-

2025-10-22 12:00:46

jabgoe2089@hub.netzgemeinde.eu

ha! ha!

Software vendors keen to monetize AI should tread cautiously, since they risk inflating costs for their customers without delivering any promised benefits such as reducing employee head count.

one of these days i'll start an orgonit business ... :-)

#ki #ai #bubble #trend #bitcoin #bluetooth #computer #it

Benjamin Carr, Ph.D. 👨🏻‍💻🧬 schrieb den folgenden Beitrag Sat, 11 Oct 2025 21:35:01 +0200

#McKinsey wonders how to sell #AI apps with no measurable benefits
#Consultant says software vendors risk hiking prices without cutting costs or boosting #productivity
Many software firms trumpet potential use cases for AI, but only 30% have published quantifiable #ROI from real customers. Meanwhile, many customers see AI hiking #IT costs without being able to offset these by slashing labor costs. The billions poured into developing AI models mean they don't come cheap.
https://www.theregister.com/2025/10/09/mckinsey_ai_monetization/

computer

2025-10-22 12:13:20

Aladár Mézga

adolar@diaspora.psyco.fr

Yeah, software that's truly useful isn't that easily made. For me personally LLMs have yet to yield a useful application. OpenWebUI does unfortunately not work, since it has been built with a specific, now obsolete python version. Ollama I haven't got working with MCP yet and in StudioLM models outright crash since I switched to CachyOS.

So the work I would like to let the software do can't be done yet and what it can do I don't really need. Waiting...(for the ghost train...)

2025-10-22 13:57:30

eeepee@diaspora.psyco.fr

@Alexander Goeres Training of LLMs needs enourmous number of resources. Using a trained model can be done on your phone.

E.g for Nextcloud there's an AI module for tagging photos with various kinds of labels. Runs also without any use of a GPU.

2025-10-22 17:38:33

Aladár Mézga

adolar@diaspora.psyco.fr

Enough power? That depends on how you define "enough"... I have been running several models on an old SFF i7 with four cores and now 48GB DDR4. Graphics card is an old MX720(?) with 2GB. You need patience, especially with "reasoning" models. But if I use a "small" one (7B or smaller), especially with not too much context, answers come pretty quickly. I can still follow the sentences appearing on the screen, but it's quite reasonable - from my point of view.

But my goal is actually to give a task and let the software perform it - on the file system, on the internet etc. I'm not very much into "chatting". Currently I use Ollama, LMStudio is not very cooperative at this time...