Shimmering Disaster

The first instance of the relevant to the headlines topic occurs on page 11 of the PDF, everything before that is mostly standard money talk and discussions of content expansion (“video games and more recently, vertical video clips and podcasts”)

"Spencer Wang, Vice President of Finance, Corporate Development & Investor Relations: Our next question comes from Sean Diffley of Morgan Stanley [Global financial firm]. What have been the early learnings from the InterPositive deal?"

Note: InterPositive being the AI film production company founded by Ben Affleck in 2022 To take a detour to that announcement for context on how that tech might be applied, here’s Ben Affleck’s statement (edits mine to omit flourishes of language)

"Together with a small team of engineers, researchers and creatives, I began filming a proprietary dataset on a controlled soundstage with all the familiarities of a full production

[...]

our first model, trained to understand visual logic and editorial consistency, while preserving cinematic rules under real-world production challenges such as missing shots, background replacements or incorrect lighting. We also built in restraints to protect creative intent, so the tools are designed for responsible exploration while keeping creative decisions in the hands of artists — and ensuring that the benefits of this technology flow directly back to the story they're trying to tell.

The results of this foundational work were deliberately smaller datasets and models focused on filmmaking techniques — rather than performances — creating tools that artists can use, control and benefit from.”

Emphasised in this statement are:

the small scale of the production (what may be the Linkedin lists the size as 11-50 employees with only one user account listed on the page.)

The custom Data Set created in-house

The use of smaller datasets and models

That last point is a more ethical element that is possible but something that runs counter to how widespread machine learning generally seems to operate when trying for accuracy to life, they don’t elaborate on how this issue of scale is handled.

Not included is if any of these models are trained on external sets at any point for supplemental data, or where the machines required for processing are operated from, presumably in house but there’s no guarantee. Also not mentioned is if using this tech means the data it collects is private and proprietary, or whether it will be used in the future.

In this Netflix interview/promotion the conversational explanation of how this works is as follows: [again, edits mine to omit hype]

“It's not about text prompting or generating something from nothing. You’re building a model from your own material. That's how this works. You have to create your movie essentially first before you can really build your model around your movie using AI. And once you do that, you have your model.

[…]

You use this technology to learn from your own film dailies. Then you're able to introduce that in the post-production process so that you're improving your editorial process, your ability to mix, color, finish your film, do visual effects. [...] It can only understand this and only build this tool because it's trained on the character that the actor has already built.

[…]

I want to take out all the logistical, difficult technical stuff that often gets in the way. You can use your own model to remove the wires on stunts, reframe a shot, get a shot you missed, shape the lighting, enhance the backgrounds. If you can take some of those problems out, yes, you can do it more quickly, you can do it more easily, […]”

So again, Not much specifics on what the actual processes are beyond a post production focus.  The implication is video generation and effects modelling.

To theorise a bit- (from my limited outsider understanding)

“Film dailies” providing visual details that their tech would presumably be using as a template and altering using physics and realism learnt through the in-house training process.

Reframing and coverage of missed shots would probably involve a generative video situation.

Shaping the lighting and enhancing the backgrounds also but perhaps in a manner closer to a smart filter or the “upscaling” of making up what pixels make the image look sharpened.

My question is: something like removing a wire, how much of that is essentially automating the established colour keying/digital painting/cgi generation tools using logic based on machine learning, and how much is frame by frame image generation where you essentially have the original frame as the input and the output is a photorealistic image of pixels based on the machine learning (what’s the difference? Layers of separation/abstraction from the pixels representing the original image sensor data captured from reality). There are plenty of tools that use neural networking for machine learning to automate results that aren’t “generators”, such as auto selection for motion tracking or adjusting effects automatically based on information in the frame instead of manually adjusting values and processes in software. It’s unclear if these are what they are talking about.

So Ben Affleck’s spin is: This is small scale, curated, specific, and replaces tedious elements of production and post production. The implication is that this supports crew and performers and cuts down on hours and labour.

