#lumoai — Public Fediverse posts on home.social

Mindaugas @[email protected] · 2026-08-22 · 07:48 UTC

If anyone is using Anthropic AI models, I would suggest not to.

There is speculation that enterprise focused Anthropic is rushing to do an IPO for trillions in valuation to overshadow spacex (like another marketing stunt), which would give them overwhelming resources to continue blow the AI bubble (with these new influx of vast sums of money) and concentrate power even more.

IF you use AI - don't pay them money (*with exceptions, because they loose a lot of money on power users), and don't add to their revenue for their investors to show how great they are doing.

They are not doing that great, as small Chinese open models have significantly improved in the last few months and can be used instead to achieve similar results for a lot less money.

Such as "Deepseek v4 Flash 0731" 250B size model, which costs 62X 😯 less than Anthropic best (available) model, and which also has higher intelligence than Anthropic's best available model from 6 months ago.

Or recently released "Qwen 3.8max" 27B small model variant, which can be run on the desktop locally, and is also more 'intelligent' than Anthropic's best model was 6 months ago.

Models don't have to be big, they can be smaller. One does not need to transfer Anthropic money for their best models, because these fresh small models can do the the job*.

*With exceptions - small models have a lot of distilled intelligence for their size, but they don't have a lot of knowledge. When some task require some deep-niche-specialized knowledge (like translating between small languages), then one can try use large variants like "Qwen 3.8max" 2.4 trillion parameters sized open model. It cost a lot less than Anthropic's as well.

You can find prices on OpenRouter https://openrouter.ai/ to choose from countless providers of open models (but OpenRouter is also becoming big tech right now, it was recently was bought by Stripe for 7 billion), so there are also French alternative AI router https://www.edenai.co/, or UK alternative AI router https://www.requesty.ai/, or privacy-crypto-chat-UI AI router https://nano-gpt.com/ to connect to open model providers). And for agentic coding software (so called "AI agent harness") use https://kilo.ai/ open source software (or choose from many other options), not "Claude Code".

And for everyday (non-coding) tasks, one can choose privacy oriented OpenAI alternative @protonprivacy Lumo with regular chat interface. They are using very capable open AI models as well.

I attached visualization how models evolved recently, and compared where Proton Lumo 2.0 currently stand.

And if you don't use AI, so continue to not use it. But if you do, these better alternatives are available. "Better" does not mean "right" - open weights AI can also become very problematic. It is at least better than to support AI companies of very concentrated power.

If movement to open models spread quicker than Anthropic IPO plans, hopefully Anthropic will go bankrupt for good. #aibubble #LumoAI

#aibubble #lumoai