December 5, 2025
Is your AI bulking or cutting?
Why are your models so big? (2023)
Big-brain bots vs pocket-size helpers — the internet picks sides
TLDR: A viral post argues simple tasks shouldn’t need giant chatbots, pushing tiny, cheap models that run locally. Commenters split: some say big models are required for solid logic, others demand small tools that run on a Raspberry Pi, with jokes and cost anxiety fueling the fight.
The post Why are your models so big? dared to ask the question everyone’s wallets are screaming: why do today’s chatty AIs—aka large language models—need to be so huge? The author says simple jobs like reminders or SQL autocomplete don’t need billions of “brain cells,” and points to tiny models you can even run in your browser. Cue the comment brawl.
On Team Big Brain, user siddboots drops a sober reality check: logic is hard, even for “simple” tasks, and size matters because models learn that logic from a massive training diet. Meanwhile, semiinfinitely turns it into a roast: a parody about “laptops being too large” that hilariously mirrors the whole argument. Team Tiny fires back. unleaded insists people are “begging a model 20x the size” to spit out basic structured text and suggests old-school text completion beats pricey chatbots for narrow tasks. lsb sets the practical vibe: their rule of thumb is “can it run on a Raspberry Pi,” bragging a 4-billion-parameter model runs comfortably on one. And then debo_ brings the meme: “2000: My spoon is too big — 2023: My model is too big.”
The vibe? Cost vs capability, jokes vs benchmarks. Fans of smaller, cheaper tools cheer for Phi-2 and browser-friendly setups; big-model believers warn tiny brains can fumble logic. Delicious drama, with a side of Raspberry Pi flex.
Key Points
- •The article questions why contemporary LLMs are very large, noting even “small” models like Phi-2 have 2.7B parameters.
- •It argues many tightly scoped tasks (e.g., SQL autocomplete, structured extraction) do not require large models or broad linguistic knowledge.
- •Inference for large models is costly in compute and operational complexity, with external sources cited to support this.
- •The author advocates for smaller, task-specific models to reduce cost and complexity.
- •Existing tools (e.g., llama2.c, transformers.js) enable building and running smaller models, including in-browser deployment.