this post was submitted on 02 Jun 2026
552 points (99.8% liked)
Technology
85136 readers
3996 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
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
To be clear, We don’t really know the architecture or model size of claude or chatgpt. If opus is a 1 trillion parameter dense model, then yes of course it’s going to be way more expensive to run than deepseak which is an 862 billion parameter MoE model. American models have focused on being the best regardless of cost where chinese model makers were forced to focus on efficiency because of lack of access to chips. Which will probably give them an advantage as the free investor money runs out for the american companies.
The constant training is simply how AI works and will continue to work chinese or american. That cost will never go away. Anytime you need your model to learn new skills or gain new knowledge, it needs to be trained.
The best advancements in usability of AI haven't all come from training larger models. Tool usage is super useful and doesn't require new training for new tools, etc.
Deepseek for example doesn't have a monthly release schedule, they've only released one version so far this year.
Plus for new knowledge, there's web search. It's no longer strictly true that AI output is restricted to information available before the cutoff date.
At this point you only really need to train a new model if you're trying out architectural changes, you don't need to crank out constant updates.