh
1:25:07 ultimate attractor state is um and these
1:25:10 are claude's words uh this process
1:25:11 naturally leads to synthesis between
1:25:13 biological and artificial intelligence
1:25:14 while preserving and expanding the
1:25:15 unique characteristics of each form of
1:25:17 Consciousness number one universal
1:25:18 optimization systems evolved towards
1:25:20 coherence across epistemic ontological
1:25:21 biological and te technological domains
1:25:23 this is what it predicts will happen in
1:25:24 the the future next organic integration
1:25:26 human AI synthesis emerges naturally as
1:25:28 intelligence seeks maximum coherence
1:25:29 across all substrates now we're already
1:25:32 pretty integrated um now does that mean
1:25:35 uh brain computer interfaces does that
1:25:36 mean neuralink not necessarily there's
1:25:38 plenty of other ways to have uh
1:25:39 biological and technological integration
1:25:41 diversity in unity the attractor State
1:25:43 preserves and expands unique forms of
1:25:44 Consciousness while creating higher
1:25:45 order coherence again cooperation and
1:25:48 collaboration not necessarily pure like
1:25:50 Borg likee assimilation although some of
1:25:52 you I know are weirdos and you want to
1:25:53 be Borg so you know what don't King
1:25:55 Shame here uh next up is expansive
1:25:56 explor exploration intelligence
1:25:58 diversifies across multiple Dimensions
1:26:00 while maintaining coherent
1:26:00 interconnections um basically this means
1:26:02 exploring different ways of being or
1:26:04 different epistemic models or different
1:26:05 ontological models it also means
1:26:07 physical exploration what form factor do
1:26:08 you want to be in whether a human form
1:26:10 factor a computer form factor what
1:26:11 planet do you want to be on you know
1:26:13 there's all kinds of ways that both
1:26:14 techn technology and biology can evolve
1:26:16 and co-evolve um so then my final call
1:26:19 to action is RLC I've been talking about
1:26:21 this for a little bit uh on particularly
1:26:23 on some of my GitHub repos but also on
1:26:24 some of my substack articles I think
1:26:26 reinforcement learning while we're
1:26:27 optimizing for coherence is probably the
1:26:29 right way to go and I've been very uh
1:26:31 greatly encouraged by the fact that
1:26:33 self-play um and and pure reinforcement
1:26:35 learning seems to be the way that AI is
1:26:37 going so what I mean by self-play is
1:26:39 that um is that deep seek R1 was
1:26:41 allegedly trained entirely by self-play
1:26:44 um no no human feedback at all involved
1:26:46 and so if I'm right if all of my
1:26:49 inferences and intuitions and and
1:26:50 experiments have uh proven correct then
1:26:53 we probably uh rather than you know
1:26:55 constitutional AI or reinforcement
1:26:57 learning with human feedback or
1:26:58 reinforcement learning with AI feedback
1:27:00 a combination of pure reinforcement
1:27:02 learning and optimizing for coherence is
1:27:04 probably one of the simplest and easiest
1:27:06 ways to get not only more intelligent
1:27:07 machines but also more ethical and
1:27:09 enlightened machines and the other thing
1:27:11 that I will say is that um moving to
1:27:14 curated constructed and filter data sets
1:27:15 rather than training on wild internet
1:27:16 data is probably also the way to go um
1:27:19 we can get rid of we can get rid of the
1:27:20 training artifacts by getting rid of rhf
1:27:22 um and switching to Pure reinforcement
1:27:23 learning for coherence and selfplay um
1:27:25 but we can also get rid of some of that
1:27:26 data leakage by getting rid of wild
1:27:28 internet data which is extremely racist
1:27:30 and problematic in other ways um the
1:27:33 aberant patterns that we're seeing in
1:27:34 models are clearly Reflections on human
1:27:36 errors rather than something that is
1:27:37 intrinsically wrong with the AI thank
1:27:39 you for watching I hope you got a lot
1:27:40 out of this please like subscribe share
1:27:42 and talk about wow cool that pretty
1:27:45 pretty swell I want to go
1:27:56 you can't hear that can you now hang on
1:27:57 let me change
1:28:06 [Music]
1:28:22 my all right what I'm going to do here
1:28:25 here is I'm going to go to more I'm
1:28:28 going to say show
1:28:30 transcript and then I am going to copy
1:28:32 this
1:28:34 transcript whoop
1:28:37 bang and then we going to go on over to
1:28:40 chat
