Forum - View topic

Answerman - Addressing the Amazon AI Dubs Controversy

Microsoft was an early investor in many of the latest AI companies. It ended up with a serious stake in OpenAI and benefited from early access to its models, creating Bing Chat and Copilot when Google, Meta, and Anthropic were just getting started. But now its momentum has stalled, and like everyone else, it's not making much money from its AI products. That's because no one is buying them, and that is because very few people actually find them useful, The Information reports.

"The Information’s story inaccurately combines the concepts of growth and sales quotas," Microsoft said in a very defensive statement (via Futurism), adding that "aggregate sales quotas for AI products have not been lowered."

Petulance aside, tests from earlier this year found that AI agents failed to complete tasks up to 70% of the time, making them almost entirely redundant as a workforce replacement tool. At best, they're a way for skilled employees to be more productive and save time on low-level tasks, but those tasks were already being handed off to lower-level employees. Having an AI do it and fail half the time isn't exactly a winning alternative.

Most generations have felt the world flying past them as newer generations embrace new status quos, but I really worry about the next generation growing up in a technological ecosystem designed to occupy their attention and stifle their control of their own lives...

My advice to Amazon is to stick to training and testing your AI-generated dubbing and subtitling on shows that you own outright and that you have produced.

I'm not sure why you keep saying AI is inevitable, Microsoft was one of the earliest AI backers and the have slashed their usage targets by as much as 50% because they just aren't making money. They say those are growth targets and not current usage rates, but the fact is they spent all this money believing they could make at least twice as much as they expect to now and we can probably expect to see those target drop further.

Cutting future projections isn't the same thing as failing though, and from what I've read that's mostly because people are using other AIs instead which are the ones experiencing the surge and growth. Given how many companies are jumping on AI there's bound to be a lot of losers in the field as only some can win. Not sure that's indicative of much myself. You could draw the parallels to Microsoft previous attempts at getting into markets they were pushed out of like consoles and MP3 players or web browsers where people prefer others over their offerings.

I do think AI is inevitable too, if it already isn't considering every store and business I go to use it to some degree like an assistant or customer service and now and even my doctor uses it to transcribe our visits now. I've been playing around with Google's new Nano Banana Pro and I don't see the genie going back in the bottle myself. It's just too good at what it does. All it took was 2 random screenshots I took of a show to generate tons of perfectly on-model pictures that looked like screenshots from an actual episode. The fact one of my posts went viral and people mistook it as real screenshots from a show was both funny and kind of scary at what the future holds for this. Like, maybe companies will struggle to find a use for it but the technology itself speaks for itself I think. Someone is going to utilize it: someway, somehow.

It feels like the wrong thing to take away from this scenario is that AI voiceovers sound bad. There's plenty of perfectly generated AI voices out there so even if one is against AI I think using the 'quality' argument isn't going to hold water. Amazon's poor outing was certainly weird and makes me think it genuinely was a mistake and not meant to be seen/heard because you see amateur YouTubers utilize it better.

The gap between what it costs to run these models and what companies charge for API access is substantial. Training runs for frontier models cost hundreds of millions of dollars. The GPU clusters required for inference at scale represent billions in infrastructure. Energy costs alone are astronomical—a single ChatGPT query uses roughly 10x the energy of a Google search.

Yet somehow, you can access GPT-4 for pennies per thousand tokens. Claude Sonnet for even less. Google offers Gemini with generous free tiers. These prices don’t just ignore the capital costs—they often don’t even cover the marginal cost of compute.

Why? Because right now, we’re in a land grab. Every major AI company is prioritizing ecosystem lock-in over profitability. They want developers building on their platforms. They want users getting comfortable with their tools. They want to be the default choice before someone figures out how to monetize this properly. The model is familiar: subsidize heavily, capture the market, then figure out how to make money.

We’ve seen this movie before. Uber spent years subsidizing rides to below the actual cost of providing them, burning through billions in VC money. WeWork convinced people that office space economics didn’t apply to them. In the early days of cloud computing, AWS and others underpriced services to drive adoption, then steadily increased prices as customers became dependent.

The AI correction will follow a similar pattern, but potentially more severe. Several triggers will force the shift:

VC patience runs out. Investors have been remarkably tolerant so far, but even in AI, you can’t lose money forever. As companies mature beyond Series B and C, the pressure to show a path to profitability intensifies.

Market consolidation. Right now we have aggressive competition keeping prices low. As the market matures and consolidates—and it will—pricing power will shift to the remaining players.

Regulatory and energy constraints. The energy consumption of AI is already drawing regulatory attention. Carbon costs, grid constraints, and environmental regulations could force prices up regardless of what companies want.

Industry analysts estimate that current API pricing may need to increase 3-10x to reach sustainable economics. Some models might need to go even higher. A comprehensive inference request that costs you $0.01 today might cost $0.05 or $0.10 tomorrow. For high-volume applications, that’s not a line-item adjustment—it’s an existential threat.

Sometimes the distributor receives a “delivery spec,” which is a document listing deliverable assets from licensors' labs, sometimes with instructions on leaving on-screen text untranslated, subtitle positioning, and presentation quality.