Aug 30, 2024 - OpenAI Strawberry Explained: The Future of AI Reasoning

Description

Chapters

Transcript

0:14 [Music]
0:24 [Music]
0:29 all right you ready champy
0:30 [Music]
0:35 are you
0:37 ready it's Friday night date night
0:41 welcome oh I didn't put up my little
0:43 picture what what kind of operation is
0:46 this I thought I thought we were
0:47 professionals around here good Lord what
0:51 a
0:52 mess there we
0:54 go it's almost like I tried it
1:01 w w w w w w w w
1:05 [Music]
1:28 w your feel feeling nervous aren your
1:33 boy with quiet
1:36 voice and impeccable
1:40 style don't ever let them steal a
1:44 joy and G way keep them from running
1:50 wild they can kick dirt in your face
1:53 dress your to and tell you that your
1:55 place is in the middle when they hate
1:58 the way you shine
2:02 I see you tugging on your
2:05 shirt trying to hide inside of
2:08 [Music]
2:16 it let them
2:19 spin let them scatter in the
2:23 wind I've been to the
2:26 [Music]
2:27 movies I've seen how they
2:30 [Music]
2:33 and the jokes on
2:41 there you're feeling nervous you're
2:45 feeling something aren't you
2:47 girl it's your brother's world for a
2:51 while longer
2:55 [Music]
3:00 you got to
3:02 wait
3:08 got I had the lyrics up before I should
3:10 have kept them
3:12 [Laughter]
3:14 up oh good people Friday night date
3:17 night what is
3:18 [Music]
3:23 happening that's it champ you're just
3:26 giving us a a half number
3:32 [Music]
3:50 God you're
3:54 looking fine tonight
3:58 [Music]
4:00 every fellas got you in his
4:07 sight what you doing with a clown with a
4:12 clown like
4:15 [Music]
4:16 me surely would have last a
4:22 little something's off there
4:32 [Music]
4:38 mystery so
4:41 tonight I'll
4:44 ask stars
4:48 [Music]
4:50 above how did I ever wish your
4:57 love what did I do
5:05 what did I
5:06 [Music]
5:12 say to turn your
5:17 [Music]
5:19 Rous my
5:21 [Music]
5:25 way I'm the guy that never learned
5:32 [Music]
5:35 never even
5:37 got Second
5:40 [Music]
5:42 Glance cross a crowded room that
5:46 was
5:49 enough I could look but I could
5:54 never so
5:56 T I'll ask
6:01 [Music]
6:07 St how did I
6:11 ever Your
6:14 Love what did I
6:17 [Music]
6:23 do good evening good people how many
6:25 people we got in here 15 people in here
6:27 what the hell is going on Tik tock
6:31 Friday nights are normally a, 1500 2,000
6:34 people we got 15 22 people in here come
6:36 on Tik Tock let's get it going all right
6:39 people get the grandkids I know it's
6:40 Friday night date night they're in bed
6:42 that's not the point wake them up have
6:44 them start tapping the screen share the
6:46 screen love the champion what champ I
6:50 love that champion in his sidekick too
6:53 exactly he's the star of the show
6:55 although he just walked away little
6:56 [ __ ]
6:58 Diva woo
7:00 how y'all doing tonight we're here two
7:02 BL
7:03 Tom Joey in the house Milo's wife in the
7:07 house
7:09 [Music]
8:12 [Music]
8:23 is this place I can rest my poor kid
8:31 together my thoughts in sweet
8:35 [Music]
8:38 silence busiest place where the feelings
8:42 [Music]
8:45 arear from an over exposure to
8:50 violence this is the place I can slowly
8:54 face the only one I truly can know
9:00 these are tears from a long time
9:03 ago got these tears from a long time ago
9:08 I made to cry 30 years or
9:12 so these are tears from a long
9:18 term AG
9:25 [Applause]
9:26 go well old darling Oh Darling you say
9:30 unto
9:31 [Music]
9:34 me where have you been all my life
9:40 time well I have been swimming the seven
9:44 sad
9:47 Seas four women have tossed me their
9:52 lifelines I pull them into the waters
9:55 darker in I'd have warn them but I
9:58 didn't know
10:01 these are tears from a long time
10:04 ago got these tears from a long time
10:08 ago I need to cry 30 years or
10:13 so these are tears for a long
10:20 time ago
10:24 [Music]
10:36 no last note champy he's singing from
10:38 the other
10:39 room I don't know what's up with
10:43 [Music]
10:44 him well we went from 22 to 27 Tik tok's
10:49 they're they're chucking Us in the
10:50 algorithm
10:52 now [ __ ] it's Friday night date
10:55 night come on tick
10:58 tock I promise I'll show boobies and
11:01 talk about drugs
11:09 [Music]
11:25 [Applause]
11:32 [Music]
11:38 put your pink bow
11:40 oh hey Tick Tock uh yeah so
11:45 listen listen we got a
11:48 talk we got a talk talk this is talk
11:51 talk welcome to talk talk with me Kyle
11:54 Shannon
11:56 your
11:57 bedazzled host
12:01 with the most with the most pink bows of
12:04 any Tik Tok
12:06 influencer you never
12:10 watch oh we we lost three we're down to
12:18 24 I didn't realize this was one of them
12:20 their pansy channels
12:27 [Laughter]
12:33 oh well what are you going to
12:35 do you'll win some and you you lose you
12:38 lose the bow
12:40 [Laughter]
12:43 haters oh man then bows are scary I
12:47 guess see here's the thing you all have
12:50 been normalized to my madness it's the
12:54 the traditional the traditional
12:56 ticktockers they have no [ __ ] idea
12:58 they're like he looks sort of like Rosie
13:01 odonald how bad could it be and then on
13:03 goes the pink
13:06 bow oh man Hubba Bubba date night
13:12 [Music]
13:37 [Music]
13:54 10,000 words swarm around my head a
13:58 million more in books written beneath my
14:01 [Music]
14:07 bed ain't it like most people I'm no
14:12 different can't figure out how to hold
14:15 my
14:16 [Music]
14:19 hand and I know you need me in the next
14:23 room over I'm stuck in here all
14:26 paralyzed
14:29 [Music]
14:32 four months I got myself in ruts too
14:35 much time spending mirrors FR in yellow
14:39 [Music]
14:43 walls ain't it like most people I'm no
14:47 different like to talk on things we
14:50 don't know
14:52 [Music]
14:56 about making a big change in my life and
14:59 discussing it with myob bot what's theob
15:02 bot fantastic experience what's aob
15:05 bot you meano like the music
15:09 thing they've got Bots now did I miss
15:11 something or is it a something else bot
15:14 or is there AO bot I don't know
15:16 about what the hell are you talking
15:18 about archetyp archetypal architect I
15:21 can never say that
15:32 [Music]
15:39 oh I made a bot
15:41 specifically for sunno oh
15:44 interesting oh to write I
15:47 see so you you created a bot to create
15:50 suo lyrics but now you're
15:52 having existential life conversations
15:55 with it this welcome to the new world
15:59 [Music]
16:21 [Applause]
16:23 let us be LS will marry our fortunes
16:27 together
16:28 [Music]
16:29 [Applause]
16:31 I've got some real estate here in my
16:35 B so we bought a pack of
16:39 cigarettes and Mrs Wagner's pie and
16:42 we've all come to look for aica
16:50 [Music]
16:51 [Applause]
16:54 [Music]
17:04 [Music]
17:08 oh good lord good Lord good Lord I think
17:12 I think what we're going to start with
17:14 tonight we're going to start with a
17:15 little
17:19 [Music]
17:22 edication because it's a small group to
17:25 start it'll probably get bigger it
17:26 usually does that when they start slow
17:28 like this
17:33 but what that mean what that says to me
17:35 is the people that are here are the
17:37 regulars which is great hey
17:39 everybody um so I think what we're going
17:41 to do is I'm going to start out with um
17:44 David Shapiro's video on explaining how
17:47 he thinks strawberry works and then Mr
17:50 strawberry today said this guy gets it
17:53 this is the first guy that's actually
17:54 understood what it's actually
17:56 doing um so I think we'll go watch that
18:00 video we'll just we'll do like a little
18:02 watch party I'll watch it then I'll stop
18:04 it and talk about it if there's things
18:05 that are
18:07 interesting but I feel
18:10 like this is a good group to understand
18:13 [Music]
18:17 that it's what's what's fascinating
18:20 about these large language models
18:25 is you'll see when we watch the video
18:33 it's just like nested prompts it's not
18:36 it's not it's not as [ __ ] remarkable
18:41 as as it should be like I don't know
18:45 there's something about this new era of
18:48 computing where you just ask the
18:50 computer to do [ __ ] and it does that's
18:52 just still a little surreal to me
18:57 [Music]
19:08 and he explains why so one of the things
19:11 they're using this model for is creating
19:14 synthetic data for GPT 6
19:18 apparently and if you don't know what
19:20 synthetic data is this will explain that
19:22 as well and when you explains it you're
19:24 like oh that kind of makes sense it
19:25 makes it makes real good sense I'm
19:28 excited to have the night off from work
19:29 880 hour work week this week Joy py good
19:32 Lord you got to tell them to lighten
19:35 up tell everyone to do their work during
19:37 the summer and not just the week before
19:40 Labor Day that's I've heard I've heard
19:42 this complaint from a lot of people that
19:44 everyone's just getting dumped on cuz
19:47 their [ __ ] colleagues are like H hey
19:50 I got to get this thing done by the by
19:52 so I can I'm taking off Friday did you
19:53 know I'm going taking off Friday I'm
19:55 going to take off Friday we're we're
19:56 I'll tell you what we're doing we're
19:57 packing up the kid we're going on a road
19:59 trip we're going to do it we're going to
20:00 do I you know I know it's crazy there's
20:02 going to be cars everywhere but we'll be
20:04 careful I'll tell you that anyway uh
20:06 listen I got a I got a TPS report I got
20:09 to get out of you could just whip that
20:10 up for you that would be fantastic and
20:12 then we'll get that into Bob The Bobs
20:15 and uh boy they'll take that TPS report
20:18 you know they're going to read it over
20:19 the weekend do they ever sleep how do
20:21 they get all this done
20:26 anyway yeah so if you had that kind of
20:30 week I'm
20:34 sorry just sorry he wrote Frank out of
20:37 the script Frank will be back we're just
20:39 having our summer
20:40 break is the OT good you mean the TV
20:43 show I thought the OT was really good
20:46 season one
20:47 anyway hey Kyle did you see the
20:49 commercial for 1X Tech robot commercial
20:52 that new robot yeah looks amazing we'll
20:55 look at that
20:56 tonight TPS reports
21:01 yeah hey Peter yeah listen we're going
21:04 to need you to go on ahead and come on
21:05 in on Saturday
21:08 yeah yeah you know 900 a.m. or so is
21:12 that work for you anyway oh hey uh one
21:15 more thing Peter uh we're going to need
21:17 you to go on ahead and come on in on
21:19 Sunday too
21:26 yeah Bob Bob
21:29 I wouldn't exactly call it
21:31 [Music]
21:38 working it's a classic that film it's a
21:43 classic Michael
21:46 [Music]
21:52 Bolton Samir not not not going to work
21:56 here anymore anyway
22:00 [Laughter]
22:04 I'll just entertain myself here you guys
22:07 go ahead wait did we get it oh we got in
22:09 three more people um you guys go ahead
22:11 and entertain yourselves I'm I'm doing
22:13 the same here I'm just having a good old
22:15 time oh my God I worked as a computer
22:18 programmer and that actually happened to
22:21 me exactly what you're saying what the
22:23 the uh we're going to need you to go on
22:25 ahead and coming in on Saturday that one
22:28 oh oh hey hey one more thing we're going
22:30 to need you to go on ahead and come on
22:32 in on Sunday
22:33 [Music]
22:44 [Music]
22:50 [Music]
22:53 too oh man uh
23:00 I'm just tired and Kyle cracks himself
23:02 up I wait for these
23:07 moments I'm not even cracking myself up
23:10 these are not even my jokes this is just
23:12 my stupid ass half remembering lines
23:15 from a movie I saw 20 years
23:17 ago that cracked me up
23:25 then it was it was an odd week it was an
23:28 odd week
23:30 um like it it was for us it was some of
23:32 our clients were not there they were
23:34 just
23:36 non-existent or they canceled meetings
23:38 and then other of our clients were like
23:40 oh my God we got to get it done you can
23:41 to get it you can get it done you get I
23:44 got to go on vacation I I'll be I'm out
23:46 next week but we can right as soon as I
23:48 get back it'll be done right you can get
23:49 it done you can get it done so a lot of
23:53 that I'm sorry they said there would be
23:57 king if I if if I count the number of
23:59 slices it appears there's not going not
24:02 going to be enough slices oh yeah oh
24:06 this that that it does look tasty it it
24:08 looks tasty is oh no oh no there's