What Ben Affleck does NOT say: What it specifically does when applied in a task, how it might do it, what is done with the information collected once that project is completed, where the processing takes place.

Let’s climb back out of this digression back to the Netflix PDF. We now know that InterPositive is an effects and production company, founded by a celebrity, using vaguely defined machine learning to train small scale datasets custom to each project.

“What have been the early [since the 6th of March 2026] learnings from the InterPositive deal? And how should we think about potential cost savings in content creation? […] could this impact your $20 billion cash content budget on a go-forward basis? Or is it more likely to be reinvested into more content and better compensating talent?”

So the question is ‘it’s been 4 months since you got this technology, is it worth it or would it be better to pay more people to do the same thing for real?"

“Theodore A. Sarandos, Co-CEO, President & Director:

[…] it's early days for InterPositive, but we're broadly seeing that gen AI is starting to have an impact across hundreds of our productions. So important to note that we have other gen AI tools in addition to InterPositive. We're thrilled with all the speed they're bringing to market for us. But we also have Eyeline [ Formerly Scanline VFX, a Canadian visual effects and animation company founded in 1989 ] and we have our animation lab. […]”

This initial statement calls InterPositive contributions generative AI by grouping it with gen AI tools Netflix already used previous to this year.

Then he groups the company in with their animation and visual effects departments.

“gen AI is scaling quickly across the entire creative process, from concept to pre-vis through post and delivery. We're making higher quality output more quickly and efficiently than we could have using traditional methods. So gen AI workflows now have been used in roughly 300 of our titles with the largest concentration right to date is on post-production. But we're leveraging gen AI for really complicated shots and sequences. […] things like enhancing crowds or historical battle scenes, those kind of things.”

THIS section, one sentence of this section, is what entire articles are spun out on.

He starts with highlighting general applications: “concept to pre-vis through post and delivery.”  Just stating they use it and it’s faster is not enough hype so the phrase “higher quality” is then applied (debatable) and “efficiently” is also used, however the application of that term to gen AI is one I have quibbled over before and still maintain actually means just paying less people for something to happen faster.

“gen AI workflows now have been used in roughly 300 of our titles with the largest concentration right to date is on post-production.” THE Headline sentence,

What it means by “workflows” is vague but post-production being singled out is at least something. “roughly 300” is frustratingly vague and again it’s not clear if they are talking about something as innocuous as sound isolation or automating motion tracking, to something subtle like artificial lighting, to something more like cgi in effect or even the type of promotional images that would be achieved through a quick collage with airbrushing/editing the appearance of the source image or generating promotional ad copy text.

“complicated shots and sequences” does that me pre-visualisation before filming? Does that mean taking the dailies and trialling different framing angles or shot sequences? does that mean using it to paint over the behind the scenes and generate a virtual set/effects?

(Side note: it is within possibility to take standard machine learning and apply it to mechanical camerawork through standard use data for essentially a more versatile steadicam/drone/rig control to automate complex camera motions. If you want an example of a non generative application of this sort of tech for “complicated shots and sequences” (there’s already normal computer powered pre-programing and tracking) )

“enhancing crowds or historical battle scenes” so what Wētā Workshop did in 2001 but I supposedly though less rendering/hardware requirements. I am unclear on how much less because what is the difference between a render farm and a data centre other than the manner in which information is processed because if anything render farms seem much smaller scale for the ‘same’ output. And again normal CGI already automates a lot of this sort of thing through both asset libraries and regular programming.

Four paragraphs might seem excessive for me to break down one paragraph but considering that paragraph still tells us so little but kickstarted the whole news flurry I felt like it was necessary.

“keep in mind that in many of the cases, productions would have left out those key shots because they just wouldn't have been able to afford them, they wouldn't have been able to do them in the time frames that they're working on.”

YOU SET THE BUDGET AND YOU SET THE TIMES! Get better at your job or accept things take time

“So those sequences are saved by the availability and access to these Gen AI tools.”

