Sora is Out, But is it a Distraction?

00:00:00.000 | After a 10 month wait, OpenAI have released Sora to paying users.

00:00:04.800 | With just a prompt, it can generate videos of up to 20 seconds in lower resolutions and

00:00:10.300 | 10 seconds at 1080p if you can fork out $200 a month.

00:00:15.800 | I've tested it and read the system card.

00:00:18.220 | The user interface is really quite beautiful, even if the videos themselves operate under

00:00:23.640 | entirely new rules of physics.

00:00:25.640 | But I just can't help wondering if OpenAI want us to focus on releases like Sora rather

00:00:33.740 | than some quietly broken promises.

00:00:36.400 | First things first, Sora is available in almost every country aside from those of the EU and

00:00:42.200 | the UK.

00:00:43.200 | And for those of you who've noticed my accent, you might be wondering how am I using it and

00:00:46.960 | I'm using a VPN.

00:00:47.960 | No, don't worry, I'm not going to do an awkward segue to a VPN sponsorship.

00:00:52.240 | As mentioned at the start, you do need to be a paying user of ChachiBT to get Sora and

00:00:57.120 | the standard $20 tier only gets 1000 credits and the meaning of that we'll see in just

00:01:02.280 | a second.

00:01:03.280 | ChachiBT+ is also capped at 720p resolution, not 1080 and just 5 seconds of duration.

00:01:09.760 | Pro at $200 a month gets you 10,000 credits, but you also get to download without the watermark.

00:01:16.240 | Now you can see how quickly you would use up those 1000 credits from the $20 tier.

00:01:21.960 | Look, 720p, 5 seconds, that's 60 credits.

00:01:25.560 | Dare I say it, even on the $200 tier, you can eat through your credits quite quickly.

00:01:30.600 | In fact, just prepping for this video, I've probably used up like 80% of my allowance.

00:01:34.840 | The allowances, by the way, don't roll over to the next month.

00:01:37.480 | By now, I'm sure that almost everyone watching this video has seen at least a dozen other

00:01:42.260 | videos from other creators, so I'm going to keep this one short.

00:01:45.200 | The first is just a standard generation and it has remembered to pick out the shard, but

00:01:51.160 | I wouldn't say this shadow is being cast over the shard.

00:01:54.600 | Next for the budding content creators out there, how about a futuristic YouTube intro?

00:01:59.440 | I'm not going to use it for the channel, but I think that's pretty cool.

00:02:02.640 | How about the storyboard feature?

00:02:04.720 | And honestly, I think the user interface is pretty amazing.

00:02:07.720 | It's very Apple-like, very sleek and clean.

00:02:10.440 | But of course, the central problem with all Sora generations, indeed, all of generative

00:02:15.080 | AI, including ChachiBT, is that it doesn't really follow physics.

00:02:19.440 | It hallucinates.

00:02:20.520 | So I prompted Sora with an image I featured in a previous video and then had the ending

00:02:25.360 | be that the sign falls off the turtle.

00:02:28.280 | The turtle doesn't notice and walks off screen.

00:02:30.600 | The sign, by the way, is supposed to stay on the ground.

00:02:33.800 | Here was the resulting video.

00:02:35.480 | Now when you're watching this video, remember that if you buy the $200 tier just for Sora,

00:02:41.400 | this video alone used up more than 5% of your allowance.

00:02:45.480 | Translated, it costs you, or in this case me, $10.

00:02:49.040 | For those listening to the podcast version of this video, the sign definitively does

00:02:53.480 | not fall off the turtle.

00:02:55.360 | The next topic that I just have to get to if any of you are considering getting Sora

00:03:00.120 | is the refusals.

00:03:01.880 | Any prompt mentioning something proprietary, in this case an Arsenal shirt, just gets blocked.

00:03:07.280 | In this case, as with the VPN, there is one way of slightly getting around that.

00:03:11.880 | What you can do is generate the relevant image in a different image generator, something

00:03:16.640 | like Ideagram or Midjourney.

00:03:18.760 | Then you feed it in as an image prompt and, lo and behold, Sora will generate the video

00:03:24.600 | based on that image.

00:03:26.200 | Notice again, though, that this hedgehog is supposed to be scoring a goal rather than

00:03:30.800 | just staring at the potato.

00:03:32.480 | Also, did you notice that massive mangling at the beginning?

00:03:35.560 | God knows what that is.

00:03:36.560 | Oh yeah, and the potato levitates.

00:03:38.920 | So far in my use of image prompts, it has been pretty hit or miss.

00:03:43.240 | I tried to make the image logo for my AI Insiders on Patreon pop out.

00:03:48.640 | So I fed in the image and the prompt as part of a storyboard.