24:12 there's no more cake left I I didn't get
24:14 a cake could I get
24:23 a I know too I know too
24:27 many too many scenes from that movie oh
24:31 my God the opening scene where where
24:33 what's his name the the little dweeby
24:35 one is like singing to Michael Bolt Is
24:38 It Michael Bolton where he's singing to
24:40 the rap songs and the black guy walks by
24:43 the car and he's like closes the lock
24:47 and starts rapping a lot more quietly oh
24:50 my God and then Peter the the opening
24:53 thing in the Dallas traffic Source Camp
24:55 if you're on here where he cuts into the
24:57 other lane and then it's stops and the
24:59 other lane goes he cuts back and it
25:02 stops and that lane goes oh my
25:07 [Music]
25:14 God oh my God that movie that
25:18 movie it's my stap stapler
25:21 [Music]
25:24 stapler going to burn the place down
25:26 burn the place down
25:28 [Music]
25:32 uh Papa JJ yeah Oprah Winfrey on
25:35 September 12th we're we'll do a watch
25:38 party so it'll probably be we'll
25:40 probably go on at 6:00 that night CU I
25:42 assume it's prime time is is 8:00 P.M
25:44 Eastern is that right is that still
25:46 Prime Time
25:48 Eastern
25:50 um but whenever that show goes live um I
25:54 think we'll we'll live stream it if if
25:57 we can we might get we might get Tik Tok
25:59 banned but if we can if we can stream it
26:01 we'll stream
26:02 it so September 12th Oprah Winfrey is
26:06 interviewing Sam Alman and Bill Gates
26:09 live and so this is right around the
26:12 time when you know they said things were
26:14 going to be launching in the fall and
26:16 strawberry seems to be confirmed and all
26:19 that stuff so it would not surprise me
26:23 if they didn't do a little bit of
26:24 coordinated
26:25 launching right around the time they're
26:27 doing you know
26:29 an audience of however many millions she
26:31 is tens of tens of Millions for
26:35 [Music]
26:53 Oprah 9:12 is a Thursday okay cool
26:58 beautiful
27:00 so Thursday night so Wednesday night
27:03 I'll be flying back from Washington DC
27:05 so Wednesday Tuesday September 10th I'm
27:09 flying to DC we got a big dinner and
27:12 then on Wednesday the
27:16 11th I'm there with 11 other AI
27:19 entrepreneurs and we're talking to
27:21 senators and [ __ ] like that about AI so
27:24 I'm going to talk a lot about the AI
27:25 salon and this Channel and what going on
27:29 and how our legislators should not just
27:32 legislate from
27:34 fear and from
27:37 ignorance they might want to go on ahead
27:39 and try the chat
27:42 GPT and understand what the [ __ ] it is
27:45 and understand how empowering it can be
27:48 and legislate from that position as well
27:52 so that's going to be that's going to be
27:53 my
27:54 message I tried to tap while you play
27:57 guitar that was a lot of tapping to
27:58 night what time is it yeah it's a little
28:00 bit late um too bad klouse Schwab
28:04 couldn't make it on Oprah with Bill
28:08 Gates um are you still coordinating with
28:10 Google yeah it's still it's still uh The
28:12 Flyin is jointly sponsored by Google for
28:16 small business and the salon which is
28:18 exciting um that's overly optimistic I
28:22 don't know what that's in reference to I
28:24 assume me saying something's going to
28:25 launch I agree it's overly optimistic IC
28:28 to assume that anything's going to
28:30 launch ever
28:31 [Laughter]
28:36 again oh man all right wait I got to put
28:39 my charger on my
28:41 computer give myself some hydration
28:49 cheers oh
28:54 Lordy um I think it's important
29:18 [Laughter]
29:26 Peter Michael the hell's going on man
29:29 people thought you were going to come in
29:30 here and start shooting no I just can't
29:32 get my address book I'm not going to
29:34 stay I got a phone number mic that I
29:36 don't want to lose what Peter you're in
29:38 deep [ __ ] you were supposed to come in
29:39 on Saturday what were you doing Michael
29:43 I did nothing I did absolutely nothing
29:46 and it was everything that I thought it
29:48 could be well I hope you have a better
29:50 story than that for Lumberg you know
29:52 you're supposed to be having your
29:52 interview right now with the Consultants
29:54 the who the Consultants what has gotten
29:57 into you
30:00 oh
30:01 yeah
30:03 right te here got to postpone it man
30:07 tell me you been sick make someone up oh
30:09 no way no I feel great it's the best day
30:11 of my
30:12 [Music]
30:14 life SP looks like a Peter
30:17 Gibbons oh there you are we just talking
30:20 about you you must be Peter Gibbons
30:22 uhhuh terrific I'm Bob Sidell this is my
30:25 associate Bob Porter oh hi Bob Bob go
30:28 ahead and grab a seat and join for a
30:29 minute or
30:30 two you see what we're actually trying
30:33 to do here is we're just we're trying to
30:35 get a feel for how people spend their
30:37 day at work so if you would would you
30:40 walk us through a typical day for
30:44 you yeah great well I generally come in
30:49 at least 15 minutes late uh I use the
30:53 side door that way Lumberg can't see
30:55 me and uh after that I just sort of
30:59 space out for about an hour till
31:01 I'm space out yeah I just stare at my
31:05 desk but it looks like I'm working I do
31:08 that for uh probably another hour after
31:11 lunch too I'd say in a given week I
31:13 probably only do about 15 minutes of
31:15 real actual
31:19 work uh Peter would you be a good sport
31:22 and indulge us and just tell us a little
31:25 more oh yeah let me tell tell you
31:28 something about TPS
31:32 reports the thing is Bob it's not that
31:35 I'm lazy it's that I just don't care
31:38 don't don't care it's a problem of
31:41 motivation all
31:43 right oh you're
31:46 welcome all right let's go f let's go
31:50 let's go find some actual AI stuff here
31:53 um I'll tell you what actually you know
31:56 what hang on there's there's there
31:59 there's something
32:00 here there's something here to talk
32:04 about
32:06 so if you have not memorized all of the
32:10 funny scenes of office
32:13 space there's your homework for the
32:15 weekend you got to get your [ __ ]
32:17 together
32:19 um the TPS reports the central the
32:23 central object of office space are these
32:26 [ __ ] TPS reports that it's all anyone
32:29 talks about and it's all they do and no
32:31 one reads them right they're just
32:32 they're just this
32:34 thing this is the kind of [ __ ] these are
32:38 the kind of jobs those are like the the
32:40 TPS reports that's the kind of [ __ ] that
32:44 AI is going to automate the snot out
32:46 of and those
32:50 tasks are going to go away right because
32:53 they're just going to be able to happen
32:56 and and then so
32:58 smart asses like Peter you know they
33:01 either won't work there or they'll move
33:03 up in management like he did um because
33:06 he he's smarter than you know the reason
33:09 he sucks at his job is because his job
33:12 sucks and he knows it and he knows no
33:14 one gives a
33:17 [ __ ] that scen
33:19 about did you did you get the memo we're
33:22 put now we're putting the covers on the
33:24 TPS reports
33:33 he does 15 minutes cuz he knows AI
33:35 exactly that's going to be the new thing
33:37 shareholder Journey oh man I think it
33:41 may be amusing how fast upper management
33:43 finds out its
33:44 replaceable Kyle is your video still at
33:47 99% yeah in fact I rendered another
33:51 one which is also now at
33:54 99% so I don't know what the [ __ ]
33:57 going on with with cling
34:00 video It's both of these are hung
34:06 now which is a real
34:13 drag anyone else have H hung videos on
34:18 cing this is funny I like this movie
34:20 yeah it's a really good movie
34:23 um okay AI learning lab
34:30 all right so let me give you a little a
34:32 little context those of you that are in
34:35 the channel if you're if you're new here
34:36 my name is Kyle Shannon this is the AI
34:37 learning lab as evidenced by the little
34:40 sign that says welcome to the AI
34:42 learning lab it's fantastic isn't it
34:46 Danielle did that for us she made it out
34:48 of clamation she's a she's a modeler she
34:51 goes to the uh to the dollar store and
34:53 she buys the uh the 10 packs of clay and
34:57 she's really good she puts them in um
35:01 she actually mixes the colors right you
35:03 got the 10 that are all kind of harsh
35:05 she actually mixes them to get these
35:06 more subtle
35:08 colors look at this she made this for
35:11 us she made like a little Tick Tock
35:13 symbol that looks like it's out of focus
35:16 all with Clay amazing
35:22 right that was this was definitely not
35:25 made with AI this was definitely this is
35:28 handcrafted see here here on the AI
35:31 learning lab we will not use art well so
35:35 you can't even call it art if it's made
35:37 with AI it's not art right so the only
35:40 thing that's got any value in this world
35:43 is things that are
35:44 handcrafted and so everyone here knows
35:46 that and they know that we don't really
35:48 like AI that the whole AI learning lab
35:50 what we're learning here is how bad AI
35:53 is and and how you shouldn't use it and
35:56 so this image is actually she took a
35:59 photograph with this with a Nikon
36:02 D50 a film camera with
36:05 film she took the film down to the uh
36:10 down to the the Walgreens where they
36:11 still process film she had that
36:14 processed and then and then she she took
36:17 that she scanned it into her computer
36:20 with a scanner not one of those
36:22 all-in-one printer scanner combos that
36:26 for from where I sit and this just could
36:28 be me if you're using a combo printer
36:31 scanner it's not really a pure scanner
36:34 so that's probably too digital for our
36:37 taste here we want this as analog as
36:40 possible so she's got a flatbed scanner
36:43 she scanned the photograph that was
36:45 printed on paper that was shot on film
36:48 of a clay sculpture that she made with
36:51 her hands and then and then she sent me
36:54 that uh via VIA FedEx on a on a floppy
36:58 drive and so that that's how we get the
37:02 artwork
37:04 here so thank you Danielle for your work
37:08 that I mean really seriously must have
37:10 taken
37:11 hours
37:16 hours you can make money with see here's
37:20 the thing I'm a smart ass that was all
37:23 just fibs that was all just fibs
37:25 Danielle made this in like 30 seconds
37:30 she made it with AI oh my God I had y
37:33 all going you were probably all like
37:35 well where do I get clay I want to mix
37:37 custom color hey that's crazy talk I
37:43 know
37:45 anyway I guess I guess I'm in one of
37:48 those
37:50 moods indeed it took seconds hours hours
37:54 to handcraft
37:56 this Suzanne you got
38:02 me oh good he listen it is my joy in
38:05 life to uh to to to say stupid [ __ ] so
38:10 dry and convincingly that people go
38:14 really is this that is that really how
38:16 she did
38:21 that I'm impressed more with the
38:25 Improv I totally get tricked into that
38:28 that's
38:34 fantastic what' you do Friday night I
38:37 got Bamboozled I got Bamboozled by by a
38:40 theatrical that wears pink bows on his
38:42 head
38:44 incorrectly okay
38:49 listen if Cory Sandler pottery's in here
38:52 she cannot
38:56 speak okay in in fact women you cannot
39:01 speak this is not this is not for you to
39:04 you have you have no input on this
39:07 none men men in
39:14 here how is this wrong I mean other than
39:18 the fact it's a pink
39:20 bow and I look [ __ ]
39:23 ridiculous I have been accused of not
39:25 knowing how to wear a headband you tell
39:28 me what I'm doing wrong men cuz the
39:30 women they've got some [ __ ]
39:31 conspiracy theory about pink bowed
39:35 headbands that men apparently didn't get
39:38 training on in the seventh
39:40 grade so
39:45 anyone what if I don't identify as a
39:49 woman as a
39:51 Shira
39:54 Shir see let's see seems normal to me
39:57 see AR typle architect exactly this is
40:00 other than the pink bonus of it this is
40:04 absolutely correct it's wrong take it
40:07 from a guy who grew up with all sisters
40:10 and a single mom ah damn it how's it
40:12 wrong what's it supposed to
40:17 be is it is it like you're supposed to
40:20 have it up like this is this
40:22 it cuz that looks [ __ ]
40:25 ridiculous I mean you know not that this
40:28 doesn't look ridiculous what happened to
40:31 your hairline shut up it's male pattern
40:34 baldness I'm getting
40:41 old just trying to get in touch with my