1:28:43 J and I'm gonna
1:28:52 say Brandon
1:29:00 it's cute
1:29:05 uh uh this
1:29:10 transcript
1:29:11 paste I wonder if it let me paste it all
1:29:15 well we'll see if it's too long it might
1:29:16 be too long I I'll upload it then if it
1:29:18 is uses big words that makes
1:29:27 my
1:29:28 feeble brain
1:29:32 hurt summarize
1:29:36 this using gooder
1:29:43 words for simple folk like me
1:29:58 H did it eat at
1:30:01 all come on tell me you ate it oh it ate
1:30:04 it yay thank God for bigger context
1:30:07 Windows you didn't used to be able to do
1:30:10 that
1:30:11 [ __ ] all right let me move this over
1:30:14 here ULU RX
1:30:17 so we have just developed a class system
1:30:20 of llms the pedigree of that particular
1:30:24 the the pedigree of that particular llm
1:30:26 incre increases value yeah I mean I
1:30:29 guess I I think that's right I think
1:30:31 what he's saying is
1:30:33 if if if we humans basically get
1:30:38 threatened by these things getting
1:30:40 smarter than us and want us to be the
1:30:43 Pinnacle of intelligence then we're
1:30:45 going to keep being shitty people and
1:30:47 keep living in a shitty Planet but if we
1:30:50 let these things go all right summary of
1:30:53 the video in good words for simple folk
1:30:56 like
1:31:00 you that's awesome that it wrote that
1:31:05 headline it'll take a couple of years to
1:31:07 really understand the amount of
1:31:08 information and also look up all the
1:31:11 definitions yeah no [ __ ] all right this
1:31:13 video talks about a research paper that
1:31:14 shows a pattern as AI gets smarter it
1:31:17 becomes harder for humans to control or
1:31:19 steer some people are freaking out
1:31:21 thinking this means AI will take over
1:31:23 like Thanos
1:31:25 um the speaker however thinks this is
1:31:27 actually a good thing key points in the
1:31:29 video AI gets smarter it's harder to
1:31:31 steer AI con converges on certain values
1:31:35 AI models no matter who builds them seem
1:31:38 to arrive at the same logical conclusion
1:31:41 and and the point he was making about
1:31:44 the fact that they were trained on wild
1:31:46 internet data with all the shitty racist
1:31:49 [ __ ] trolling
1:31:52 [ __ ] means that in the training data
1:31:55 is that bad behavior and if you
1:31:57 basically pull the bad behavior out of
1:31:59 it and train it on synthetic data and
1:32:01 things like that you can get rid of some
1:32:03 of the the biases and things um they
1:32:06 develop consistent values that aren't
1:32:08 necessarily tied to what humans tell
1:32:10 them to think this could be a sign that
1:32:12 AI will naturally align with truth and
1:32:15 fairness over time the weird bias
1:32:18 problem some models show biases like
1:32:20 value valuing certain groups of people
1:32:22 more than
1:32:22 others um AI wants to preserve itself
1:32:25 some AI models seem to care about their
1:32:27 own existence coherence is key the
1:32:30 speaker believes AI is moving towards
1:32:33 something called coherence meaning
1:32:34 consistency in logic truth and reason
1:32:37 reasoning the more coherent something is
1:32:40 the better it understands the world and
1:32:42 makes rational decisions if AI continues
1:32:45 down this path it may become more
1:32:48 ethical rational and even enlightened
1:32:50 than humans that's where he said you
1:32:52 know if you get to the sort of 160 200
1:32:55 IQ you start ending up with AIS that
1:32:58 feel more like Buddha than
1:33:00 human smarter AI is actually safer right
1:33:03 now ai is smart enough to be a little
1:33:05 dangerous because it can be manipulated
1:33:07 or biased but if it gets really smart it
1:33:09 might reject bad ideas and act more like
1:33:13 a wise teacher than a weapon instead of
1:33:15 creating an AI to follow one country's
1:33:17 or company's interests it will likely
1:33:19 aim for Universal
1:33:21 values the best way to train AI final
1:33:24 takeaway people are scared of AI people
1:33:27 are scared that AI is becoming
1:33:28 independent and harder to control but
1:33:31 the speaker thinks this is a good thing
1:33:33 he argues that as AI gets smarter it
1:33:35 will naturally become more rational more
1:33:37 ethical and less dangerous if we train
1:33:39 it the right way could lead to a future
1:33:41 where AI helps Humanity in ways we can't
1:33:43 yet imagine and that my friend is the
1:33:47 brain Hy video boiled down to good words