They could also have been saved by better scheduling and funding. And there are PLENTY of successful film and tv productions made on lower budgets and shorter production schedules. Sorry to be like Get Good but apparently Stranger Things series 5 cost up to $400-480 million, took 2 years of pre-production, took 1 year to film, and 10 months post production, before airing with a month intermission splitting a single season.  And it wasn’t even good. So forgive me if I find the budget and time frame argument a bit dubious from this company.

“On the content side, we believe it takes great artists to make something great, and AI is not changing that. AI will give creators better tools to bring their visions to life. Movies are being made by people who make movies. AI provides them with better tools to make them even better.”

Again that sticky use of the word “better”, like “efficient” I am extremely sceptical of it. The use of “tools” is also questionable depending on what job the artists actually want to be doing since most artists like actually doing things themselves.

“Movies are being made by people who make movies” is asinine, of course, but also if I can get deep with it: Advertisements are also made by people who make movies. Movies are sometimes made by people who are only there for the pay-check on the way to the movie they want to be making. Movies are sometimes made by people contractually obliged to step in and finish the thing. Sandwiches are being made by people who make movies. Let’s be for real here. Sometimes people who make movies are using AI because their parent company told them to and not because they actually need it. This false sequentiality of technology will never not be annoying. The type of person who thinks 3D animation is better than or the successor to 2D, or that a photograph of a flower is better than a painting, when they are simply two different mediums with different aesthetic priorities. The same sort of mindset is being pushed by every person who claims the final result of generative AI is better than something created using regular computing or physical work. Just highlights that either they don’t know what the technology actually is or does or they want people who don’t know what the technology is or does to believe them and buy into the hype.

Not to mention the constant gen AI propaganda of  ‘it’s just a tool’ and ‘this won’t replace artists it will help them’ which is so widespread. Is it ‘just’ a ‘tool’ or is it a huge innovation? People argue it’s both when it might be less than either- a different method of processing existing tools using massive resources.

Back to the talk:

“our talent leverages tools for things like set references and pre-vis and VFX and sequence prep and shot planning, which all makes the production itself so much more smooth and efficient and fast.”

Ah! Specificity at last (kinda) ! so it IS for rendering imagery beforehand (an entire job), running through what a shot might look like or could have added in post (a normal part of the iterative process and probably something you want your creative team actively thinking about rather than offloading for consistency if nothing else), and shot planning (again vague how much of this is ‘streamlining’ a process versus replacing the work entirely) “So much more smooth and efficient and fast”. Smooth depends on what is actually being done which isn’t really described. Efficient again is debatable and might just mean paying less people. Fast, I will grant you, that is the big appeal.

“ We're seeing it across the entire production life cycle and AI -- those use cases are scaling faster and faster.”

Yes sure the way this sort of technology develops is acceleratory and when it does hit a plateau resources are expanded. Is the cost of resources worth it? (see various ethical concerns) Is achieving something indistinguishable from regular human labour worth it? (what is THE POINT if it isn’t visually distinct)

“our documentary series we just released […] That series features 17 minutes of AI-enhanced footage. It enabled us to expand the scope of the series in ways that just wouldn't have been feasible before. Those 17 minutes, Sean, they were produced twice as fast and at half the cost of previous options.”

I’m not interested in hunting down AI in Netflix properties (they’ve caught flack for ai in the opening credits of a marvel show before so it is obvious sometimes) but we FINALLY got given a name of a SINGLE title of 300 supposed use cases. 17 minutes does not mean much as still images can be on screen for a while, but the fact that a documentary is presenting things that have been essentially altered without being too obvious or open about it- in the way you might a re-enactment for true crime- is concerning, just definitionally.

Anyway here’s an example of one of the enhancements I got from clicking to a random spot: they took a WOOD ENGRAVING and “enhanced” it by trying to make it look like a paper illustration in a method that removed the vertical etchings and left behind muddy writing and cropped out the name of the artist. Howard Pyle from Harper's Magazine, March 1882, btw. I checked and neither on screen nor in the sources in the credits was this information found.  