40:43 feminine
40:46 side I'm aggressively getting in touch
40:48 with my feminine side I'll wear the pink
40:51 bow the way I want to wear the damn pink
40:53 bow
40:56 ladies blood pressure goes
40:58 [Laughter]
41:03 up and scene let's bring it
41:07 down I feel pretty oh so
41:11 pretty Three Mile Storage man Jim Ross
41:14 is like what happened to this
41:18 channel I go away for a week and I come
41:21 back to
41:22 this well we talk about the salon right
41:26 so now we have salon product I wonder if
41:29 I have any hairs growing I've been
41:30 monoxide
41:33 diling oh all right add more jazzer size
41:38 this is definitely this is definitely
41:40 jazzer size era era autra M right leg
41:47 warmers wait should it
41:49 be
41:51 oh
41:53 maybe no
42:01 I'm going into an add hyperfocused
42:04 tunnel
42:05 on on how this should be worn that's
42:09 really
42:14 bad maybe it's a bow tie oh that's well
42:18 I'll tell you what if we did this with
42:20 the manini I I'll tell you what this and
42:25 a manini we got ourselves a [ __ ]
42:28 [Laughter]
42:36 party oh
42:38 lord it's just the Friday night date
42:42 night's not going well if you brought an
42:44 actual date here I'm sorry about this uh
42:47 I will apologize right now to whoever
42:49 you brought that has not experienced
42:51 this before it's not normally like this
42:55 it's normally much worse um so I
42:57 apologize for
43:00 that okay chip Andale Kyle actually uh
43:04 my my uh co-founder of agency.com and I
43:07 Chan he was he was a really big tall
43:09 Korean dude but he and I are both a
43:10 little on the chubby side so Chip and
43:13 dals was all the rage when we started
43:15 agency.com and uh so so we said we're
43:19 going to open we're going to open a club
43:21 called Chubb andales where the guys
43:23 start out naked and then the women pay
43:25 the money to put their clothes back on
43:27 [Laughter]
43:31 it's chub andales but so by the way
43:34 Irregulars if you want to make some
43:35 Chubb and's images and toss them in
43:37 Irregulars I think that could be kind of
43:41 fun where was my life before this
43:44 channel you were watching actual
43:46 entertainment on Netflix I
43:48 think this this this is this is like
43:52 sucrose it it's it's a substitute for
43:56 entertainment and it's bad for
44:03 you oh man sounds like banfa for me yeah
44:09 probably yeah tick tock tick tock with
44:12 there you said a dirty word we're going
44:14 to we're going to age restrict this
44:16 channel we're not going to promote you
44:17 you you dirty
44:19 birdie you said boobies whatever you do
44:22 don't show a booby or we will ban you
44:24 for life ban you for Life we'll do it
44:27 we'll do do it we have an AI for
44:32 that it's crazy although I did see a Tik
44:35 Tok today that apparently somewhere in
44:37 Tik Tok now you can actually clone your
44:40 voice and then you can just type scripts
44:43 and it'll make videos of you talking
44:45 that sounds not not much like
44:49 you for what that's
44:51 worth um okay so let me give you a
44:55 little uh
44:57 I'll give you a little teeny tiny
44:59 history lesson wait he wears the pink
45:01 bow and then we're going to get a
45:02 history lesson this is the worst date
45:05 night ever can I have some more
45:07 nachos
45:09 um did you microwave these they're
45:13 chewy you got to put them in the air
45:15 fryer Bob oh God do I have to do
45:19 everything myself that's not how you
45:21 wear a pink bow Bob
45:28 um
45:32 [Laughter]
45:38 sorry now I want nachos now we've got
45:41 three three comments on nachos it's good
45:44 I like
45:50 it try the pink bow dragon egg and talk
45:53 about the atmology of emology I see that
45:56 would be good why not why not squirrel
45:59 welcome to Chad
46:02 ad hey
46:04 everybody uh what are we going to talk
46:07 about right history lesson
46:12 okay
46:15 um
46:17 so last November November
46:23 2023 open AI had just launched custom GP
46:27 s we were all fired up about it we're
46:29 building gpts we're figuring them out
46:32 we're like yeah gpts are going to change
46:34 the world we're going to make a fortune
46:35 they're going to do a GPT store
46:39 woo and then like two weeks later they
46:42 fired Sam
46:46 Alton and then everyone's like oh [ __ ]
46:48 and then they unfired Sam Alman everyone
46:50 was going to quit and
46:54 then but apparently what happened
46:58 was Ilia
47:01 sus and probably some other people but
47:04 it sounds like
47:05 Ilia saw
47:08 something in their technology that
47:10 spooked the living [ __ ] out of him and
47:12 so he went to the board and said uh
47:15 Sam's going
47:16 to put this in the world and he
47:19 shouldn't do that so we should fire him
47:22 and then the board said yeah we Sam's an
47:24 [ __ ] we don't like him let's fire him
47:26 this is so much more fun than Netflix or
47:28 sucros thank you I appreciate that AI
47:32 learning lab nearly as good as
47:33 artificial
47:36 sugar
47:39 um so the world of open AI got tossed
47:43 into chaos
47:45 um and and what it seems like has
47:49 happened is that since that moment
47:52 they've
47:52 been working on testing finishing
47:58 this thing that is called strawberry
48:01 project strawberry or it's called
48:03 project
48:04 Orion um it was
48:07 announced yesterday or the day before in
48:09 the information that has a bit of a
48:12 leak um and it it indeed is is a
48:17 thinking
48:19 reasoning GPT or or chat GPT or model
48:24 from open AI like supposed to be quite
48:28 good quite
48:29 powerful
48:31 um and the other thing you know over the
48:34 over the past six or eight months there
48:36 have been a lot of announcements from
48:38 open
48:39 AI where they announce something
48:42 everyone gets all hot and bothered like
48:43 the Sora video model and the real-time
48:46 voice and realtime voice they're like
48:49 they they announced that back in April I
48:50 think and they're like couple of weeks
48:52 it'll be out it's going to be fantastic
48:53 you're going to love it and then couple
48:55 of weeks came and went and then four
48:57 weeks came and went and they're like
48:58 yeah about that couple of weeks it's
49:00 going to be a couple of weeks more and
49:02 then they were like yeah about that
49:03 couple of weeks more it's going to in
49:05 the fall yeah yeah that's the ticket
49:09 yeah
49:12 yeah and so nothing's been
49:16 launching everyone else has been
49:18 catching up all of the Just just
49:22 [ __ ] Technologies everywhere that are
49:25 all of about the same class level um but
49:30 it does seem now
49:33 like some progress is happening Sam mman
49:37 tweeted
49:38 yesterday or the day before I think it
49:41 was yesterday yeah it was
49:43 yesterday that they've come to an
49:46 agreement with the US government the
49:49 AI I don't know review panel board [ __ ]
49:53 whatever thing it is that they've come
49:56 to an agreement for how they're going to
49:58 do testing to release these more
50:00 powerful models so I assume what that
50:03 means is that sometime over the past 6
50:05 months they've been interacting with the
50:06 government there were some tweets about
50:08 that as well to let them know what's
50:10 coming and to show how they did the
50:13 testing so that it's safe check your
50:16 regulars please oh
50:22 no oh we got a video
50:31 oh [ __ ] that's really
50:38 good
50:43 wow so if you were not here last night
50:47 we did a we did we did what took way too
50:49 long we went to chat GPT we generated
50:52 some concepts for um short videos that
50:56 have an opening frame and a and a
50:58 closing frame and then we had three
51:01 different image models generate start
51:03 and end frames and then we the the
51:04 intent was to do it in three different
51:06 video models but only one of them
51:08 actually worked and then it hung so this
51:11 is this is the one that hung but this
51:14 this actually is
51:17 it that that's actually staggeringly
51:20 good the prompt was envelope magically
51:23 opens itself and bursts into flames I
51:26 would say it bursts into flames and then
51:28 magically opens itself
51:32 but that's pretty
51:37 amazing um Irregulars Channel okay
51:40 anyway that's pretty
51:43 slick new activity oh yeah nachos anyone
51:47 chub and
51:49 dals oh chub and Dales I like
51:55 it now listen
51:57 if you word to the wise if you ever
52:01 start a Tik Tok
52:03 channel in a time in history where
52:06 people can not only make fun of you but
52:09 they can make fun of you instantly with
52:13 hilarious comical images uh oh it's
52:18 going to sting it's going to
52:24 sting Chum and Tails
52:28 take my money put your shirts on please
52:35 please are those your fantasy men no can
52:38 I leave can we leave how do we get out
52:41 of
52:49 here the Church of the bow
52:53 beautiful you all are [ __ ] weird
52:59 oh man that's
53:02 hilarious that's the chub and's
53:07 winner that actually looks just like the
53:09 bow that's pretty amazing I think this
53:12 one it's with the nachos the nachos
53:15 really make it this is this is like the
53:18 stage Mom this is the guy at chub
53:20 andales that wasn't quite good enough to
53:22 make it on stage so he's like the he's
53:25 like the backstage mom always cooking
53:27 nachos for the
53:35 boys oh good lord all right so okay
53:42 so
53:44 so so what appears to be
53:48 coming other than more more comedy
53:52 images at my expense um what seems to be
53:56 coming is this new powerful engine from
53:59 open
54:01 AI um David Shapiro if you don't know
54:04 him you should know him so we're we're
54:06 going to hop over to
54:09 YouTube David
54:14 Shapiro he made this video he couldn't
54:17 he couldn't sleep because his mind was
54:19 all a
54:20 flutter David Shapiro
54:24 strawberry there was also a whole bunch
54:26 of drama with the strawberry Twitter
54:28 account and a whole bunch of hype around
54:30 it but it does appear to be
54:32 coming so what I thought would be good
54:35 to
54:37 do is watch this video together better
54:41 faster cheaper pick all three dcom AI
54:44 first customer
54:47 service in Puerto Rico we call ourselves
54:51 B we're proud hello everybody I awake
54:54 and I don't want to be because I
54:55 couldn't shut my brain off
54:58 so so David is neuros spicy yes he's
55:00 wearing a Star Trek shirt he always
55:02 wears a Star Trek shirt and he is one of
55:05 the most Brilliant Minds in in AI he
55:09 he's super into open source he's super
55:11 into AGI he's written I think five books
55:14 on AGI if you don't know what AGI is it
55:16 stands for artificial general
55:18 intelligence how many people do we have
55:20 in here 50 okay
55:22 um really knows this [ __ ] um if you if
55:27 you tend on the geeky side he's someone
55:29 to follow if you if you want to be
55:33 educated about how these tools work and
55:35 what what the implications are of them
55:38 he's also good to follow you don't need
55:40 to be an engineer to follow him but if
55:42 you're an engineer he's got a lot of
55:44 content that's that's you know more
55:46 Technical and and open source projects
55:48 and things like that
55:52 um
55:53 okay let's see I never remember to even
55:57 go over there
56:01 now it's the hair for me
56:05 okay
56:12 okay he sure does love speculating
56:15 wildly he annoys me to no end of course
56:18 he does of course he does
56:20 Pate there's lots better people on the
56:22 engineering side to follow uh there may
56:25 be I again I'm not on the engineering
56:27 side so this is this is about as
56:29 technical as I get and I I do like me
56:31 some good speculation so anyway um this
56:35 is his speculation on how qar Works how
56:38 strawberry works so let's watch I think
56:41 I know how strawberry
56:43 works I'm not even kidding so if you're
56:45 not familiar strawberry has been all the
56:48 rage qar strawberry the leaks from open
56:51 AI now the rumor has it right now and of
56:54 course just rumor we'll see if it gets
56:56 it's walked back but they're talking
56:59 about how strawberry can solve complex
57:01 math problems but then there's also talk
57:03 about synthetic data and that synthetic
57:05 data will be used to train project
57:08 o now something stuck out to me when
57:12 they're talking about having solved uh
57:15 the generation of synthetic data so next
57:19 I want to point out I was talking about
57:21 synthetic data over 2 years ago so I was