00:03:53.680 | And the prompt was the robot, and this is around four seconds in, it should kick in.

00:03:58.080 | The robot holding the AI Insiders book, looks at the book, then brings it closer to the

00:04:02.640 | screen.

00:04:03.640 | You can move those prompts earlier or later in the timeline to adjust when they kick in.

00:04:07.560 | The only problem is, as you can see, the robot instantly transforms into a different robot,

00:04:12.480 | albeit holding the right book.

00:04:14.280 | So that's really cool, but not quite what I was intending.

00:04:18.420 | Big plug for my $9 a month Patreon, where I release videos like this one on the media's

00:04:23.720 | misreporting over O1's quote escape.

00:04:26.360 | My favorite creation had to be this one, a hedgehog showing off its vegetables.

00:04:31.360 | I did prompt it with an image, but I think the resulting 1080p video is crisp and clear.

00:04:37.480 | You might know that your prompt can be a video, not just an image or some text.

00:04:42.680 | And so I got Sora to extend this scene both backwards in time and forwards in time.

00:04:47.920 | For Kling AI, I used the motion brush tool to get the turtle to move, but Sora quite

00:04:52.580 | literally took the turtle in a different direction.

00:04:54.920 | Obviously, I'm not the best text-to-video prompter, but you can see some hand-selected

00:04:59.440 | videos in the featured section on the product page.

00:05:03.320 | Some of these are pretty incredible.

00:05:05.060 | This is 1950's Suburban Bliss, generated just 40 minutes ago, 1080p of course.

00:05:11.480 | This is a drone shot of a container ship on docks, loading containers.

00:05:16.880 | Not bad at all.

00:05:17.880 | If you were just looking for a bit of cuteness in a quick short video at say 720p, there

00:05:22.800 | are competitors though, like Runway and Pika.

00:05:26.200 | But yes, overall, Sora is the best and especially at higher resolutions, I wouldn't say it's

00:05:31.880 | that close.

00:05:32.880 | Just don't bank on it being reliable as Sam Altman found out tonight when they were

00:05:36.960 | doing the live demo.

00:05:38.400 | In neither of the videos did the crane catch a fish.

00:05:41.800 | Then there are rivals like Google DeepMind who've produced Veo, their most capable

00:05:46.340 | generative video model.

00:05:47.800 | The only problem is, it's not accessible to anyone hardly, and definitely not me.

00:05:53.320 | So that, if you will, is my initial review of Sora.

00:05:57.020 | The best currently available, although very pricey with a limited amount of generations

00:06:02.720 | and hardly any adherence to physics.

00:06:05.520 | The system card, by the way, even though I read it in full, doesn't tell us that much,

00:06:09.760 | so I've just picked out five highlights.

00:06:11.920 | After covering these highlights, I'll touch on why I think there is an element of distraction

00:06:16.140 | in play.

00:06:17.140 | First is the team keeps repeating "we made Sora to understand and simulate the real world",

00:06:22.160 | a capability we believe will be an important milestone for achieving AGI.

00:06:26.240 | Here's my cynical theory on that.

00:06:28.480 | OpenAI was funded by people like Elon Musk as a non-profit to create AGI that benefits

00:06:34.720 | humanity.

00:06:35.720 | It is really quite hard to link the creation of an entertaining video generator to creating

00:06:41.920 | AGI.

00:06:42.920 | It's great for signing up new subscriptions, and indeed new signups are currently blocked

00:06:47.580 | because it's so popular.

00:06:49.000 | And great, of course, for building a ton of revenue, but really quite hard to link to

00:06:53.520 | the creation of AGI.

00:06:54.520 | Call me cynical, but Sora is miles away from understanding the real world, if it ever could,

00:07:00.280 | and certainly further than a model like O1 in ChatGPT.

00:07:03.880 | Just seems like a justification for why they're doing Sora.

00:07:06.640 | For all the lawyers watching, they again don't go into where they get the data from, I covered

00:07:11.400 | this back in February of this year.

00:07:13.120 | They just say "mostly collected from industry standard machine learning datasets".

00:07:18.320 | They are very well aware that they are a big target for lawsuits, which is why they're

00:07:23.400 | not going to be any more specific.

00:07:25.300 | One interesting bit of technical detail is that they customise their own GPT to achieve

00:07:30.260 | high precision on the moderation for certain topics.

00:07:33.520 | And the reason that they could afford that latency is because we're all waiting anyway

00:07:37.880 | for the video generation, so they use that window of time to run this precision targeted

00:07:43.400 | GPT.

00:07:44.400 | What that model can do, by the way, is identify third party content as well as deceptive content.

