Midjourney v6, Altman 'Age Reversal' and Gemini 2 - Christmas Edition
00:00:00.000 | It's almost Christmas in the world of AI, and what else could that mean, other than
00:00:05.080 | another step forward in the march toward lifelike text image, rumors of an imminent extra ten
00:00:11.520 | years of health span funded by Sam Altman, and new state-of-the-art models being trained
00:00:17.480 | behind the scenes, presumably by elves.
00:00:20.440 | But let's start briefly with the new Mid Journey version 6 announced in the last few
00:00:25.920 | days.
00:00:26.920 | I've been playing about with it for far too many hours, and the main difference that
00:00:30.120 | I can notice with version 5.2 is that it adheres to your prompts far more sensitively.
00:00:36.320 | I could give you dozens of examples, but take this Roman Triumphal Arch.
00:00:40.800 | I recently saw one of these in person, but the problem with Mid Journey version 5 is
00:00:45.440 | that it might capture the main topic of your prompt, but misses out on key relational details,
00:00:51.480 | like something being bigger than, next to, or on top of.
00:00:54.840 | Version 5 also misses out on parts of the prompt, like here the stream running through
00:00:59.960 | the center of the arch.
00:01:02.120 | With version 6, at least one of the outputs will likely adhere to what you're asking
00:01:07.000 | for.
00:01:08.000 | Of course, the difference with even earlier versions of Mid Journey are even more stark.
00:01:12.960 | Look at the progress just since 2022, and that's before I even upscale the v6 output.
00:01:20.080 | We really are getting tantalizingly close to photorealism.
00:01:24.520 | But now let me show you the upscaled version of that image using Magnific.
00:01:28.880 | Talk about getting pretty close to photorealism.
00:01:32.040 | And no, this is not sponsored, but honestly, I found it really fun to play about with Magnific,
00:01:36.640 | albeit it's a little bit pricey.
00:01:38.480 | Basically because all you have to do is drag and drop your Dali 3 or Mid Journey v6 image
00:01:43.380 | into the top left, and then basically scroll down.
00:01:46.160 | I don't even think you have to change any of the settings to get amazing results, and
00:01:49.560 | then you just go to upscale, and then it's done.
00:01:52.040 | It takes less than a minute and it turns the AI slightly glossy images into something much
00:01:57.920 | more lifelike, photorealistic.
00:02:00.440 | But while we're on the topic of tips, here's another one that I found really useful.
00:02:04.760 | Instead of adding words like photorealism or 4k, 8k to your Mid Journey prompt, try
00:02:10.120 | --style raw, and you'll notice that the images go from, again, a kind of glossy, clearly
00:02:16.960 | AI style image, and they become much more photorealistic already, before you even try
00:02:22.240 | Magnific.
00:02:23.240 | Dali 3 is still, for me, a little bit better at text, but they try to avoid photorealism.
00:02:29.760 | My guess is because OpenAI want people to know if an image is AI generated, and is of
00:02:35.240 | course watermarking its images.
00:02:37.360 | Now yes, I did experiment with putting Dali 3 images through Magnific, and it kind of
00:02:42.320 | works.
00:02:43.320 | So I put Dali 3 Roman Arch over a quiet stream, and then this is after one round of upscaling.
00:02:50.560 | And of course, with the magic of the new Suno AI web app, you can turn your prompt into
00:02:55.160 | a backing track.
00:03:20.720 | Then throw in a bit of runway ML to get some movement.
00:03:25.680 | And at this point, it doesn't take much imagination to see where the industry is going.
00:03:32.120 | Photorealism has almost reached text to image, and it's coming for text to video.
00:03:36.880 | But there are a few more things to get to before the end of this Christmas episode of
00:03:42.480 | AI Explained.
00:03:43.680 | Sam Altman, the CEO of OpenAI, yesterday came out with this list of the things he wished
00:03:49.000 | someone had told him.
00:03:50.400 | I'm just going to pick up on three of them as they relate to new stories that have broken
00:03:54.520 | in the last few days.
00:03:56.120 | The first one being how he mentioned concentrating his resources on a small number of high conviction
00:04:01.840 | bets.