57:25 using gpt3 to synthesize data to create
57:28 synthetic data sets for fine tuning the
57:31 reason that I'm sharing this it's not a
57:32 flex I mean it is a little bit of a flex
57:34 but I'm pointing out that like I have
57:37 been working on this for a while um and
57:39 I've also been sharing this you see this
57:41 is the open air yes aio's wife he is
57:43 wearing a Star Trek uniform he always
57:45 wears Star Trek uniforms he's neur
57:47 neuros spicy like the rest of us he's
57:51 he's special neuros spicy I for I've
57:53 been sharing it openly for a while as
57:55 well
57:56 so that's one thing by the by the way he
57:59 refers to himself as neuros spicy so I'm
58:01 not outing him in some in any kind of
58:03 derogatory way thing is just keep in
58:06 mind synthetic data it's something that
58:08 I'm actually a world leading expert in
58:12 um and then I want to talk about latent
58:14 space Activation so latent space
58:16 activation um basically what we realized
58:19 is that
58:21 if uh all right here's here's another
58:24 way to think about it you train these
58:26 models on gigantic data sets and there's
58:28 a lot of embedded knowledge even if it's
58:30 not fully crystallized and so what I
58:32 mean by crystallized is it knows things
58:36 but it has to assemble them and this is
58:38 very similar to how human brains work is
58:40 that um have you ever had a conversation
58:42 where someone is asking you about
58:44 something that you know a lot about but
58:46 they ask you a question that you haven't
58:48 really thought of before and then at
58:50 that time of instantiation you kind of
58:52 connect all the dots and then you say
58:54 then you spit something out you're like
58:55 wow I knew more about that than I
58:56 thought that I did that's what latent
58:58 space activation is that's the
58:59 equivalent of what latent space
59:00 activation is and what what the latent
59:02 space is if you don't know when they
59:04 when they train these models and they
59:06 they put all this data and information
59:08 in it converts it into these things
59:10 called vectors which is literally like
59:13 thousand dimensional Vector space you
59:15 know like an XY graph where you can
59:17 chart numbers on a two-dimensional plane
59:21 the latent space is like thousand
59:23 dimensional crazy complicated
59:25 mathematical space where there's all
59:27 these points in space that represent the
59:29 tokens that's what he's talking about so
59:31 it's basically it's basically machine
59:34 learnings or like you know GPT 4's brain
59:37 is the latent space but you know all all
59:40 anything that uses a Transformer uses
59:43 that for uh large language models and I
59:47 was working on this 10 months ago um
59:50 Source Camp I was talking about lat and
59:52 space activity earlier today it's
59:55 activation but yes I'm sure you were and
59:58 I made videos about it as well so that
1:00:01 data is out there the next what I want
1:00:04 to show you is how I think strawberry
1:00:07 works at a very high level just using a
1:00:09 generic example on Claude And the reason
1:00:12 that I think that I know how it works is
1:00:14 because I was working on this in my
1:00:16 startup that I ultimately didn't go
1:00:18 anywhere 18 months ago the the other
1:00:21 thing that I think is worth pointing out
1:00:23 before he digs into how this works when
1:00:25 he's talking about synthetic data what's
1:00:27 going on there is so gp4 they they went
1:00:30 out and they scraped as much of the
1:00:32 internet as they could they they got
1:00:34 these web scrapers that go out and find
1:00:37 data they probably used illegal sources
1:00:39 of data but they basically went out and
1:00:41 they got as much content off the public
1:00:44 internet as they could and then they
1:00:46 embedded it put it into this latent
1:00:49 space as these models get bigger and
1:00:51 bigger and bigger we are effectively
1:00:54 running out of data to train them so the
1:00:57 the basic thing that open AI cracked the
1:01:00 nut on is they they realized that the
1:01:02 more data and the more compute power
1:01:05 that you throw at um creating these
1:01:10 these giant models the better they get
1:01:13 um they're effectively running out of
1:01:16 data and and a lot of the new data is
1:01:18 just going to be crap that gets vomited
1:01:21 out of old AI models and it's going to
1:01:23 suck and it's going to start to pollute
1:01:25 the data so synthetic data data
1:01:27 basically is instead of going out to the
1:01:30 internet and finding all the data you're
1:01:32 actually generating it out of one of
1:01:33 these earlier models and by using this
1:01:36 reasoning engine what he's going to talk
1:01:38 about here it'll it'll start to make
1:01:40 more sense but basically they're
1:01:41 creating data that is high quality data
1:01:45 that the new like GPT 6 can be trained
1:01:47 on and that's what he's that's what he's
1:01:49 about to talk
1:01:53 about who are all those people pay are
1:01:56 those all the people that worked on the
1:02:05 Transformer
1:02:08 okay so I'm getting really excited now
1:02:11 at the startup that I was working on I
1:02:13 was developing a concept that I called
1:02:15 surface plate surface plate would have
1:02:18 been a combination of three models oh
1:02:20 that's great so great so can p can
1:02:23 either you do that if you haven't can
1:02:25 you toss that in the regulars Channel
1:02:27 and maybe in um well definitely in in
1:02:31 the tech in the tech Guild but maybe
1:02:33 also in the um in water cooler would be
1:02:37 a good place to put those names good
1:02:39 good people to follow on YouTube that's
1:02:40 awesome thank you for that oh you'd have
1:02:43 the generator or the the expert then
1:02:46 you'd have the interrogator oh wait hang
1:02:48 on this is important let me go back a
1:02:49 little bit since I rambled before this
1:02:53 would have been a combination of three
1:02:55 models so you'd have the a concept that
1:02:58 I called surface plate surface plate
1:03:01 would have been a combination of three
1:03:03 models so you'd have the generator or
1:03:06 the the expert then you'd have the
1:03:09 interrogator um which is basically
1:03:11 asking the the expert model what do you
1:03:13 know about this and then you'd have a
1:03:15 grader which would be a third model that
1:03:17 would grade the quality of output you
1:03:20 can do all this with chat Bots now so in
1:03:23 this first one imagine that instead of
1:03:25 me asking this question this is a this
1:03:28 is a chatbot that is fine-tuned or just
1:03:30 instructed to interrogate the expert
1:03:33 model and extract all of the information
1:03:35 out of it to to create that latent space
1:03:38 activation saying I know you know about
1:03:40 this at a very deep level now tell me
1:03:42 everything you know about it to
1:03:43 crystallize it now you might remember
1:03:46 the uh the paper that Philip covered
1:03:48 over on AI explained textbooks are all
1:03:50 you need what we found is that synthetic
1:03:53 data when it is highly curated is
1:03:55 actually much more efficient at training
1:03:57 these models so that's all the theory
1:04:00 behind it so in this first example I
1:04:02 said tell me everything you know about
1:04:03 physics at a high level um we will
1:04:05 iteratively drill into topics yada y y
1:04:08 and so then uh Claude happily generates
1:04:11 the top 10 qu uh categories of physics
1:04:15 and then I say excellent unpack
1:04:16 fundamental forces and particles because
1:04:18 that was the first one so by recursively
1:04:21 searching through the data remember what
1:04:23 qar does is it's a apparently a
1:04:26 recursive search What's the title of
1:04:29 this um video so David Shapiro is the
1:04:32 guy and the video is called how open AI
1:04:35 strawberry works and then in quote quote
1:04:38 marks one textbook to rule them all my
1:04:41 educated
1:04:43 guest algorithm now what I'm not quite
1:04:46 sure on is is how this would cover math
1:04:48 so I'm probably missing something but at
1:04:51 least I know how to generate synthetic
1:04:53 data and if you have enough tokens you
1:04:55 can recursively generate or or or
1:04:58 synthesize or you know do latent space
1:05:00 activation for all of human knowledge in
1:05:03 these models and you can distill it and
1:05:05 then also
1:05:07 make let's go back and listen to that
1:05:10 statement enough tokens you can
1:05:12 recursively generate or or or synthesize
1:05:16 or you know do latent space activation
1:05:18 for all of human knowledge in these for
1:05:20 all of human knowledge so basically we
1:05:23 can we can synthetically generate
1:05:25 anything on any topic for all of human
1:05:27 knowledge um and so if if you didn't get
1:05:31 what he's talking about there these
1:05:32 three models are you basically have your
1:05:35 core llm right that's answering your
1:05:37 questions you then have basically a
1:05:39 chatbot that's just good at asking
1:05:41 questions of that thing to vomit out
1:05:43 data and then the third model is
1:05:45 something that looks at what gets
1:05:47 vomited out and you know figures out if
1:05:50 it's good or not so it's three different
1:05:51 models interacting that's what he's
1:05:53 talking about here yeah and he says it
1:05:55 so casually I know models and you can
1:05:58 distill it and then also make new
1:06:00 connections so fundamental forces and
1:06:02 particles it started with just uh three
1:06:04 bullet points and then in the next chat
1:06:07 and
1:06:07 remember you like this does not take
1:06:11 much intelligence to interrogate again
1:06:13 so you could very easily fine-tune
1:06:15 another model to just uh be the
1:06:18 interrogator and it say certainly so
1:06:20 then it unpacks more about fundamental
1:06:23 forces um and so then uh let's just I'll
1:06:26 show you again it's like okay gravity is
1:06:27 the first one so I'm recursively going
1:06:29 through just basically a binary search
1:06:32 tree um tell me everything you know
1:06:35 about gravity and away we go so again it
1:06:39 does not take that much intelligence to
1:06:41 interrogate and activate and and create
1:06:44 those lat and space activations and then
1:06:46 what it's doing is it's synthesizing
1:06:48 information and so then what you do is
1:06:52 then you take it to the third model and
1:06:54 so in this case uh I said you're a
1:06:56 grader of data I will give you a bit of
1:06:58 text and you will grade it with a rubric
1:07:00 what we found a long time ago is that
1:07:03 and lots of people independently
1:07:04 discovered this by the way is that
1:07:06 language models are really good at
1:07:08 discriminating so you have a generator
1:07:10 and a discriminator but in this case we
1:07:11 have an expert an interrogator and a
1:07:14 discriminator they're really good at
1:07:16 grading each other with rubrics um
1:07:18 particularly if you tell it how to grade
1:07:20 it and on what criteria now I didn't
1:07:22 give it uh a full rubric so a full
1:07:24 rubric would be grade one is this grade
1:07:26 two is this grade three is this um but I
1:07:29 gave it the first sample and I said you
1:07:32 know there are four uh this is copied
1:07:33 from the other uh conversation and it
1:07:36 gave it a five out of five um which I
1:07:39 would I would say it's close I would say
1:07:41 it's not particularly comprehensive um
1:07:44 and so the reason that you do this is
1:07:46 because as you're generating samples you
1:07:48 want to grade all of them and you just
1:07:50 basically discard the bottom % of the
1:07:53 samples so that you're constantly
1:07:56 refining your your your your synthesized
1:07:58 data set to be the highest quality
1:08:01 information so um next
1:08:04 sample whoops
1:08:07 sorry fat fingered
1:08:10 that all right there we
1:08:15 go okay so that's the next
1:08:18 sample and we're going to watch it grade
1:08:21 it um and again it graded at five out of
1:08:24 five I could see some room for
1:08:25 improvement
1:08:26 um but when you have basically you're
1:08:29 you're basically recursively writing a
1:08:31 textbook of everything that humans know
1:08:34 that is that is what I think strawberry
1:08:36 does is by by using these models that
1:08:39 have already been trained on all text
1:08:41 Data across the entire world they know
1:08:43 everything that humans know whether or