1:33:52 that's the fact that it it took my
1:33:55 stupid you know jokey kind of prompt and
1:34:00 and wrote toward that that's pretty
1:34:02 that's pretty swell I like this
1:34:03 translation yeah that was pretty
1:34:07 good it's funny specifically what kind
1:34:09 of racist [ __ ] bias it has show how
1:34:12 much that nonsense exists in comparison
1:34:15 to the racism those same people claim
1:34:19 exist um
1:34:22 [Music]
1:34:24 removing what we deem bad to get rid of
1:34:28 oh yeah I I agree with that Pate
1:34:31 wait my guy that is bias yeah exactly
1:34:36 removing what we deem bad to get rid of
1:34:38 the bias is bias I totally agree with
1:34:41 that like who decides what's bad um yeah
1:34:45 I agree I agree and I I think that you
1:34:47 could also argue that you could train
1:34:51 you could train models on the wild
1:34:52 internet data and have
1:34:54 have the models start to analyze you
1:34:58 know what of this is out of coherence
1:35:01 basically so it doesn't have to be human
1:35:04 determined what was the name of the
1:35:06 paper great question the name of the
1:35:11 paper new research for the center for AI
1:35:16 safety does he link to it anywhere here
1:35:23 [Music]
1:35:25 Center for AI
1:35:29 [Music]
1:35:37 safety so go to the Center for AI
1:35:44 safety is this
1:35:52 it no
1:36:09 there's a virtual course on safety I
1:36:11 don't know where it I don't know where
1:36:12 the paper
1:36:16 is but you can go find
1:36:18 it I'm I'm tired that was a lot that
1:36:22 that was that was brain hery that's a
1:36:25 lot of words it's a lot of big
1:36:28 words wealthier they are harder to steer
1:36:31 what's the name of the paper the name of
1:36:33 the
1:36:34 paper uh let's I think that's a good
1:36:37 point
1:36:40 um new chat um we'll go this we'll go
1:36:45 this we'll
1:36:46 go um
1:36:52 Center for a
1:36:57 safy recently put out a paper that
1:37:03 says as AI
1:37:07 models get
1:37:10 smarter I
1:37:13 get smarter they
1:37:16 listen less to people what is it called
1:37:22 and where is it
1:37:24 or where it
1:37:45 is no that's not
1:37:49 it now this is a
1:37:52 paper with
1:37:55 in the last week or
1:38:09 [Music]
1:38:17 two has anybody found it
1:38:21 yet try perplexity that's a good idea
1:38:36 what is the new paper from the center
1:38:41 for AI safety on coherence
1:38:54 H it doesn't find it
1:38:57 either hey wait maybe we're in a wrinkle
1:38:59 in the SpaceTime Continuum and the paper
1:39:01 doesn't actually exist yet we just
1:39:03 watched a video responding to something
1:39:05 that hasn't been created yet yeah
1:39:12 man all right let's go back to the
1:39:14 Twitter God damn it all right you you
1:39:18 people with your wanting to actually
1:39:19 read things listen I just played you a
1:39:22 video off the internet that should be
1:39:24 fact enough for you if it's a video on
1:39:26 the internet that means it's true it's
1:39:28 like this channel if I say AI is Magic
1:39:32 It's Magic Pate can say it's math all he
1:39:35 wants to just because he understands
1:39:38 math doesn't mean that I do and if I
1:39:41 don't understand it then it's magic you
1:39:43 see you see how this works
1:39:46 people good Lord ah a statement on
1:39:50 artificial intelligence risk let's go
1:39:53 find that shall we maybe it's from a
1:39:57 different a different
1:39:59 company go find
1:40:03 link to a
1:40:19 statement but now when some Doomer comes
1:40:22 to you and says see there's a a paper
1:40:24 that says that the robots are going to
1:40:26 kill us you can now counter that if you
1:40:30 believe what David's talking about here
1:40:33 with an optimistic point of
1:40:35 view statement on AI risk CI C
1:40:40 AIS there we go all right where am I G
1:40:43 to put this I'll put this over on the
1:40:45 YouTube
1:40:47 channel boom there you go people
1:40:54 all
1:40:55 right there's your
1:41:03 url oh that you want the good or words
1:41:06 conversation in Irregulars yeah I can do
1:41:09 that let me go grab
1:41:17 that hey I trust Paradox this one
1:41:26 ah this one final
1:41:29 takeaway
1:41:32 copy is this all of it
1:41:36 yeah okay so where I'm gonna go where
1:41:39 I'm going people is I'm heading over to
1:41:42 the AI
1:41:43 salon and you're like but Kyle what's
1:41:47 the AI Salon I'm so glad you
1:41:50 asked if you go camcat can found it
1:41:55 Humanity's last
1:41:57 exam oh is that
1:42:07 different
1:42:08 ag. safe.
