The AI dooooooom thread

[Syt]

https://www.pcgamer.com/software/platforms/microsoft-doesnt-know-what-to-do-about-the-memory-pricing-crisis-microsoft-is-causing/

Due to the memory pricing crisis that its own AI ambitions are helping cause, Microsoft does not know how to sell a new Xbox that feels cutting-edge at a price that regular people can afford.

The next Xbox, codenamed Helix, will have "leading-end performance," new Xbox boss Asha Sharma said in a recent interview with Fortune, but I'm left wondering how Microsoft defines that, because everything else she said downplays its likely technical capabilities.

What the console business needs, Sharma said, is "new business models" rather than "just the most premium, high-performance console in the world."

"I think we've reached a point where it will be hard to imagine that mass audiences can afford thousands of dollars to spend on a console generation," she said. "So I think we will see radically different business models that we never expected come into orbit later this year."

It's an astonishing place we're in. Consumer devices always involve compromises for the sake of affordability, but this is the first time I can remember a tech company lowering expectations for its next big gadget so dramatically.

Sharma went on to say that Microsoft will "have to think very differently about storage and memory going forward."

"We will have to apply new techniques so that we can compress [games]," the Xbox CEO said. "We will have to empower customers to have very flexible storage offerings. We will have to empower new types of games so that they can fit on-device."

There are outlines of ideas there, but Xbox has already done "flexible storage options," and what it means to "empower new types of games so that they can fit on-device" is anyone's guess. (Are they going to put "Now with smaller games!" on the box?)

As for "new business models," perhaps Microsoft will offer financing? A rent-to-own plan?

Another guess is that we're talking about cloud streaming. It's not strictly new—RIP Stadia—but it fits in well with big tech's AI obsession. At a GDC session I attended in March, Nvidia DLSS pioneer Bryan Catanzaro said that "AI is fundamentally much more efficient in the cloud." That's a useful premise if you want to move toward a world where games are served like Netflix shows and at-home devices can get by with less RAM and SSD space.

[HisMajestyBOB]

Why not just sell a box with AI on it? Why would anyone want anything else?

[The Minsky Moment]

Helix is confirmed to use an RDNA 5 GPU.  It's going to perform.  The question is pricing given RAM and storage costs.   But those costs could look different by the time Helix is ready to release.  OpenAI may not even exist by then in its present form.

[The Minsky Moment]

Let's put some data points on the board.

Data Point 1: the IPOs

Venture capital Mark I was the private equity get in early, take the company through the tricky early stages, then use the public markets to cash out at a nice premium, go on the next deal.

Venture equity Mark II is that the tech companies stay private even well after financial viability because the owners want to keep control and avoid dilution.  Space X was the poster child for this. Anduril. Stripe. Etc.  The PE players either stay in or trade stakes in each other's investments to record a monetization event and collect fees.

The big AI ventures are still quite young and far from profitable.  Under Mark I or Mark II they would be far from going public.  Yet now both OpenAI and Anthropic have filed confidential S-1s.  SpaceX has filed openly, and to be clear, the only reason it is going public is because of the AI assets Elon forcibly stapled to the legacy business.  It is an AI IPO.

There is only one logical reason for these filings.  The AI cos can't get the funding they need to sustain future operations from the private markets. And given the massive capacity of the private funding markets, it looks like they are trolling for sucker money.

Data point 2: the Softbank financing fail.  Softbank tried to raise a $10 billion loan on OpenAI worth $64.5 billion based on the price they paid for their stake (which is around half what is rumored to be the targeted IP price).  No takers.  They lowered it to $6 billion.  No takers. Further evidence that the smart money doesn't want more exposure to the LLM model cos.

Data point 3: The "tokenmaxxing" fad that swept corporate America seems definitively over, with clear indication that the benefits did not justify the cost. Thus, the revenue model needed to support the hypothetical IPO valuations has collapsed.

Data point 4: The gap between data center build out commitments and the actual build out is enormous and increasing. A key component of that gap is - to the surprise of no one with a clue - the long lead time involved in developing, deploying and procuring the massive electrical capacity required. That has not prevented the circular chain of AI players from burning lots of resources in the effort.  Given how far the current US administration's energy policy is up its own ass, and how politically toxic the AI hyperscalers have become  - little help from the government is to be expected.  The clearest path for eliminating the gap is large scale cancellation of build committments.

At this point, the competing private co. LLM players are all Cinderellas frantically racing to secure their prince (a huge war chest of IPO cash) before the ball ends and the true nature of their status becomes undeniable.  It's not the best business proposition.  Life usually isn't a fairly tale.

[Savonarola]

I volunteer as a reviewer for the IEEE EPICS initiative; which is a program where the IEEE gives small grants (under $10,000) to student groups who partner with non-profits to provide technical solutions for local projects.  Most of the projects are from south or southeast Asia (though I have gotten ones from Africa before.)

The first one I got this year was obviously written by an AI Chat of some sort as no human being could possibly have used the terms "Inclusive," "Community development," "Dignity," "Empowerment," "Underserved communities," and "Innovation" so often.

I just reviewed another round of proposals.  One of the proposals this time was from a group in Ecuador for an AI teddy bear.  They wanted to make a teddy bear that would interact with a child and foster learning and development.  As I was reading it I kept thinking "You want $10,000 to design a Furby?"

There are already AI teddy bears.  I didn't reject it, but I sent it back for revisions telling them they had to explain how their AI teddy bear would be different than the ones on the market.  I think that any interactive child's toy that connects to an open AI would be a horrible idea (and one AI teddy bear was already pulled from the market for disturbing interactions with children), but that's not a criterion that I can use to evaluate proposals.

[crazy canuck]

The Canadian research tri-council funding agencies have forbidden funding applications to be written by AI.  I think that sort of prohibition will catch on quickly given the sort of nonsense that can end up in an AI written grant proposal.

[The Minsky Moment]

Hilarious/horrifying case out the District of Missisippi, where both sides were submitting AI drafted papers on the same set of motions.  Basically, AI arguing against AI, replete with hallucinatory legal authorities on each side.

https://storage.courtlistener.com/recap/gov.uscourts.msnd.50181/gov.uscourts.msnd.50181.123.0.pdf?ref=404media.co

The court tossed all the lawyers off the case and postponed the scheduled trial.

[garbon]

Hilarious/horrifying case out the District of Missisippi, where both sides were submitting AI drafted papers on the same set of motions.  Basically, AI arguing against AI, replete with hallucinatory legal authorities on each side.

https://storage.courtlistener.com/recap/gov.uscourts.msnd.50181/gov.uscourts.msnd.50181.123.0.pdf?ref=404media.co

The court tossed all the lawyers off the case and postponed the scheduled trial.

Read the thread.

[The Minsky Moment]

How about that, same case.