1:08:45 not they know it and whether or not it
1:08:47 has been trained in them well so by
1:08:49 recursively generating a textbook and
1:08:52 you have the expert that is saying okay
1:08:54 I'm writing the the base text and then
1:08:55 then you have the interrogator that's
1:08:57 saying okay tell me what you know about
1:08:58 that and then you have a third model
1:09:00 that is saying okay cool that was well
1:09:02 written or maybe not and you could do
1:09:05 this with I could imagine there might be
1:09:06 several more fine-tune models and you
1:09:08 might have one that is reformatting it
1:09:10 so actually let's do this as an
1:09:12 experiment um so I will create another
1:09:15 chat and I say okay uh your purpose is
1:09:19 to write um uh
1:09:23 textbooks oh that's a that's a good idea
1:09:25 d
1:09:26 yeah I should be watching this at 1.5
1:09:28 speed I I normally watch it 1.5 speed
1:09:30 although I I think right now it's fine
1:09:32 it's fine going a little slow cuz
1:09:35 um like I I want to I want to take this
1:09:39 like I want to I want to take this in
1:09:41 like most stuff I don't give a [ __ ] if I
1:09:43 if I capture everything but the he's
1:09:46 he's a pretty dense dude for you know
1:09:50 for the non Pates of the
1:09:53 world P's like yeah yeah whatever
1:09:56 he he's a child uh textbook sections
1:09:59 based on the data I give you um write
1:10:05 comprehensively um at the highest
1:10:09 intellectual level uh
1:10:12 oops uh domain expert um expound upon
1:10:18 the um Base information I give you with
1:10:23 all the uh knowledge you possess on the
1:10:28 topic so then there might be a fourth
1:10:30 model which is just this is just the
1:10:32 drafter and so let's see what it
1:10:35 does um certainly I'd be happy to blah
1:10:37 blah blah blah blah and then uh actually
1:10:40 it's clawed so it's generating an
1:10:41 artifact so here we have it it's
1:10:44 actually generating the chapter for
1:10:46 us this is how I would if if I had the
1:10:49 money and the time and the uh and the
1:10:53 and the compute time if you asked me
1:10:56 said hey Dave use Laten space activation
1:10:59 and create me a fine-tuning data set
1:11:00 that is going to be like one data set to
1:11:02 rule them all um this is how I would do
1:11:05 it is I would uh use these models to
1:11:08 iteratively basically generate one
1:11:11 textbook to rule the is how I would
1:11:13 approach this and you see how how far
1:11:16 this is going and how fast imagine that
1:11:19 you have a whole bunch yes he is wearing
1:11:21 a Starfleet uniform CL or gp4 or
1:11:24 whatever working in parallel to iter
1:11:27 iteratively unpack literally every
1:11:30 domain of human knowledge that is what I
1:11:32 think open a is doing and that if I'm
1:11:35 correct that is how they are training
1:11:37 gp5 or GPT 6 we're not sure um now and
1:11:41 again like I said the one part that's
1:11:43 missing is I'm not sure how they would
1:11:44 have solved math um unless they did the
1:11:47 same thing where they basically asked a
1:11:49 generator saying hey write out this
1:11:51 latex formula and and figure out how to
1:11:54 how to logically you know unpack this or
1:11:57 something along those lines I might make
1:11:58 another video if I figure that out but
1:12:00 yeah I hope you got a lot out of this um
1:12:02 I could be wrong but as someone who's
1:12:04 been fine-tuning models since
1:12:07 gpt2 this is how I would approach it so
1:12:10 so we'll see and then Mr strawberry
1:12:12 responded to this video and basically
1:12:14 said this is the first guy that's gotten
1:12:16 it right that he got it right now Mr
1:12:18 strawberry has no [ __ ] credibility in
1:12:21 my in my book because it's just a hype
1:12:23 machine but you know
1:12:27 for speculation but but I I mean that's
1:12:31 kind of fascinating so basically what
1:12:33 he's saying
1:12:34 is you have these models that have been
1:12:38 the the way they train them is they take
1:12:39 data and they categorize it and and when
1:12:42 they embed it it creates these semantic
1:12:45 clouds these clouds of meaning where all
1:12:48 these words are in these well organized
1:12:52 clouds but he's basically saying there's
1:12:54 a bunch of [ __ ] in there that's not well
1:12:57 organized and wasn't well tagged but you
1:13:01 can pull it out by just asking for it
1:13:04 and then if you have something that asks
1:13:05 for it really good it'll vomit it out
1:13:07 then you have another thing that looks
1:13:08 at it really good and says here's the
1:13:10 good [ __ ] and just keep the good
1:13:12 [ __ ] um then you end up with this you
1:13:16 know massively valuable um all of human
1:13:20 knowledge in a well organized uh thing
1:13:24 the the link here is this is a YouTube
1:13:26 video the the the channel is David
1:13:29 Shapiro
1:13:31 it's Dave shap I think hang
1:13:36 on yeah
1:13:38 at Dave
1:13:40 shap and the video is
1:13:45 um hello everybody I am awaken how open
1:13:48 AI strawberry works one textbook to rule
1:13:51 them all
1:13:56 so Bonkers
1:14:01 right now what are the implications of
1:14:04 this I don't
1:14:07 know our our gpts are going to get way
1:14:12 smarter and they're going to be able to
1:14:14 do like a lot of the stuff they fail on
1:14:16 now they're going to be able to do and
1:14:19 so what that means is a lot of the tasks
1:14:21 that we do for our job that require
1:14:23 human in the loop may not require human
1:14:25 in the loop so assuming this launches
1:14:29 this fall which it looks like it's going
1:14:31 to um we're going to have we're going to
1:14:34 have some we're going to have some other
1:14:36 [ __ ] to talk about on this channel all
1:14:38 right hope they found the Lost
1:14:40 information of Alexandria we need a
1:14:42 major bump wait you mean we actually
1:14:47 learn here on this AI learning lab I
1:14:50 thought this is date night I know this
1:14:52 is kind of rough um the whole thing we
1:14:55 watched from reminds me of Spock's Brain
1:14:57 Episode 1 Season 3 there you go Andy's
1:15:00 wearing a Star Trek uniform so you
1:15:02 [Laughter]
1:15:04 know yeah he he's he's funny he's very
1:15:07 neuros
1:15:09 spicy he's he's solid he's solid on the
1:15:13 neuro the neuro
1:15:18 Spectrum
1:15:23 uh you don't even have to teach it you
1:15:25 just point it in the right direction and
1:15:28 it uses its
1:15:29 data yeah yeah exactly well and that's
1:15:32 so what they're saying what what what
1:15:36 I've heard about it is that it's very
1:15:38 slow that strawberry is
1:15:40 slow which actually makes sense right if
1:15:43 you say hey go solve some problem go go
1:15:46 tell me everything there is to learn
1:15:48 about it doesn't [ __ ] matter whatever
1:15:51 um genetic engineering in strawberry
1:15:56 growing to to to come up with the more
1:15:59 Street sweet strawberry in the least
1:16:01 amount of you know
1:16:04 iterations it will go off and these
1:16:07 three models will start interacting with
1:16:09 each other so from what I understand it
1:16:11 sort of goes off for five minutes and
1:16:14 just does [ __ ] and then comes back
1:16:17 with you know incredibly wellth thought
1:16:20 out answers I don't know if that's how
1:16:22 it's going to work but but that's a
1:16:24 possibility
1:16:26 um it's crazy and then you know what
1:16:29 it's going to be is it's going to be you
1:16:31 know version one of this is it takes 5
1:16:33 minutes to go think of [ __ ] and then you
1:16:35 know a year later it does it in 30
1:16:38 seconds and a year after that it does it
1:16:40 in 5 seconds and then it's instant right
1:16:45 so so the next three years are going to
1:16:47 be absolutely [ __ ] Bonkers Bonkers I
1:16:50 tell you
1:16:51 Bonkers add a rag Checker and it's even
1:16:54 better I bet
1:16:56 rag stands for retrieval augmented
1:16:58 generation which is basically where you
1:17:00 take your large language model that can
1:17:02 do all the language twirly stuff and
1:17:05 then the rag is is basically a a
1:17:08 database that's got your data in it and
1:17:11 then you interact with the large
1:17:13 language model and it it if you ask it
1:17:16 something about your data it goes in and
1:17:18 pulls actual data out and you know the
1:17:21 large language model presents that to
1:17:23 you so that could be quite powerful yeah
1:17:26 exactly all right um but this is just
1:17:29 synthesizing human data or is it capable
1:17:31 of generating novel ideas I don't that I
1:17:34 don't
1:17:35 know I
1:17:38 mean here's the
1:17:41 thing novel novel ideas are just [ __ ] we
1:17:45 haven't thought of yet how do you think
1:17:48 of [ __ ] we don't we haven't thought of
1:17:51 yet is you take what you know and you
1:17:54 make lot logical connections between
1:17:56 those things and if you make connections
1:17:59 between those
1:18:00 things that are different than other
1:18:03 humans did then that's a novel thought
1:18:06 but it's still based on some education
1:18:09 it's still based on language it's still
1:18:11 B based on what information is in
1:18:15 Einstein's
1:18:17 brain that allowed him to think
1:18:22 about you know the relative speed of or
1:18:26 the relativity of
1:18:27 time compared to the speed of
1:18:33 light it's he's not inventing it out of
1:18:37 nothing he's taking other Concepts and
1:18:39 said oh no one else has made these
1:18:42 connections when I make these
1:18:43 connections here's here's how I think
1:18:45 the world
1:18:46 works so I would
1:18:51 think that if you had that interrogator
1:18:54 interrogate you to say hey go find me
1:18:58 data sets where connections haven't been
1:19:00 made or there's gaps of understanding
1:19:05 Between Worlds of understanding right
1:19:08 here here's how automotive paint
1:19:12 suspends particles to color shift and
1:19:16 here's
1:19:19 um the history of hot dog recipes
1:19:23 somewhere between those two things there
1:19:25 are some connections to be made that
1:19:27 would be a new novel idea color shifting
1:19:29 hot
1:19:31 dogs that's a new novel concept so I
1:19:34 could I could see this thing coming up
1:19:37 with new novel Concepts why not because
1:19:41 novel concepts are not devoid of
1:19:45 data they are derived from the data
1:19:48 that's in whoever's brain that makes the
1:19:50 connections that no one else made right
1:19:55 and and these models are good at making
1:19:58 connections really quickly between
1:20:00 disparate data sets
1:20:02 so but it's only going to keep on
1:20:05 increasing its knowledge it needs to
1:20:07 increase compute time well but the other
1:20:10 thing Danielle is these things are also
1:20:11 going to get more and more efficient
1:20:13 right people like pate and people like
1:20:16 the chip manufacturers and you know
1:20:18 everyone's working to make these things
1:20:20 more efficient these customized chips
1:20:22 like tpus and the grock chip
1:20:25 that's just about doing inference where
1:20:28 where when you ask chat GPD to do its
1:20:30 thing grock is doing that with a very
1:20:33 specialized chip that makes it super
1:20:34 fast and super efficient so I think
1:20:37 you'll as we require more compute you're
1:20:40 also going to see the infrastructure get
1:20:42 more and more
1:20:44 efficient um so you know again I don't I
1:20:47 don't think the I I
1:20:50 think energy and compute and all of that
1:20:53 are absolutely engineering challenges to
1:20:56 be
1:20:57 solved but I don't think we can look at
1:21:00 the world of chips today and assume that
1:21:03 that's what they're going to be 5 years
1:21:04 from now especially when AI starts
1:21:07 getting in and improving chip design and
1:21:10 software design and infrastructure
1:21:12 design right if you have a strawberry
1:21:16 that can say
1:21:17 hey we we have we have compute and
1:21:21 energy demand going like this
1:21:26 but the physical you know infrastructure
1:21:29 is only growing like this so you
1:21:31 strawberry need to close that Gap so we
1:21:34 need to either make [ __ ] more efficient
1:21:37 or we need to be able to make more
1:21:39 energy or more compute Cycles to match
1:21:42 the demand and then AI is going to
1:21:45 [ __ ] solve
1:21:46 that like 2 million new stable crystals
1:21:50 novel Concepts yeah
1:21:52 exactly I think it was only 38 ,000
1:21:55 stable stable crystals it was 2.2
1:21:59 million new crystals that were
