Things we learned about LLMs in 2024

Energy use and climate impact

• Several comments link 2024’s LLM boom to a surge in methane/gas power plants, arguing AI is extending fossil fuel lifetimes when emissions should be falling.
• Others say AI is just one of many growing loads (EVs, reshoring, population) and that gas is still better than coal per kWh, though methane leakage may erase much of that benefit.
• Some call for strict rules that data centers use renewables or fully internalize their carbon and water costs; a carbon tax (with either green subsidies or per‑capita rebates) is widely endorsed but seen as politically hard and needing global coordination.

Economics, business models, and AGI speculation

• Thread debates whether current model inference is sold below energy cost; one correction says cheap models like Gemini/Nova at least cover energy, possibly helped by subsidies.
• OpenAI’s very high valuation is questioned; some think it assumes regulatory capture or winning an “AGI race.” Others argue AGI, if real, would commoditize everything and erase most individual AI firms’ moats.
• There’s broad agreement that model quality is converging and open weights (e.g., Llama-family) plus many hosts will push inference prices down, making long‑term margins thin.

Usefulness, slop, and criticism quality

• Many see LLMs as power-user tools: extremely helpful but unreliable, requiring good prompting, context management, and manual verification.
• Others report frequent hallucinations, shallow or incorrect summaries, and flood of low‑value “slop” content, especially when users are lazy or indiscriminate.
• Some argue that dismissing LLMs outright is a mistake; what’s needed is better, more specific criticism and clearer guidance on where they work and where they don’t.

Coding and developer experience

• Strong split: some report “spookily good” productivity gains (fast scaffolding, bug-spotting, DSL snippets, ad‑hoc tools, refactors); others see subtle bugs, fake APIs, and degraded code quality from overreliant colleagues.
• Consensus that LLM-written code must be tested and code‑reviewed; they’re likened to overconfident junior devs.
• Different models perform better in different stacks; Python/JS/React praised, Rust and some math-heavy or niche areas fare worse.

Agents, tooling, and local models

• “Agents” is viewed as poorly defined marketing jargon; suggestions range from “multi-step workflows using tools” to “semi-autonomous software with goals.”
• People like editor/CLI integrations and custom scripts for feeding codebases/docs into models.
• Local models on high-RAM Apple laptops impress some, but GPU VRAM and power limits keep best models in data centers for now.

Social impacts and governance

• Concerns include job displacement (especially knowledge workers), worsening inequality, content authenticity, medical/misinformation risks, and climate trade-offs.
• Others highlight benefits: letting people ask “stupid” questions without judgment, tutoring, therapy‑like conversations, and faster access to complex information.