The AI dooooooom thread

[Baron von Schtinkenbutt]

Microsoft dropped the anthropic licenses. It was using internally.  Why?  Because it's too expensive for the benefit it produces.

If one of the most wealthy companies in the world cannot use AI tools economically, who can?

Microsoft is joined at the hip with OpenAI.  There's not much reason for them to use anything by anyone else.

[Tonitrus]

I used ChatGPT the other day to look at my crash log from (my overly modded) Skyrim and tell me what went wrong. I haven't iterated on the process to know if it'll get me out of the hole (i.e. I made changes based on the feedback, but haven't really played since so I don't know if it works). If it does I'll happily concede that that's a good use case for AI.

I've done similar things with Rimworld and mods, and ChatGPT has been really hit-or-miss (mostly miss), as it has often pointed me at problems or solutions that were completely made up...espcially with Rimworld's XML coding ("use this XML tag!"...that XML tag doesn't actually exist..."oh whoops, you're right...I cannot help you").

It seems like AI-coding in Rimworld modding is growing...and there also seems to be some extreme, no-concessions backlash against it...even against good ideas coming from those who just don't have the coding skills.  Some anti-AI prejudice in this sphere is well founded...some is just reactionary.

[crazy canuck]

Microsoft dropped the anthropic licenses. It was using internally.  Why?  Because it's too expensive for the benefit it produces.

If one of the most wealthy companies in the world cannot use AI tools economically, who can?

Microsoft is joined at the hip with OpenAI.  There's not much reason for them to use anything by anyone else.

No, they dropped it because of cost

Microsoft's internal Claude Code pilot is over. The company launched the program in December 2025 across its Experiences & Devices division, then watched token-based billing consume the team's entire annual AI budget within months. The cancellation takes effect June 30. The mechanism is a classic enterprise cost trap: flat seat licenses kept token spend invisible. The moment Microsoft switched to usage-based pricing, the true cost became immediately visible and unmanageable. Internal developers are now being redirected toward GitHub Copilot, which Microsoft owns outright. Essentially: (Microsoft, Anthropic) are a cautionary case study in how enterprise AI pricing models create structural budget exposure that procurement teams aren't equipped to manage. - Claude Code pilot launched December 2025 in Experiences & Devices; cancellation deadline is June 30, 2026. - Full annual AI budget was consumed in months under token-based billing. - Microsoft is steering affected developers to GitHub Copilot as the internal replacement. This lands at the worst possible moment for Anthropic, which is simultaneously raising at a $900B valuation on the premise of accelerating enterprise Claude adoption

[Baron von Schtinkenbutt]

No, they dropped it because of cost

Sorry, I didn't mean to imply cost wasn't a motivator.  In fact, OpenAI's models are going to be inherently cheaper for them because they're hosted in Microsoft datacenters.

Also, as noted in that excerpt, they're still using "AI".  They're just using their own, built on OpenAI's models and their infrastructure, internally.  They're also pushing Copilot pretty hard externally as well, so it makes sense to dogfood their own system.  I don't agree with this author's conclusion that the reason is entirely financial, nor with the conclusion that this is a "cautionary tale".  Microsoft is not a neutral purchaser of such services.

[crazy canuck]

I think the key Takeaway though is that they blew through the entire year that they had budgeted in just a few months.  Presumably, but for that, they would have kept the licenses in place for at least a year.  It had already been budgeted.

I think the second important Takeaway is that companies or not properly budgeting the true cost of these tools which is I suppose understandable given that there isn't a lot of experience to draw from

[Sheilbh]

Thanks, that is a really interesting perspective.

I am coming to the view that some limited AI agents, when used properly, can be helpful.

But those are big caveats.  An example is an AI tool being developed at the UBC law school to assist with legal research. It is not designed to give answers, but it is designed to get the researcher to think about the questions they ask themselves and the topics they are going to research.  That sort of collaboration between a researcher and an AI tool could prove very beneficial to ensure that an area that should've been researched isn't missed.

But what I'm seeing in practice is people relying on generative AI tools to spit out research answers, and that's where we see fabrications of cases and legal principles that simply do not exist.

On a purely legal side there's been a slightly horrifying case in the UK recently around contempt of court from the insolvency courts. But it was kind of interesting because it was a law firm (Pinsent Masons) with an actual license to use the AI system which meant that the prompts and the entire conversation could be entered into evidence.

For good reasons the partner and senior associate were named while the junior associate wasn't. But what was really striking in the way the junior associate used it was that the AI system basically repeatedly said "you should check this and actually verify against statute - and they didn't. The firm has self-referred to the regulator and the judge was very critical at insufficent oversight by the senior associate and partner (while acknowledging no deliberate attempt at misleading the court). But it was slightly interesting in the context of conversations about "hallucinations" when the AI system itself was saying repeatedly that it needs to be verified.

[Sheilbh]

Sorry, I didn't mean to imply cost wasn't a motivator.  In fact, OpenAI's models are going to be inherently cheaper for them because they're hosted in Microsoft datacenters.

Also, as noted in that excerpt, they're still using "AI".  They're just using their own, built on OpenAI's models and their infrastructure, internally.  They're also pushing Copilot pretty hard externally as well, so it makes sense to dogfood their own system.  I don't agree with this author's conclusion that the reason is entirely financial, nor with the conclusion that this is a "cautionary tale".  Microsoft is not a neutral purchaser of such services.

I agree I think it's quite tough for Microsoft to rely on certain other systems. It's like their delightfully quaint insistence that Bing and Bing Search are things.

They can't be neutral and just use the "best" service. I agree that costs are a problem - I suppose the question to ask is the extent to which we think those costs are fixed and are sticky.

[crazy canuck]

Thanks, that is a really interesting perspective.

I am coming to the view that some limited AI agents, when used properly, can be helpful.

But those are big caveats.  An example is an AI tool being developed at the UBC law school to assist with legal research. It is not designed to give answers, but it is designed to get the researcher to think about the questions they ask themselves and the topics they are going to research.  That sort of collaboration between a researcher and an AI tool could prove very beneficial to ensure that an area that should've been researched isn't missed.

But what I'm seeing in practice is people relying on generative AI tools to spit out research answers, and that's where we see fabrications of cases and legal principles that simply do not exist.

On a purely legal side there's been a slightly horrifying case in the UK recently around contempt of court from the insolvency courts. But it was kind of interesting because it was a law firm (Pinsent Masons) with an actual license to use the AI system which meant that the prompts and the entire conversation could be entered into evidence.

For good reasons the partner and senior associate were named while the junior associate wasn't. But what was really striking in the way the junior associate used it was that the AI system basically repeatedly said "you should check this and actually verify against statute - and they didn't. The firm has self-referred to the regulator and the judge was very critical at insufficent oversight by the senior associate and partner (while acknowledging no deliberate attempt at misleading the court). But it was slightly interesting in the context of conversations about "hallucinations" when the AI system itself was saying repeatedly that it needs to be verified.

Yeah, when computer based research came along, we were forever, telling juniors not to start with a computer search. We kept telling them that they needed to read the texts and the digest to understand what they need to be searching for, and to learn the vocabulary of the area of law that they were searching.  The good ones did that but of course the majority didn't and it showed.  I think AI tools are becoming a lot like that where people looking for a shortcut go with whatever the AI tool tells them based on whatever input they put into the AI tool and call it a day, with disastrous results.