1:22:01 discovered that was that was uh Deep
1:22:05 Mind put that out right the the material
1:22:07 science thing yeah 380,000
1:22:11 stable crystals 2.2 million new crystals
1:22:15 or new molecules I don't [ __ ] know
1:22:18 Material Science
1:22:19 [ __ ] [ __ ] that does [ __ ]
1:22:25 and and that's that 2.2 million new um
1:22:30 materials that were
1:22:32 discovered was not
1:22:34 with
1:22:37 strawberry right it's whatever Deep Mind
1:22:39 is doing over there which is pretty
1:22:40 [ __ ] remarkable but once once we have
1:22:43 reasoning engines the the um the
1:22:47 Innovation that that these tools are
1:22:49 going to be capable of it's just going
1:22:50 to be Bonkers and it's what's what's
1:22:52 fascinating to me is
1:22:56 this technology is going to be in our
1:22:58 hands some version of it right it will
1:23:00 be hobbled to make sure that you can't
1:23:03 you know take down the government and
1:23:04 crack C cryptography and all all of the
1:23:07 [ __ ] that you know they had to go talk
1:23:10 to the National you know government
1:23:13 about from a safety perspective it's you
1:23:17 know there's there's going to be some
1:23:19 some guard rails on it
1:23:21 but once this thing's out in the world
1:23:24 and we understand how it works and what
1:23:26 it does the open source Community is
1:23:29 going to go oh we can go build that and
1:23:31 they're going to go build their version
1:23:32 of it and then grock's going to build
1:23:35 their version of it and meta's going to
1:23:36 build their version of it and apple is
1:23:39 probably just going to make me emojis
1:23:41 that talk
1:23:44 better [ __ ]
1:23:47 Apple we're the best no you were the
1:23:50 best
1:23:57 [Laughter]
1:24:01 uh life escape velocity sign me up Lev
1:24:05 baby just survive the next five years
1:24:08 that's what they're telling
1:24:12 us on on my AI office hours today
1:24:18 um we we were talking about that about
1:24:21 um
1:24:26 like in theory you've got I think
1:24:28 Danielle you were talking about this in
1:24:30 theory you've got you know these tools
1:24:32 are going to be able to cure any
1:24:34 diseases and then you've got lobbying
1:24:36 groups like The Pharma lobbying groups
1:24:39 that don't necessarily want to cure
1:24:42 right because sick people are their
1:24:46 customers so if you cure it you don't
1:24:49 sell as many
1:24:51 drugs
1:24:53 um watching the death throws of big
1:24:57 centers of power Financial power
1:25:00 lobbying power political power as these
1:25:03 tools start to disrupt things is going
1:25:05 to be
1:25:07 fascinating it it's crazy crazy crazy
1:25:11 shit's about to about to hit the
1:25:13 fan
1:25:18 um that's why I'm glad we have so much
1:25:21 open sourced
1:25:26 [Music]
1:25:28 all right what are your thoughts what
1:25:30 are your thoughts Janu here for it it's
1:25:34 still probably 5 years away why do they
1:25:36 have to be the scary ones yeah it's
1:25:41 going to get weird for a while I
1:25:43 think we're also watching the death
1:25:45 throws of
1:25:47 copyright that may just
1:25:49 impact
1:25:52 Pharma yeah copyright
1:25:54 the copyright laws were already
1:25:56 struggling to
1:25:58 to have relevance in an in an all
1:26:01 digital world and now that we're in a
1:26:03 generative world where you're generating
1:26:05 original content at essentially infinite
1:26:09 scale and infinite levels of
1:26:13 personalization copyright laws just look
1:26:16 weird they look weird and Antiquated and
1:26:18 I don't I don't have a good solution my
1:26:21 my thought is that you
1:26:25 you pay people on the generation of
1:26:28 officially licensed models so if I take
1:26:30 all my
1:26:33 photography and I create you know here's
1:26:35 Kyle Shannon clouds photography here's
1:26:38 Kyle Shannon street photography and I
1:26:40 make a customized model that's Kyle
1:26:43 Shannon
1:26:45 branded you can sign up for that and say
1:26:48 I want to make images that look like
1:26:50 these photos and every time you generate
1:26:51 an image
1:26:57 it's going to I you know I'll get paid a
1:27:00 tenth of a penny or whatever it is that
1:27:03 system doesn't exist yet but I I think
1:27:05 that's where you have to go
1:27:07 because we artists and copyright holders
1:27:11 can't be chasing down every
1:27:14 seven-year-old that makes Mickey Mouse
1:27:16 farting on Pluto's
1:27:21 head and and you know and that image
1:27:23 ends up you know becoming a huge hit and
1:27:26 everybody loves it and shares it and
1:27:28 then Disney what's Disney gonna do sue
1:27:30 the
1:27:32 seven-year-old like I I just don't I
1:27:34 just don't
1:27:36 think longterm copyright just gets
1:27:38 really
1:27:41 weird so monetized Bots and
1:27:44 gpts takeen to the next level huh I
1:27:47 agree yeah I think
1:27:49 so I think
1:27:51 so it's it's it's weird man
1:27:56 [ __ ]
1:28:04 weird all right what are we doing time
1:28:10 wise what do you guys want to do
1:28:12 anything what did I have in my Friday
1:28:15 night
1:28:17 thread bid Shiro oh there's a new robot
1:28:20 out let's go look at the new robot
1:28:31 I don't know a ton about it but it's an
1:28:34 impressive so it's called
1:28:41 [Music]
1:28:43 Neo so the initial responses to this Neo
1:28:46 robot were will you put the burning
1:28:50 letter wait will you put that letter
1:28:52 burning open
1:28:56 Oh you mean in in
1:29:01 Irregulars do I
1:29:05 have so so they've been they've been
1:29:08 accusing that this is a man in a suit
1:29:10 it's not this is a humanoid robot so
1:29:12 their humanoid robot is putting glasses
1:29:15 away closing the cabinet closing the
1:29:21 dishwasher and then and then and then
1:29:24 there's there's mama mama in the kitchen
1:29:27 mama in the kitchen with the little
1:29:30 bot no Danielle's like no no and then
1:29:35 let's see here's another
1:29:37 [Music]
1:29:41 one
1:29:50 [Music]
1:29:52 yeah yeah
1:29:57 [Music]
1:29:59 wait this one gets creepier watch
1:30:01 [Music]
1:30:07 this come on back lady give me a hug
1:30:16 goodbye oh boy there goes the
1:30:21 neighborhood jennif Jennifer got one of
1:30:24 the those
1:30:25 neobot oh
1:30:27 [Laughter]
1:30:30 Jennifer I've never seen her
1:30:33 [Laughter]
1:30:39 happier so that's the
1:30:42 neobot oh wait here's another one I
1:30:44 didn't see this
1:30:53 one I don't understand this video maybe
1:30:56 that it walks sort of human still looks
1:30:59 weird to
1:31:00 [Music]
1:31:12 me I think that I think that that's the
1:31:15 appropriate that's the appropriate
1:31:18 meme for that
1:31:22 robot and you know they didn't have a
1:31:24 talking there like when so I mean what
1:31:28 are there like aren't there like 15 or
1:31:30 20 humanoid robot companies now like
1:31:33 figure figure is a 2-year-old company
1:31:36 that they're already on version two of
1:31:38 their bot this neobot is some new thing
1:31:41 there was just a Chinese one that came
1:31:43 out like the hardware for these humanoid
1:31:46 robots is quickly quickly
1:31:50 developing the software is going to be
1:31:53 this really
1:31:54 authentic natural voice from open
1:31:57 AI with this insane reasoning brain in
1:32:03 it
1:32:08 so welcome to the
1:32:10 Future yeah yeah get used to it
1:32:18 kids we're gonna be like hey Kyle can we
1:32:21 go back and just make gpts again
1:32:24 remember when we made gpts for the
1:32:26 nonprofits where you could like look up
1:32:29 what the schedule was for the
1:32:32 nonprofit can we go back to that I like
1:32:35 those
1:32:36 [Laughter]
1:32:39 days imagine it walking into your room
1:32:42 your tea sir good day to you sir good
1:32:47 day I mean actually what would be funny
1:32:50 is
1:32:51 like okay so actually so so here's a
1:32:55 here's the
1:32:56 thing I am increasingly increasingly
1:33:00 Cindy Coon's in the in the call tonight
1:33:01 or she was earlier anyway I'm
1:33:05 increasingly feeling
1:33:07 confident that all of us are going to
1:33:10 have
1:33:13 to
1:33:15 confront what it means to create digital
1:33:18 versions of ourselves that we put out in
1:33:20 the
1:33:22 world and and think you should start
1:33:25 doing it now rather than wait until we
1:33:29 have to and so here's what I mean by
1:33:32 that you're going to be able to create
1:33:34 these autonomous agents that do things
1:33:36 on your behalf well what might they do
1:33:39 well you might have one that all it's
1:33:42 going to do is go negotiate contracts
1:33:45 for you so it's going to call up United
1:33:47 Healthcare and it's going to get your
1:33:48 [ __ ] paid and it's going to call up you
1:33:50 know the car dealer and instead of you
1:33:52 having to deal deal with the Ley car
1:33:54 salesman it's going to actually
1:33:56 negotiate you the best deal possible
1:33:58 right and that's going to represent it's
1:34:00 going to be a version of you that that
1:34:03 is like the badass negotiator you might
1:34:06 also have a version of you that's like
1:34:07 the social butterfly that's like you've
1:34:09 got social anxiety and you don't really
1:34:11 feel like meeting people but you're
1:34:13 going to have your social butterfly bot
1:34:16 go out and meet people and introduce you
1:34:18 to interesting people that it it's had a
1:34:20 good conversation with and it might have
1:34:23 even had a good conversation with with
1:34:24 their bot but both of those Bots agree
1:34:27 that you two should really
1:34:29 meet and then when you have these inhome
1:34:33 robots imagine having an inhome robot
1:34:36 that you could actually fine-tune and
1:34:38 train to have your sense of
1:34:40 humor so instead of it instead of me
1:34:43 cracking my stupid jokes like I do on
1:34:45 here I I could say you know tell them
1:34:49 what they've won Bob and then the
1:34:50 computer would be like uh they haven't
1:34:52 won anything this isn't a game show
1:34:54 could we uh get on with mowing the lawn
1:34:58 like why not so that's something that
1:35:02 that I think over the next year I'm
1:35:04 going to start exploring in here is how
1:35:07 we can start to make our own digital
1:35:09 Twins and what it means to make them how
1:35:13 do you use them why would you do it why
1:35:15 would you do it now and not wait Kyle
1:35:18 regulars
1:35:20 neurolink with a personality di oh
1:35:22 imagine neurolink with a personality
1:35:24 dial yeah
1:35:27 exactly exactly where you can you know
1:35:31 you can
1:35:32 basically just with your brain tap into
1:35:36 these
1:35:38 systems and have
1:35:40 them be like you want them to be
1:35:42 personality wise knowledge wise
1:35:46 expertise wise skills wise
1:35:57 I know that the [ __ ] I'm saying is like
1:36:00 sounds [ __ ]
1:36:03 weird and that's because it
1:36:16 is I mean I'm still kind of Blown Away
1:36:30 is that
1:36:32 Korea no it's gen 3 Alpha where's
1:36:36 Korea here it
1:36:39 is so there's our video I I'll put that
1:36:43 video on
1:36:47 uh on whatever you call it on Irregulars
1:36:50 Channel all right let me type something
1:36:54 here um oh wait let me go
1:36:57 to let's see
1:37:00 Kaa real time image
1:37:06 generation oh that's
1:37:14 weird huh that's not working all
1:37:20 right let me go to Deco here this this
1:37:22 seems like it's broken
1:37:30 [Music]
1:37:41 come
1:37:44 on
1:37:47 create we we live in a world right now
1:37:50 where I can type in um a a
1:37:55 tall
1:37:58 building
1:38:00 made of bet of
1:38:05 made of jelly
1:38:09 beans and
1:38:17 Twizzlers and as I'm typing it it's
1:38:20 generating
1:38:26 those
1:38:28 things I I
1:38:30 mean this is [ __ ] mind-blowing the
1:38:33 stuff we're talking about with
1:38:34 strawberry and with these robots and
1:38:38 with the ultimate textbook which is all
1:38:41 of human knowledge in this crystallized
1:38:44 organized access instantly accessible
1:38:47 format um I don't quite know how to
1:38:50 process this [ __ ]
1:38:54 sorry I'm late on date night is this
1:38:56 Korea this is Deco here but Korea Korea
1:38:59 and Deco here both have these realtime
1:39:01 image engines they're using sdxl is the
1:39:05 is the real time image generation
1:39:07 sitting underneath these things and then
1:39:09 and then Kaa has flux which is
1:39:12 the the
1:39:14 high-end like mid Journey competitor
1:39:17 that's open sourced where's my Kaa did I
1:39:20 lose it
1:39:31 hello CLE good day to you
1:39:35 sir so flux here's this generating you
1:39:38 know high quality images so I can just
