Devs clash over local AI: 8 years of Claude subs vs a quantized DeepSeek at 30tps
brandon_galang · x · 2026-10-14
Developer brandongalang argues local AI makes no sense unless you absolutely need privacy — and if you do, the task is probably too important to trust to dinky quantized models, so do it by hand.
He quotes theo's cost jab: eight years of $200/month Claude subscriptions buys you a team running a lousy quantized DeepSeek v4 Flash at 30 tokens per second — an 'incredible deal' for local inference fans.
More from Fun
- A 1960s Dune quote on outsourcing thinking to machines feels eerily prophetic in the AI era — Particular_One_1764 · 2026-10-14
- Eerie 'The Next Floor' corridor video packed with mannequins captivates Reddit — Public_Sleep7969 · 2026-10-14
- Meme: RSI is easy — just treat the model like Goofy and prompt it — andrew_n_carr · 2026-10-14
- Meme: hard environments aren't required for learning, per a Hearthstone card — HanchungLee · 2026-10-14
- 14-year-old GPU rigid body pipeline revived in CUDA simulates 2.05M dominoes — erwincoumans · 2026-10-14
- Researchers warn AI-posed humans on social media still sound like chatbots — CShorten30 · 2026-10-14