00:04:02.840 | And it seems on this front, Sam Altman is putting his money where his mouth is.
00:04:06.920 | He has personally invested $180 million on healthy lifespan extension.
00:04:13.000 | The life extension company, Retro, is quick to point out that that's not about living
00:04:16.800 | forever.
00:04:17.800 | It's about adding 10 good years to your life.
00:04:19.960 | It was a fascinating read in Bloomberg, and of course, I can't get to all the details,
00:04:24.800 | but the company has just one investor, that's Sam Altman, CEO of OpenAI, who's invested
00:04:29.680 | that $180 million.
00:04:31.560 | Now I don't know how I missed this, but back in March, MIT Technology Review reported that
00:04:36.080 | Sam Altman had emptied his bank account betting on energy and extended lifespan.
00:04:41.400 | He said, "It's a lot.
00:04:42.480 | I basically just took all my liquid net worth and put it into these two companies."
00:04:46.640 | Retro is betting on five different tracks, thanks to Sam Altman being willing to do something
00:04:51.320 | different and throw lots of money at a bunch of things in parallel.
00:04:55.280 | Their main bet is on partial cell reprogramming, and I read several papers to understand it
00:05:00.560 | before making this video, but I still can't fully summarize it.
00:05:04.120 | Nevertheless, here's my best attempt.
00:05:05.400 | In a nutshell, what it's doing is reprogramming mature cells by introducing a set of what's
00:05:10.440 | called transcription factor genes, basically full of special instructions, which turn those
00:05:15.920 | mature cells into something more youthful without changing their type or function or
00:05:20.900 | making them stem cells.
00:05:22.240 | A bit like changing the code of a program, not the function.
00:05:25.760 | Now, any biologists watching this can let me know if that was a good summary, but the
00:05:29.680 | main purpose of including this article in this video was this paragraph.
00:05:35.040 | Retro, the company, and its peers really do think that this time is different.
00:05:40.000 | Many researchers in the field contend that the science behind cell reprogramming in particular
00:05:44.800 | has been solved and that therapies are now an engineering problem.
00:05:48.640 | They see full-on age reversal as not only achievable, but also perhaps imminent.
00:05:54.680 | Sam Ullman thinks what's needed is an open AI type effort in longevity.
00:05:59.320 | Now, I'm old enough to have been reading longevity research for over a decade, so I
00:06:04.400 | have heard claims like this before.
00:06:06.200 | In fact, I remember reading an article about Calico Company being formed over 10 years
00:06:10.680 | ago.
00:06:11.680 | That was funded in part by Larry Page.
00:06:13.360 | They've been working on increasing healthy lifespan for over a decade, and I haven't
00:06:17.960 | seen much literature on the success of their work.
00:06:21.680 | And even the Bloomberg article reports that Sam Ullman will not be able to fund this all
00:06:26.440 | the way, according to the Retro co-founder.
00:06:29.280 | Getting to the final products will require more investors and going public at some point.
00:06:33.480 | But I haven't seen language like things being imminent before.
00:06:37.120 | And even the PubMed article that I read while researching for this video used language that
00:06:41.760 | was pretty interesting.
00:06:43.180 | In the final sentence, it said, "It remains to be seen whether cycles of epigenetic rejuvenation
00:06:48.480 | by age reprogramming could confer biological immortality in mammals, including humans."
00:06:55.400 | But as many of you may know, it's not just Sam Ullman in big tech who's focusing on
00:06:59.600 | longevity and also, by the way, personally taking metformin to boost his own health span.
00:07:04.600 | There's also Jeff Bezos, who has pumped billions into Altos Labs.
00:07:09.120 | One of the leaders of that lab is also using some dramatic language.
00:07:13.460 | He said last year that within two decades, we will be able to prevent diseases and ageing.
00:07:19.060 | Now, of course, the point of this Christmas video is not to go deep into the science and
00:07:23.300 | test out all of these theories.
00:07:24.960 | It's more to describe the shift in language that I've been seeing in recent months and
00:07:29.700 | years.
00:07:30.700 | As I say, I've been following the literature for more than a decade, and it's the first
00:07:34.060 | time I've heard things like imminence being seriously spoken of.