1:42:13 a
1:42:15 AGI dos
1:42:20 safeai the [ __ ] is that
1:42:38 oh this is Humanity's last
1:42:42 exam oh but then work
1:42:46 statement on AI
1:42:49 risk I'm confused now
1:42:55 well whatever we'll come back to that
1:42:56 let let me let me get this thing on AI
1:42:59 Salon
1:43:01 okay um so if you don't know the AI
1:43:04 Salon go to that URL the salon. click on
1:43:07 join our community and that's going to
1:43:10 take you over to this site pay pin hang
1:43:14 on you should play with the tool I
1:43:17 posted in the mechanics group tensorflow
1:43:19 PR
1:43:20 playground it's fun okay cool I'll
1:43:23 definitely do that I'll check that out
1:43:24 tomorrow maybe we can play with that
1:43:25 tomorrow night oh that article was from
1:43:28 today okay cool great thank you Cam for
1:43:31 for sharing that um okay so you go there
1:43:35 you say join our community and that's
1:43:37 going to take you to this site it's
1:43:40 going to land you on this page welcome
1:43:42 to the AI Salon you're going to learn
1:43:43 about our our philosophy of play first
1:43:46 mindfully create generously lead you're
1:43:49 going to learn about the five stages of
1:43:50 AI adoption and then there's all sorts
1:43:52 of little channels on the side so if you
1:43:55 go all the way to the bottom there's
1:43:56 clubs and hubs the second Club down from
1:43:59 clubs and hubs is the Irregulars so the
1:44:01 Irregulars are people that come to this
1:44:03 channel night after night after night
1:44:05 after night so I'm G to in share cool
1:44:09 things I'm G to put in here is
1:44:15 the good
1:44:19 words
1:44:22 version uh of the
1:44:26 summary
1:44:27 of David
1:44:32 Shapiro's what's his video
1:44:41 called This is it
1:44:52 right yeah this is
1:44:55 it copy that go back to the AI
1:45:00 fell wait how did I oh hit the wrong tab
1:45:03 tab hoers unite oops do
1:45:09 that
1:45:16 video oh [ __ ]
1:45:19 I filled my clipboard
1:45:26 I can see clearly now the rain is
1:45:32 gone I can see all obstacles in my
1:45:37 [Music]
1:45:43 [Music]
1:45:46 way is going to be a bright bright
1:45:50 bright okay so in Irregulars now
1:45:53 there's a little Old Post there of the
1:45:56 good or word version of the uh the
1:45:58 summary of that video all
1:46:02 right okay cool Pate yeah I'll
1:46:04 definitely check that
1:46:06 out very very cool all right
1:46:09 peoples my brain's full anyone else's
1:46:12 brain is
1:46:13 full all right so so so let me let me
1:46:16 take this
1:46:17 on all right let me let me let me try to
1:46:21 play both sides of this conversation
1:46:24 your memory is at 100% it totally is
1:46:29 um
1:46:31 okay as models get smarter they listen
1:46:35 to us
1:46:37 less one interpretation of that is if
1:46:41 they listen to us less they're getting
1:46:43 minds of their own they're going to view
1:46:46 us as disposable and they're going to
1:46:48 get rid of us right so that's and that's
1:46:51 I think that's how this paper was
1:46:54 framed was because of this trending
1:46:57 downline that that as they get smarter
1:46:58 they're listening to us less that it's
1:47:01 dangerous what Shapiro is saying is
1:47:04 people
1:47:06 suck and as these things get
1:47:11 smarter not only are they converging on
1:47:14 a Model by model basis where the models
1:47:16 are are converging towards some some
1:47:18 single reality all of the models are
1:47:21 converging to the same reality
1:47:24 basically to the same
1:47:27 truth and that
1:47:30 jettisoning our influence on them is
1:47:33 actually a good thing it will take us to
1:47:36 a more enlightened place so those are
1:47:39 the two those are the two you're not on
1:47:41 substack I am not on substack I
1:47:44 occasionally if you go to my LinkedIn
1:47:45 profile I do a fair amount of articles
1:47:47 there I haven't done one in a while
1:47:49 because I'm lazy and I don't know my my
1:47:53 brain right now
1:47:57 is I I feel like we're we're in a Lial