“by equipping creators with these tools, we believe they're going to enhance their abilities, and we are going to have better and more impact for every dollar we spend on our programming. So content creation time lines can be shortened and quality can be enhanced.

So the cost savings will likely be reinvested into more content on the service, which fuels high-quality engagement and that whole kind of revenue profit flywheel that's going to come from that, that we've been talking about from day 1.”

Sick of hearing the term “enhanced” which could mean anything at this point. “better and more impact” another moment of strange vague hype phrasing. “time lines can be shortened” I still think probably the main draw being pushed here. “cost savings will likely be reinvested into more content on the service,” tricky thinking here- cost savings is how you promise money for investors, more content is how you promise to create that money, but is this not circular reasoning? Has the profits helped any of the 11 shows cancelled after only one season so far only halfway into this year? How much “high-quality engagement” can you make if your model trajectory seems to be cheaper and faster but also more money but it’s spread out to more shows?  

SO WHAT DO WE KNOW?:

Netflix have multiple AI services they use for multiple stages of production

They bought a small scale AI company from an A list celebrity and broadly advertised this acquisition of a company that had no prior attention over the first 2 years of existence until it’s sale. This created press for the actor, Netflix, and added to attempts to legitimise the use of AI in media production without actually giving concrete examples.

When asked about if this acquisition was worth it they said they use it alongside other AI services and regular production

They then went on to advocate their use of AI due to how it is faster and also only a tool and also makes results better and also makes things run smoother and cheaper and also doesn’t take control away from artists.

They also listed a bunch of jobs and tasks that are already accomplished by people and computers without machine learning.

Of 300 supposed use cases they listed a single miniseries of which 17 minutes out of 369 minutes was “enhanced” by AI (could mean anything but I confirmed at least one case of “hand drawn filter” on a wood carving)

Generative AI is apparently used across the entire pre-production, production, post production, marketing process.

It is not clear what is actually being done

No directors or crew actors who don’t own AI companies gave their opinion on working with this tech in the 3 articles I read or the earnings meeting.

“that we've been talking about from day 1.”

Okay, well I got a lot of rambling from this one short response turned headlines, so just for fun let’s look at day one: let’s do a little comparison between the Q2 release for 2007 versus 2026 (financial results and earnings call)

Reed Hastings, Netflix co-founder and chief executive officer:

“Online DVD rental is a large and attractive opportunity and we remain committed to investing in our long-term growth. With yesterday’s price cuts in two of our most popular subscription plans, together with the reductions in February and June, we are choosing to lower price and reduce marketing as the most efficient means of sub growth and retention in the current competitive environment, and we are lowering our full-year guidance for revenue, subscribers, and earnings accordingly.”

Cute.

“As we look at how to grow our business most efficiently, we make trade-offs between service levels, marketing levels and price levels.”

“At lower prices, we will spend less acquiring subscribers than we otherwise would have because of the increased attractiveness of our lower-priced programs. In other words, marketing reductions largely fund these price cuts.”

“The tactics of our increased growth investments are higher service levels, lower pricing and slightly less marketing because we believe that is the most efficient combination for Netflix at this point.”

I see a theme here.

Also something before their shift from hosting to producing media:

“Adding titles doesn’t particularly cost us. It’s the actual viewing, so the more people watch, the more we owe the studios, which is fair and appropriate.”

And more ominously-

“The next big step -- delivering Netflix online video to the television -- we expect to come together next year with a number of partners and we’ll have more to say about that next year. One subscription service with two delivery methods is we believe the path to long-term leadership and profits.”

“I’d say our competitive view is that we should focus forward in online video, so if you see us doing various partnerships, we would be probably more inclined to do that. In the area of which -- we’re an Internet company, we look forward on these things, rather than trying to go tit-for-tat with Blockbuster on more their core area being store logistics.

Today, stores are probably more relevant to most consumers than online video, partially because it’s only to the laptop. But stores are going to be less relevant every year over the next five years and online video is going to be more relevant every year over the next five years.

We have a great balance sheet. We’re investing in online video.