00:07:37.740 | And even setting aside the possibility that something like AGI could speed up this research,
00:07:43.060 | the 2030s are more and more being spoken of as the decade that could be revolutionary
00:07:48.500 | for healthspan.
00:07:49.500 | That's healthy lifespan as opposed to just living longer in your hundreds, decrepit.
00:07:54.060 | Of course, not everyone shares the excitement of living indefinitely.
00:07:57.060 | Here's one OpenAI employee who might have an interesting chat with Sam Ullman.
00:08:01.340 | He said, "In the pursuit of immortality, we are losing the benefits of Darwinism.
00:08:06.200 | Things dying and new things being born brings health to the ecosystem.
00:08:10.020 | I feel that for the misguided quest of stability, we often keep things alive longer than they
00:08:15.580 | should be."
00:08:16.580 | And it does seem like there is a big split in big tech about whether the goal of living
00:08:21.100 | for longer is one that we should pursue in the first place.
00:08:24.420 | I personally wonder what Santa Claus has to say, who is himself presumably immortal.
00:08:29.460 | Anyway, back to Sam Ullman who said, "Fight BS and bureaucracy every time you see it and
00:08:34.860 | get other people to fight it too.
00:08:36.780 | Kind of ironic in the light of what happened with the OpenAI board, but nevertheless, do
00:08:40.980 | not let the org chart get in the way of people working productively together."
00:08:44.800 | And there is one more point on this list which nicely dovetails with the next news item that
00:08:49.820 | I wanted to mention.
00:08:50.820 | Toward the end, Sam Ullman wrote this, "Compounding exponentials are magic.
00:08:55.740 | In particular, you really want to build a business that gets a compounding advantage
00:09:00.540 | with scale."
00:09:01.540 | And that was straight after saying that scale often has surprising emergent properties.
00:09:05.900 | The question of course though, is whether it's OpenAI who will get that compounding
00:09:10.660 | advantage of scale, or maybe is it going to be that old titan, Google?
00:09:15.660 | Just this week, we learned from the information that according to one person familiar with
00:09:19.460 | the matter, Google is already training its next big model, Gemini 2.
00:09:23.740 | So they're training Gemini 2 before they've even released Gemini 1 Ultra.
00:09:28.660 | But how does Gemini 2 relate to compounding scale?
00:09:32.200 | Well, according to a report from Semianalysis, the compute power available to Google is set
00:09:37.500 | to far outstrip OpenAI in the near future.
00:09:41.020 | As they put it, "Even when giving OpenAI every benefit of the doubt, Google's compute
00:09:45.260 | capabilities make everyone else look silly."
00:09:47.820 | Google will quite literally have more TPUV5s, that's Tensor Processing Units that they're
00:09:52.860 | using to train Gemini 2, than OpenAI, Meta, CoreWeave, Oracle and Amazon will have
00:09:58.500 | GPUs combined.
00:10:00.100 | They go on, "Of course, on performance per chip, there is a significant deficit for TPUV5
00:10:05.260 | versus H100.
00:10:06.260 | But even when shaking that out, OpenAI's compute is a fraction of Google's, and that
00:10:11.380 | the incredible TPUV5 build-out is going to drive significantly more training and inference
00:10:16.980 | capacity than anyone else on the planet."
00:10:19.980 | So it seems certain that whether we're talking about Gemini 2, GPT-5, Healthspan or Photorealism,
00:10:26.300 | 2024 looks set to be interesting.
00:10:29.300 | One thing we can all agree on though, is that working with great people is one of the best
00:10:34.300 | parts of life.
00:10:35.300 | And for me, that's all of you who watch these videos, even if you don't like or
00:10:39.500 | comment.
00:10:40.500 | I really do appreciate every view, every comment in this Christmas season and all year.
00:10:44.580 | A special thanks of course to all of the AI Insiders and Legendary Supporters who keep
00:10:49.780 | the channel going on Patreon.
00:10:51.940 | A massive and enduring thanks to all of you.
00:10:55.380 | So from me, thanks for watching all the way to the end.
00:10:58.500 | Have a very Merry Christmas and a wonderful day.