00:07:49.720 | I tried clipping out a section of Lord of the Rings and feeding it in as a video prompt

00:07:54.600 | and it was flagged immediately.

00:07:56.000 | Next, in case you were wondering or intending to do this, you can't actually ask for a video

00:08:01.880 | generation in the style of a living artist.

00:08:04.720 | They thought about allowing this, but then opted for the conservative approach.

00:08:09.080 | Also, you can't use as an image prompt, a photo or video of a real person.

00:08:14.500 | They say given the potential for abuse, we are not initially making that capability available

00:08:19.040 | to all users.

00:08:20.040 | I know that's not many highlights from a long report, but that is as much as I got.

00:08:25.080 | And I think you've got to admit that that was at least a fairly decent segue to this

00:08:29.040 | video sponsors 80,000 hours.

00:08:31.500 | One of their recent podcasts from the 27th of November was covering what one critic calls

00:08:36.260 | OpenAI's theft of the millennium.

00:08:38.540 | As you can see, they are super highly rated, but they also cover other topics too.

00:08:43.020 | I'm also subscribed to their YouTube channel and tend to listen when I'm doing long drives

00:08:47.800 | or long walks.

00:08:48.980 | They also produce a career guide, so do check them out with the links in the description.

00:08:53.340 | Just as I end though, what's that distraction element that I mentioned at the beginning?

00:08:57.140 | For some of you, this will be the most important part of the video.

00:09:00.140 | Well, we have all of these so-called 12 days of shipmas, OpenAI releasing product after

00:09:05.940 | product, one after another.

00:09:07.940 | And all of us are starting to think about Christmas and breaks to come, so that would

00:09:12.660 | be a good time to bury bad news.

00:09:14.860 | Just one example would be a coincidence, but three, let me know what you think.

00:09:18.940 | First, on December 2nd, OpenAI opened the door to doing ads.

00:09:23.300 | That doesn't sound like too big of a deal though, right?

00:09:25.820 | Their CFO just said that they're looking into ads.

00:09:28.220 | But remember, Sam Altman has said in the past, "I think that ads plus AI is sort of uniquely

00:09:33.540 | unsettling to me."

00:09:34.980 | On Lex Friedman, he also called them a last resort.

00:09:37.820 | That's just the warm-up though, because on December 6th, in the Financial Times, we

00:09:42.180 | got this.

00:09:43.180 | This story for me, if it turns out to be true, is far more, quote, "unsettling".

00:09:47.500 | OpenAI have promised that when they reach AGI, their commercial contract with Microsoft

00:09:52.420 | would be void.

00:09:53.500 | All the profits thereafter would go to their non-profit to distribute to everyone.

00:09:57.780 | That was key because as Sam Altman himself has said, AGI would break capitalism.

00:10:02.780 | It's even in their very definition of AGI that it can do half of the world's jobs.

00:10:07.380 | They made this commitment repeatedly, I've covered it many times before on the channel

00:10:11.460 | and it's in their bloody charter.

00:10:13.340 | Well apparently, they are now discussing removing that provision.

00:10:17.740 | AGI, according to this article, might well be misused then for commercial purposes.

00:10:23.540 | Just for a moment, I want you to picture Microsoft having a monopoly over AGI.

00:10:28.740 | You might not believe they would do so, and you might not even believe AGI is coming,

00:10:33.180 | but just imagine if those two things happen.

00:10:36.300 | Why are they trying to ditch this provision that shuts Microsoft out of its most advanced

00:10:40.700 | models?

00:10:41.700 | A provision they've promised repeatedly to uphold?

00:10:44.020 | Well, they are seeking to unlock billions of dollars of future investment.

00:10:48.140 | Just in case you think it's me making up all of these clauses, the FT directly quotes

00:10:52.700 | them many times.

00:10:54.220 | According to OpenAI's own website, AGI is explicitly carved out of all commercial and

00:10:58.980 | IP licensing agreements.

00:11:01.100 | But there's a problem with that, dear viewer.

00:11:03.020 | That would limit the potential profit and value for Microsoft.

00:11:07.240 | After all, they've pumped in $13 billion into OpenAI.

00:11:11.120 | This would disincentivise the big tech group from further investment.

00:11:15.420 | Oh no.

00:11:16.420 | By the way, this isn't just rumour, this is according to multiple people familiar with

00:11:20.420 | the conversations.

00:11:21.420 | Altman said, apparently, "We've left ourselves some flexibility because we don't know what

00:11:25.960 | will happen."

00:11:26.960 | Again though, the FT quotes from OpenAI's own history.

00:11:30.280 | They told anyone investing in them to consider their investments in the spirit of a donation,

00:11:35.220 | with the understanding, as I've quoted before on this channel, that it may be difficult

00:11:38.720 | to know what role money will play in a post-AGI world.