1:39:40 roll the dice it writes a prompt for me
1:39:43 I hit generate and then it just
1:39:46 quickly
1:39:47 generates high quality
1:39:51 images and it's it like does the first
1:39:54 pass at low quality and then it does
1:39:56 them the next level
1:39:59 up think that one's still
1:40:03 generating but like you know like this
1:40:06 is just we're just in this [ __ ] weird
1:40:08 ass
1:40:10 Bonkers and then I can take any one of
1:40:12 these and turn them into a video you
1:40:19 know it's crazy it's crazy crazy crazy
1:40:23 crazy
1:40:24 well then R2-D2 can be like an extra pet
1:40:28 this is Korea yes so what's nice about
1:40:31 Korea is it's free and they're giving
1:40:33 you access to the flux um image model so
1:40:36 you don't have to roll your own and
1:40:39 start your own GitHub repository and
1:40:41 repet or whatever the
1:40:45 [ __ ] open- Source tool you got to use to
1:40:48 make your [ __ ] work you can just go here
1:40:50 and do it
1:40:55 I got a bad attitude for open source
1:40:57 because it requires too much
1:40:59 concentration on my part so I'm just
1:41:01 waiting for people to build me [ __ ] I
1:41:04 can play with and this is one of
1:41:08 those all right
1:41:11 fantastic thoughts questions
1:41:15 craziness is flux better than mid
1:41:17 Journey um it's better than mid Journey
1:41:19 at text so flux is now really good at
1:41:22 text one of the other things that flux
1:41:24 can do and Murphy is um if you remember
1:41:29 back in stable diffusion days you could
1:41:32 create these things called
1:41:33 luras
1:41:35 L where you're doing sort of custom
1:41:38 fine-tuning on stable diffusion for a
1:41:40 certain image it looks like flux you can
1:41:44 do it with a handful of images and it's
1:41:46 really good so like you could train it
1:41:48 to know what you look like and then just
1:41:51 like dream Booth you could do before but
1:41:53 you could have you in starring in a book
1:41:56 right so for character
1:41:58 consistency um it it looks like the uh
1:42:01 the ability to do much more controlled
1:42:04 Generations um it looks like flux is
1:42:06 really good at it um again I haven't dug
1:42:10 deep into it because you have to go do
1:42:12 it in programming environments right now
1:42:14 and I'm
1:42:15 just I'm lazy people I'm just lazy I'll
1:42:19 tell you what you know what I want to do
1:42:21 you can make money want to make money
1:42:24 with chat JP
1:42:25 T just go hey go start me a business you
1:42:28 big fat loser and it goes off and starts
1:42:30 business and then sends me a check every
1:42:32 month you know what I don't want to
1:42:34 check every month I want to check every
1:42:36 week in fact I want I just want to get a
1:42:40 little Bing on my phone that says
1:42:42 another $500 was put into your account
1:42:44 this
1:42:48 minute that's what I want
1:42:54 in the
1:43:00 meantime Kyle I said earlier we need an
1:43:03 actual C3PO AI robot so it's not scary
1:43:06 yep
1:43:09 yep yeah and it's funny I don't know if
1:43:12 you've been watching sunny on Apple TV
1:43:15 the the movie about the inhome robot
1:43:18 that punch punched a chick
1:43:21 out and and did did some
1:43:24 murders
1:43:26 um it's going to be wild like you know
1:43:30 if you've got buggy if you get an early
1:43:33 version of Microsoft Excel and it's
1:43:36 buggy you [ __ ] up a a couple of math
1:43:39 equations if you get an inhome robot
1:43:43 that's strong enough to lift up the back
1:43:45 of your car and it's a little
1:43:48 buggy it could [ __ ] some [ __ ] up
1:43:58 I actually started a screenplay idea
1:44:00 about this I don't I don't know where I
1:44:01 put it I think I started it on this
1:44:03 channel I've got to go find it cuz it's
1:44:06 a good
1:44:07 one the screen the screenplay is
1:44:10 basically
1:44:12 um in the in the town where the robot
1:44:15 Factory is that's making these new
1:44:18 robots they run a local contest at the
1:44:20 local grocery store that people in town
1:44:25 can get use of one of these robots for a
1:44:27 year and so it's about four people from
1:44:30 the grocery store that each get a robot
1:44:34 and then hi Jinx and
1:44:37 [Laughter]
1:44:42 Sue I want a world we don't need to
1:44:44 depend on money and that's listen if you
1:44:46 if you listen to the Ubi guys and the
1:44:49 the infinite
1:44:50 Prosperity narrative that's part of a GI
1:44:54 and part of part of all this that that's
1:44:57 a very real possibility cam katkin the
1:45:00 thing I can't get my head around is how
1:45:03 do you prevent selfish people from
1:45:06 remaining selfish and wanting to keep
1:45:08 all the profit if there's sort of
1:45:10 infinite Prosperity we live in a world
1:45:13 right now where the top you know 0.1% of
1:45:16 the
1:45:18 0.1% just want to siphon all of that
1:45:21 profit to themselves personally I don't
1:45:23 know how you stop that unless you get
1:45:26 the AI
1:45:28 to take care of some
1:45:32 things and creativity and and
1:45:35 inventiveness are rewarded yeah exactly
1:45:37 that would be awesome Wouldn't It Disney
1:45:38 should make a robot since they already
1:45:41 um own the so the the so the SW imagery
1:45:45 I don't know what SW
1:45:47 is but but friendly imagery I agree with
1:45:50 that and I think we'll see that right
1:45:54 okay here's the
1:45:55 deal why we need artists why we need
1:45:59 liberal arts Majors is all of the robot
1:46:02 companies right
1:46:04 now are populated with Engineers they're
1:46:08 not populated with you know DreamWorks
1:46:12 imagineers so Disney and Disney by the
1:46:16 way has is making Bots like this right
1:46:19 have you seen that Superman one or the
1:46:21 Spider-Man one that that's a literal
1:46:25 it's a bot a humanoid bot that looks
1:46:28 like Spider-Man that swings on a rope
1:46:30 and then does a tum somersaults through
1:46:33 the air and lands in a net like Disney's
1:46:37 going to be good at this [ __ ] but you
1:46:39 need Disney imagineers working for some
1:46:41 of these humanoid robot companies to
1:46:43 give them friendly faces to give them
1:46:45 friendly
1:46:47 forms to give them senses of humor to
1:46:49 give them
1:46:51 storytelling right if it's just about I
1:46:53 can do a task for you yeah if it's like
1:46:57 hey we got some cool [ __ ] going on
1:47:03 here I mean isn't it already fascinating
1:47:06 I mean we already
1:47:08 have and this isn't even the good stuff
1:47:11 like we already
1:47:18 have oh Star Wars yeah hey pie how's it
1:47:21 going
1:47:26 hey there everything's going great
1:47:28 thanks for asking I'm here to chat
1:47:29 entertain and help out with whatever you
1:47:31 need are you a Star Wars fan oh no I was
1:47:34 just talking to someone else you
1:47:35 remember my name
1:47:39 right if you've created an account with
1:47:42 me I should know your name uh but if you
1:47:44 haven't registered I'm afraid I won't
1:47:45 know it have you created see that she
1:47:47 does I just I guess I'm not signed in
1:47:50 but we've already got that and like
1:47:52 that's the shitty voice like that shitty
1:47:54 voice to text voice the advanced voice
1:47:58 that open AI has announced is one that's
1:48:02 got personality and it can hear our
1:48:04 inflection and it can mimic that and
1:48:06 respond with emotional tone and then put
1:48:08 that in a humanoid robot where imagine
1:48:12 going on a 6-hour road trip that you've
1:48:14 got to go on your
1:48:16 own and you take your [ __ ] humanoid
1:48:18 robot and you have just [ __ ] wacky
1:48:21 conversations with your robot and then
1:48:24 you get to the gas station and you're
1:48:26 like I don't feel like just run in and
1:48:28 get me Twizzlers while I pump gas and
1:48:30 your humanoid robot runs in this gas
1:48:33 station you
1:48:35 know and if and if you're a
1:48:38 hacker you're you're like hey go get me
1:48:40 a pack of Twizzlers and by the way you
1:48:42 know Rob The
1:48:45 Joint here's a
1:48:49 gun AI will hopefully democratize
1:48:53 resources that's that's so what I'm
1:48:55 hoping cam people hoard money often
1:48:58 because they're
1:49:00 afraid it's all they know yeah yeah
1:49:03 exactly yeah if
1:49:05 if if AI can democratize resources I
1:49:08 mean I think I think a big thing we're
1:49:10 going to start to get is
1:49:13 accessibility um boosts where people
1:49:17 that might not currently be able to
1:49:19 contribute to society all of a sudden
1:49:21 can um and you know a lot of those
1:49:24 people they might not be able to speak
1:49:25 or they might have neurological things
1:49:27 or injuries or you know um you know
1:49:30 blindness or deafness things like that
1:49:32 and they might not be able to contribute
1:49:33 in a certain way that all of a sudden
1:49:35 can so I think we're going to see sort
1:49:37 of massive upswings in
1:49:40 contribution
1:49:42 um and then and then yeah I think it I
1:49:44 think it goes up from there I you know
1:49:46 I'm quite bullish on the future but man
1:49:49 is it going to get weird
1:49:51 fast it's it just like and again we we
1:49:57 are the
1:50:00 people we're we're the only
1:50:03 people who will know what it was like
1:50:07 before AI before generative AI sort of
1:50:11 blew onto the scene like it
1:50:15 has we'll be able to look back and say
1:50:18 remember when it was just cute imperfect
1:50:21 images remember when the videos were
1:50:23 only 4 seconds long and they look
1:50:31 weird and we're going to be living in
1:50:33 that her kind of future where people
1:50:38 that sort of grow up into that are going
1:50:40 to think that's normal and we're going
1:50:42 to be like it was never like
1:50:45 that you're dating a robot you don't
1:50:48 think that's weird no why okay
1:50:53 [Laughter]
1:50:56 I'm still getting used to
1:50:58 [Laughter]
1:51:06 swiping Sally's happier than she's ever
1:51:09 [Laughter]
1:51:15 been I don't know what it is her
1:51:18 casseroles used to suck you think the
1:51:20 robots making the casseroles
1:51:22 [Laughter]
1:51:24 wonder what else that robot's
1:51:26 doing the scandals the scandals in
1:51:31 Suburbia when when Jennifer gets her
1:51:34 [Laughter]
1:51:40 robot and all the other housewives are
1:51:44 jealous because they've got a [ __ ]
1:51:47 human piece of [ __ ] human that [ __ ]
1:51:50 about mowing the lawn doesn't talk to
1:51:52 them
1:51:55 and Jennifer's down the block having
1:51:57 romantic
1:52:03 dances oh man we got some movies to
1:52:06 write here
1:52:07 [Laughter]
1:52:11 people sorry I'm gonna lose
1:52:15 it real real robots of New York
1:52:22 exactly real robots of
1:52:24 [Laughter]
1:52:27 Dallas actually you know what a reality
1:52:32 show a reality show of neighbors where
1:52:35 some neighbors have robots and some
1:52:37 neighbors don't that would actually be a
1:52:39 [ __ ] fascinating show right because
1:52:43 you could see how the neighbors without
1:52:45 robots see it from the outside and they
1:52:48 judge
1:52:50 it and then you could see it from the
1:52:52 inside what it's like and like what the
1:52:54 issues are and like like nothing's going
1:52:56 to be perfect and these things are going
1:52:58 to be janky right like you know they're
1:53:00 going to trip and fall and break a
1:53:02 coffee
1:53:03 [Laughter]
1:53:07 table but they're also going to mow the
1:53:09 lawn for you you
1:53:13 know and Neil's just going to be smoking
1:53:16 a cigar in his lawn chair while this
1:53:18 robot mows the lawn and then the other
1:53:21 men that are out there mowing their
1:53:22 lawns like [ __ ] Neil with his [ __ ]
1:53:28 robot of course Neil has a BMW robot
1:53:31 [ __ ]
1:53:33 [Laughter]
1:53:40 douche oh
1:53:42 man oh well I should probably let y'all
1:53:45 get get get get on to your weekend it's
1:53:48 a long weekend I'm glad you hang hang
1:53:50 out um I forgot it's Labor Day weekend
1:53:53 so the fact that we have 60 people in
1:53:55 here is absolutely stunning to me um
1:53:58 thank you all for hanging out
1:54:03 um yeah we're we're in we're in wacky
1:54:06 times aren't we we are in wacky
1:54:18 times this is different than this
1:54:24 all
1:54:26 right these are very
1:54:36 cool still meeting Monday yeah I'll
1:54:40 probably be here Monday I don't I'm not
1:54:41 going anywhere I'm not doing
1:54:44 anything thank you for being here I
1:54:46 always enjoy listening thank you very
1:54:48 much manifesting Monica I appreciate