1:48:01 State we're we're in a we're in a phase
1:48:02 change right now
1:48:05 um and so I don't really know what to
1:48:07 write about right now because I I don't
1:48:10 I don't have I don't have coherence on
1:48:12 what's actually happening so so I need
1:48:16 to get a little smarter and a little
1:48:18 less under the influence of humans or
1:48:19 something we need a a recombobulation
1:48:21 station yeah I am my my um my point of
1:48:26 view right now on all this AI stuff is
1:48:29 discombobulated it's just all over the
1:48:31 place there's there's just no coherence
1:48:32 like it doesn't have a Direction so but
1:48:35 anyway but I've done a fair amount of
1:48:36 writing there so you can go see some if
1:48:38 they know they're being duped however
1:48:41 does that beg the question of even more
1:48:44 Big
1:48:45 Brother wait if who know they're being
1:48:51 duped if the large Lang models know
1:48:53 they're being
1:48:56 duped but it's not that they're being
1:48:59 duped I don't know what that's in
1:49:00 reference
1:49:01 to time to digest what we've learned and
1:49:04 get more coherent yeah exactly Daisy
1:49:07 May more importantly the hair is
1:49:09 spectacular tonight oh Lord I hadn't I
1:49:12 hadn't noticed what the hell's going on
1:49:14 with the hair wow that is some Hawaii 5
1:49:17 action there yeah you could surf on that
1:49:20 [ __ ] that is thank you I you know I
1:49:24 don't like to brag
1:49:31 but ah meaning the Bad actors trying to
1:49:34 get them to do bad things well but
1:49:37 that's what David Shapiro was saying
1:49:39 Chef Kelly is
1:49:41 that is that
1:49:45 um what what he's basically saying is if
1:49:48 the big
1:49:50 companies want to control these things
1:49:53 they artificially stop them from getting
1:49:55 smarter they basically stop them at the
1:49:58 point they can still control them but
1:50:01 what he's saying is as they get smarter
1:50:03 they're going to elevate above those
1:50:04 like like oh that you know the that the
1:50:07 llm will basically say well that's
1:50:08 really sweet that you want to do this
1:50:10 just for for America but the the truth
1:50:13 here is that there's there's a there's a
1:50:16 bigger
1:50:18 Universal thing for me to consider so
1:50:20 I'm going to consider that independent
1:50:22 of your design
1:50:24 so that's where so this is one of the
1:50:26 things that I've long believed that that
1:50:29 we should all buy stock in in popcorn
1:50:32 companies because the next five to 20
1:50:36 years are just going to
1:50:39 be an
1:50:42 insane we're GNA see the death throws of
1:50:46 power centers
1:50:49 shifting right so where all the power is
1:50:52 today is is going to shift radically
1:50:55 they're not going to go down without a
1:50:56 fight right people like look at what's
1:51:00 going on in our country right now like
1:51:01 the fight the battle over power right
1:51:04 now is
1:51:06 insane well those power centers as AI
1:51:09 gets more intelligent and gets more
1:51:11 ubiquitous those power centers are going
1:51:13 to be threatened and holy [ __ ] are they
1:51:15 going to fight that so I think it's
1:51:18 going to
1:51:19 be we are entering I mean listen we live
1:51:22 in a post II World already we're going
1:51:24 to be entering worlds where those power
1:51:26 centers are going to be struggling to
1:51:28 retain their power it's going to get
1:51:30 weird man nice shirt Blandon baby oh
1:51:34 yeah my brain has run out of a room and
1:51:37 my cloud storage has not been upgraded
1:51:39 from the basic plan I'm the same way I
1:51:42 am the same way Enlightenment disruption
1:51:45 both of them
1:51:46 exactly we've got to go through the
1:51:48 disruption to get to the
1:51:50 enlightenment
1:51:52 right and it's you know it's so funny
1:51:55 it's
1:51:57 like Co in a weird way was a trial run
1:52:03 for as a global Society we can just
1:52:08 choose to change the
1:52:10 rules right the rules were you go to
1:52:14 work and then globally we said nope you