00:11:42.180 | All that seems out of the window though, and there's only one thing they're promising.

00:11:46.340 | OpenAI's non-profit, which does still exist, they say, will continue to exist and thrive

00:11:51.780 | and receive full value for its current stake.

00:11:54.720 | We don't know the exact figure, but I think their current stake is in the region of 30%

00:11:58.820 | of the current value of OpenAI.

00:12:00.740 | In other words, that non-profit will be bunged $50, maybe $80 billion, and receive quote

00:12:06.140 | full value.

00:12:07.140 | Yes, that will surely give it an enhanced ability to pursue its mission, but that's

00:12:11.780 | not the same as Microsoft not owning AGI.

00:12:15.300 | That's the same Microsoft, by the way, that was reported just today in the information

00:12:19.440 | was boasting about how much labor costs you could save if you adopt AI.

00:12:23.620 | They're being explicit about making sales for their co-pilot service based on showing

00:12:28.740 | the example to their customers of how they laid off 10,000 people last year.

00:12:33.540 | That's according to three current sales employees.

00:12:36.180 | We've been able to improve our throughput per customer service agent by 12% using co-pilot.

00:12:41.560 | That's real money.

00:12:42.560 | That means we don't have to hire as many people, Microsoft's Spatero said.

00:12:47.400 | Is that the kind of company that you want controlling AGI?

00:12:50.440 | Oh, and by the way, I haven't even got to the third thing that OpenAI are potentially

00:12:55.200 | trying to distract us from.

00:12:56.820 | They have pivoted to work inside the military-industrial complex, albeit with some small caveats.

00:13:03.360 | As MIT reports at the start of this year, OpenAI's rules for how armed forces might

00:13:08.720 | use its technology were unambiguous.

00:13:11.520 | No one could use their models for weapons development or military or warfare.

00:13:16.360 | Then that changed in January to "don't use our technology to harm yourself or others"

00:13:21.960 | by "developing or using weapons" or "destroying property".

00:13:25.520 | That "destroying property" went out of the window quite quickly.

00:13:28.400 | In October, they changed the terms to "you can only use it to protect people and deter

00:13:32.960 | adversaries".

00:13:33.960 | Now, though, their technology will be deployed directly on the battlefield.

00:13:38.640 | To help, apparently, the US and allied forces defend against drone attacks.

00:13:42.560 | Now, you can argue the rights and wrongs of this, but you've got to admit it's a shift

00:13:46.500 | in policy.

00:13:47.500 | As one analyst said, "Defensive weapons are indeed still weapons.

00:13:51.400 | They can often be positioned offensively, subject to the locale and aim of a mission."

00:13:56.680 | OpenAI, MIT says, has long pontificated about how to steward AI responsibly, and they will

00:14:02.080 | now work in a defence tech industry that plays by an entirely different set of rules.

00:14:07.200 | In that system, when your customer is the US military, tech companies do not get to

00:14:11.160 | decide how their products are used.

00:14:13.000 | And according to the Washington Post, even their own employees are pushing back on the

00:14:17.200 | deal and asking for more transparency from leaders, wherever we heard that before.

00:14:21.880 | Those OpenAI employees apparently want assurances that their technology won't be directed against

00:14:28.160 | human piloted aircraft.

00:14:30.080 | It's a great point, right?

00:14:31.320 | Because a defensive weapon could still be used against humans if those humans are piloting

00:14:36.020 | offensive aircraft.

00:14:37.440 | One other OpenAI employee said that "Defensive use cases still represented militarization

00:14:42.520 | of AI" and noted that the fictional AI system Skynet, which turns on humanity in the Terminator

00:14:48.200 | movies, was also originally designed to defend against aerial attacks on North America.

00:14:53.080 | OpenAI executives quickly acknowledged the concerns.

00:14:55.840 | It should immediately be pointed out that they are not the only ones with people like

00:14:59.960 | Anthropic and Meta changing their policies to allow military use of their technology.

00:15:04.440 | If you were looking for a purely positive review of Sora, I'm sorry to have disappointed

00:15:09.460 | you.

00:15:10.460 | The videos can be amazing if you ignore the physics and the user interface is exemplary.

00:15:16.100 | But it's quite pricey for the number of credits you get.

00:15:19.580 | And I just wonder if these 12 days of releases might be distracting us a little bit from

00:15:26.140 | other news about OpenAI.

00:15:28.100 | As always though, let me know what you think.

00:15:30.340 | Thank you so much for watching to the end, and have a wonderful day.

Sora is Out, But is it a Distraction?

Chapters