90
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 20 Jun 2026
90 points (100.0% liked)
Technology
85604 readers
3350 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
The cost will come down in 3 generations a lot, but the 1to3 trillion spent now to then will be worthless.
I just don't see the cost coming down. It appears the expenses go up fairly linearly with their revenue, and if they lose subs, they'll have to recoup that difference.
There was an article about open AIs finances the other day, and they lost 20 billion just last year, which is up about 4x from the previous year. Yes their revenue also when up 4x, but that the dollar amount for the loses grew more. If they keep on a similar trajectory, even investors are going to step away.
Prices have already gone up for at least Copilot and Gemini to token use pricing from borderline unlimited use. It's going to, if it hasn't already for many, reach a point that just hiring someone to do they AI work will be cheaper and more reliable.
There will be some sectors that benefit, like IT security monitoring, but I think optimized local models will take that spot. It already can for most personal users, it's easy to setup, and new hardware is coming out with local LLMs in mind.
There was just an open source subquadratic model released that performs about as good as the last round of frontier models, with a 12 million token window, but with 3 orders of magnitude less resource usage. Google also has their compressed models which greatly reduce processing. I think at some point the stuff we use frontier models for today will be running locally on our PC's and all this buildout will have been for naught or for people who have such good business cases that paying top dollar for a better model makes sense.
My question is this model is it open source or is it open weight because there is a difference. Open source gives you access to the training data that was used for the model. It gives you the code used to make the model. It gives you checkpoints during training of the model and methodologies that were used to train the model, etc.
Open weight models, however, don't give you access to the training data or snapshots during training. Mainly, they just give the methodology and what they weighted the data as.
The only fully open source model I am aware of, personally, is OLMO from AI2.