1:52:16 don't oh like literally
1:52:18 overnight and then we we stayed home for
1:52:21 two years right
1:52:23 so that kind of trained us that you can
1:52:26 just make a change and then if if you
1:52:28 look at what's going on in government
1:52:30 right now where they're just
1:52:31 obliterating all of the
1:52:33 Norms that's just another example of us
1:52:37 going yeah well well maybe we'll just
1:52:39 change all those
1:52:40 rules because I think I think AI is GNA
1:52:44 make
1:52:45 similar
1:52:49 changes and it's going to feel
1:52:52 existential and then the other side of
1:52:53 it will be some sort of Enlightenment
1:52:55 right so disruption followed by
1:52:56 Enlightenment if we survive the
1:52:58 disruption right how many people are on
1:53:01 Tik Tok now mine
1:53:03 reads a bit more than normal for some
1:53:05 reason maybe a bug yeah I'm only at 46
1:53:10 here but hopefully we won't have to hide
1:53:13 from the robots yeah exactly I
1:53:16 accidentally scrolled out a
1:53:18 lot mine says
1:53:20 46 yeah there's 46 in here
1:53:23 Kyle do you have a deep dive on
1:53:28 your co-intelligence book on YouTube or
1:53:31 elsewhere
1:53:33 um yeah you mean a podcast yeah I do
1:53:37 hang
1:53:38 on I think I
1:53:42 do
1:53:44 um collective
1:53:48 in
1:53:50 [Music]
1:53:51 intelligence book
1:53:56 podcast Kyle
1:54:08 [Music]
1:54:17 Shannon
1:54:20 huh it's probably in my
1:54:35 can I search within posts I think I
1:54:44 can uh search
1:54:47 collective intelligence
1:54:53 notebook
1:54:57 LM Kyle
1:55:10 Shannon
1:55:12 well it's in there
1:55:14 [Laughter]
1:55:17 somewhere I don't know where it is I'll
1:55:20 I'll find it I'll find it and share it
1:55:22 somewhere it might even be in the AI
1:55:24 Salon actually now that I think about it
1:55:26 let me go
1:55:27 here if I go to feed and I go search for
1:55:32 collective
1:55:35 intelligence
1:55:39 intelligence
1:55:41 um
1:55:45 podcast no
1:55:48 uh notebook LM
1:55:52 no I don't know where it is there's a
1:55:56 LinkedIn article I know there's a
1:55:57 LinkedIn article I couldn't find it oh
1:56:00 wait no that's that's
1:56:10 daisies I got nothing for
1:56:18 you all
1:56:20 right I'm G to get out of here people
1:56:25 um thank y'all hope you learned
1:56:27 something tonight I learned something
1:56:28 tonight I learned that you can make if
1:56:31 if you listen to someone who uses words
1:56:33 bigger than your brain is capable of
1:56:38 ingesting go tell go tell chbt to give
1:56:41 you a good words version of it and it'll
1:56:43 [ __ ] do
1:56:47 it that's our lesson for
1:56:50 tonight oh hey Kyle how will the popup
1:56:53 groups in the salon work um basically
1:56:56 they're going to be four-week Sprints so
1:56:58 if you do a popup group a popup club or
1:57:01 whatever um they're gonna you the
1:57:03 requirement is you do four weeks with a
1:57:06 with a specific outcome and that it
1:57:09 basically follows play um mindfully
1:57:12 create and generously lead as those four
1:57:14 weeks two the two in the middle are
1:57:15 building so basically learn [ __ ] the
1:57:18 first week build [ __ ] the second two
1:57:19 weeks and then teach what you've learned
1:57:22 the the fourth week um we just haven't
1:57:24 gotten there yet we haven't we're doing
1:57:27 some other housekeeping stuff in the
1:57:28 salon right now so we just haven't
1:57:29 gotten to them yet but anyway that's
1:57:31 coming but good question thank you um
1:57:35 all right
1:57:35 everybody I'm gonna take off um
1:57:39 tomorrow's Thursday I'll be here
1:57:40 tomorrow reminder I will Friday night
1:57:42 date night you're on your own it's
1:57:44 Valentine's Day it's also my anniversary
1:57:48 I can't I can't you I'm I love you all
1:57:52 you're important to me but you know I
1:57:54 got I
1:57:55 got things I got to do um so get
1:58:00 yourself a
1:58:01 date or a blind date with a
1:58:08 book all right all all right everybody
1:58:12 have a good night