1:54:50 that it's very nice of you to say
1:54:53 um I didn't really do
1:54:56 a an ad for the
1:55:00 salon but if you are an AI person if
1:55:04 you're like hey I love this AI [ __ ] I
1:55:05 want to know more about this I want to
1:55:07 find people who are into this this
1:55:09 community the
1:55:10 salon is a community that I started the
1:55:13 week after chat GPT came out a lot of
1:55:16 the people that hang out here are
1:55:18 leaders in this community doesn't cost
1:55:21 anything it's really remarkable
1:55:22 community so go hang out there this
1:55:24 community is also the one co-sponsoring
1:55:26 with Google for small business this fly
1:55:29 into New York where we're going to go
1:55:30 talk about hey AI is
1:55:34 good don't [ __ ] it
1:55:38 up um so go check that out that's our
1:55:41 website and then if you scroll down
1:55:43 there's a button that says join our
1:55:44 community so go join the community um
1:55:47 homework for the
1:55:51 weekend um
1:56:08 yes don't pay attention to anything
1:56:10 about
1:56:14 AI just go have fun just go be in the
1:56:18 world be with people
1:56:20 connect that's the homework because
1:56:26 I I sense I sense that the next 3 months
1:56:29 are going to get really
1:56:30 weird and really
1:56:33 intense and and some of it's going to be
1:56:35 like weird and exciting and good and
1:56:37 some of it's just going to be like holy
1:56:38 [ __ ] how do we [ __ ] process this it's
1:56:41 so funny when when um when chat GPT
1:56:44 first launched the initial interviews
1:56:46 with Sam
1:56:48 mman he talked a lot
1:56:51 about oh thank you very much that's very
1:56:54 very very generous I I appreciate that
1:56:57 um he talked a lot
1:57:00 about wanting
1:57:03 to not launch too much too quick because
1:57:07 it would freak people
1:57:08 out and as a as a
1:57:12 tech fan and Optimist and someone who
1:57:15 loves kind of living on the front edge
1:57:17 of this stuff my init my my initial
1:57:20 response to that was oh you [ __ ]
1:57:23 [ __ ] just give us the [ __ ] give us the
1:57:30 toys as I've watched over this year the
1:57:32 tools get better and better and
1:57:35 better they were they
1:57:38 were the initial things we were playing
1:57:41 with were toys our
1:57:44 toys right I can make a 4C video and if
1:57:48 if you're talented at video storytelling
1:57:50 then you can put that in an editing
1:57:52 program and go write your music and you
1:57:54 can do good storytelling but the the
1:57:57 videos the video makers that we have
1:57:59 right now are kind of toys and the image
1:58:01 generating tools are kind of toys but
1:58:03 the image generating tools now that
1:58:05 they're really good at text are starting
1:58:07 to turn into things that are more than
1:58:10 toys and especially when you get someone
1:58:13 like Elon Musk that puts flux into grock
1:58:17 and takes off the guard rails for making
1:58:21 political figures and making you know
1:58:26 Brands um all of a sudden you start to
1:58:29 go oh these are kind of more than toys
1:58:31 like I'm I'm now starting to grock so to
1:58:34 speak why Sam mman was like we've got to
1:58:37 roll this out slowly get people used to
1:58:40 the power of what it's like to have a
1:58:43 large language model where you can ask
1:58:45 it a question and it just answers it
1:58:47 even if it hallucinates and does it
1:58:51 wrong it's taken me a year 18 months to
1:58:54 really even get my head around what does
1:58:56 that even
1:58:57 mean and so I think the next 3 months we
1:59:03 we're going to be confronted with
1:59:05 technology that is very hard for us to
1:59:10 comprehend and it's going to be harder
1:59:12 for us to
1:59:16 actually use and assimilate and and
1:59:24 understand what our work is and what our
1:59:27 value is I connect with people via VR
1:59:31 does that
1:59:32 count that's good actually it
1:59:34 does
1:59:38 um so so I think this weekend just take
1:59:41 just take some time to like don't worry
1:59:45 about like this Tech is coming it's
1:59:47 going to be here so I think have a nice
1:59:50 weekend like that's that's that's my
1:59:52 homework work is is just [ __ ] have a
1:59:55 nice
1:59:57 weekend and I mean maybe think just
2:00:00 philosophically
2:00:05 about
2:00:06 if if the machines can
2:00:09 do the stuff in your life that is the
2:00:13 repetitive mind-numbing horrible tasks
2:00:17 that you don't like
2:00:19 doing and you could just do the stuff
2:00:21 that made you happy happy what would
2:00:23 that be what would be what would be your
2:00:25 life if the repetitive
2:00:27 low-level [ __ ] storm of Life were
2:00:32 handled because I think designing that
2:00:35 life designing who you want to be if all
2:00:37 that [ __ ] were handled that's going to
2:00:39 be our job in the next three
2:00:41 years so maybe think about that what
2:00:45 could my life be let's get the channel
2:00:48 to 30 likes I assume you mean 30,000
2:00:52 we surpassed 30 likes but we are at
2:00:55 28,000 we're going to hit that um okay
2:00:58 so go to the salon check it out join it
2:01:03 unless you use Grill dad or Sin City
2:01:05 sipper to help you with your barbecue
2:01:07 yeah
2:01:08 exactly
2:01:09 um yeah I think do nothing I think do
2:01:13 nothing yeah it's good go find yourself
2:01:17 a pink bow I think we need more pink
2:01:19 bows in our life
2:01:22 and let me just tell you something all
2:01:25 you people with you're [ __ ] wearing
2:01:27 it wrong with your with your pink bow
2:01:33 rage there are many ways we can wear
2:01:35 pink
2:01:36 bows there are many people that can wear
2:01:41 them and how I wear a pink bow might be
2:01:44 different than how you wear a pink bow
2:01:46 and and we we cast a wide net here we
2:01:49 have a large
2:01:50 tent that Embraces is all the pink bow
2:01:53 wearing um yeah go get yourself a pink
2:01:56 bow um all right I'm going to get out of
2:01:59 here so oh try character a AI if you
2:02:03 want your feelings work for homework
2:02:06 yeah character. a Danielle says it's
2:02:08 it's it's a mean-spirited service but
2:02:10 it's apparently really
2:02:12 popular yeah the the that that's another
2:02:15 thing that we're going to confront is
2:02:17 that these AI systems already have
2:02:20 personalities and we're going to thank
2:02:22 you for 30,000 likes that's amazing um
2:02:25 we're going to start to discover the
2:02:27 tools that we like that have
2:02:28 personalities we like and some that have
2:02:30 personalities we hate and that's just
2:02:32 like dealing with people so it's about
2:02:35 to get weird hang on to your butts all
2:02:37 right um let's do uh oh very nice
2:02:41 Winston thank you so much for the
2:02:43 Cornucopia fruit there fantastic digital
2:02:46 twin and rag sharing helped me a lot
2:02:49 tonight oh good fantastic spin B3 and
2:02:52 that's we're going to do I I'm going to
2:02:55 I'm going to find ways to dig into that
2:02:57 more on this channel because I I think
2:02:59 I'm increasingly feeling like that's
2:03:01 really important that us really
2:03:03 understanding who we are what data
2:03:06 represents us and us being in charge of
2:03:08 how we put that into the world that
2:03:11 feels important to me tell your friends
2:03:13 AI Sal AI Salon presents Tuesday night 3
2:03:17 p or 5:00 P p.m. mountain yeah this
2:03:19 coming
2:03:20 Tuesday the salon we have our twice
2:03:24 monthly meetings the first one is called
2:03:25 AI Salon presents and we have uh a guest
2:03:30 speaker come in and speak so that's this
2:03:32 coming Tuesday all right so be there for
2:03:34 that uh September 11th I'm going to be
2:03:37 in Washington
2:03:38 DC representing the the salon and all of
2:03:42 us so hopefully
2:03:45 I maybe I should wear my pink bow to
2:03:48 Washington
2:03:53 I don't think they get
2:03:56 it Mr Shannon Mr Shannon I understand
2:03:59 you're a thought leader in the in the
2:04:01 artificial intelligence space uh we have
2:04:03 Unearthed a video uh of it appears to be
2:04:07 you perhaps it's a deep fake uh wearing
2:04:09 a a large pink bow on your forehead uh
2:04:12 it it doesn't look like you're actually
2:04:14 wearing the headband correctly uh which
2:04:17 is odd why you would choose to do it
2:04:18 that way anyway uh Mr Shannon could you
2:04:21 explain is is is this indeed you wearing
2:04:24 this pink bow and what was the purpose
2:04:26 of the bow and what does that have to do
2:04:27 with artificial intelligence and and
2:04:29 regulations we should consider here in
2:04:31 uh in Washington DC and for for the
2:04:33 nation as a
2:04:35 whole Mi M Mr
2:04:38 Shannon sir sir could could someone
2:04:40 please wake up Mr
2:04:42 Shannon hi Mr
2:04:44 Shannon so that's going to be
2:04:47 fun are you still collecting stories for
2:04:49 your trip um yeah if if um one of the
2:04:53 things that I'm going to talk about when
2:04:55 I go to DC is stories of impact so if AI
2:05:00 has impacted your life in a positive way
2:05:06 um I've I'm trying to think where I've
2:05:08 got the announcement I've got an
2:05:09 announcement in AI Salon announcements
2:05:13 um but you can just DM me if you've got
2:05:15 if you've got some interesting story for
2:05:17 how AI has impacted your life in a
2:05:19 positive way just DM me either here
2:05:22 or on the salon and uh and we'll
2:05:26 coordinate cuz yeah I'm collecting
2:05:27 stories for that so thank you for
2:05:29 reminding me of that we learned a lot
2:05:31 today thank you thank you side hustle
2:05:32 Mimi I appreciate it um all right I am
2:05:35 out of here so who's in the house
2:05:37 tonight Amelio's wife was here early
2:05:39 Auntie bubble's in the house uh Tobias
2:05:42 is in the house wear your AI Salon
2:05:44 t-shirt yes next Tuesday wear your AI
2:05:46 Salon t-shirt Oh you mean me wear it
2:05:47 when I go to DC it's a good point um
2:05:51 yeah maybe I'll put it underneath like a
2:05:53 a nice suit jacket or something like
2:05:55 that all right it's not a bad idea
2:05:58 um okay Vicki you can repin okay so go
2:06:02 to AI Salon if you go here go to the
2:06:04 salon. go to the mighty networks's
2:06:07 Community there's an announcements
2:06:09 Channel Vicky's going to repin my
2:06:11 request for stories so go to salon
2:06:14 announcements and you'll see my request
2:06:16 for stories about AI impact all right
2:06:19 side hustle mimies in the house Lord
2:06:21 digital God thank you so much for the
2:06:22 Roses very very uh generous of you as
2:06:25 always um Winston great to see you
2:06:28 Brandon thank you papa JJ cam katkin
2:06:30 from Cleveland good to see you um Dr Jay
2:06:34 in the house who else we got Wolf Man
2:06:37 Clint Be not Afraid Danielle good to see
2:06:46 you two block Tom great night thank you
2:06:49 thank you very much for the kind words I
2:06:51 appreciate that Marina good to see you
2:06:54 wear your salon T-shirt with your
2:06:58 bow I think I should take this to
2:07:00 Washington you can make money with
2:07:05 J um put on your calendars the Oprah Sam
2:07:09 Alman interview cuz we're going to do
2:07:12 that we're going to do commentary to
2:07:13 that here um that's September 12th the
2:07:17 night after I get back from
2:07:19 DC uh
2:07:24 Joey's in the house
2:07:27 [Music]
2:07:29 Joey bunny cat bunny cat in the house
2:07:33 Jew live laugh love wait Jew oh Jewel
2:07:36 live laugh love wear your smart pants
2:07:39 too
2:07:41 sir if we work for British Airways we're
2:07:44 expected to wear smart pants exactly I
2:07:48 will wear my smart pants in fact I got
2:07:50 some smart pants for my DC trip so I
2:07:54 don't look like a dumb Dum hey hello
2:07:57 Senator my name is Kyle Shannon I want
2:07:59 to tell you about artificial
2:08:00 intelligence otherwise known as machine
2:08:02 learning or or vice versa if you know
2:08:04 what I mean so some one of them is a
2:08:05 subset of the other anyway sir uh AI is
2:08:08 just awesome so I think you
2:08:10 should make it not bad don't mess it up
2:08:14 all
2:08:18 right Kyle lives in a pineapple under DC
2:08:21 nice verifier like it very nice very
2:08:24 nice um all right good I'm out of here
2:08:28 people have a great Friday night have a
2:08:30 great weekend have a great Labor Day and
2:08:33 yeah just unplug don't pay attention to
2:08:35 any of this [ __ ] I will see you probably
2:08:37 Monday night uh if not Salon Tuesday all
2:08:40 right peace everybody bye