Groq speed lane via llm -m groq-llama-3.3-70b-versatile

Groq runs Llama 3.3 70B at ~300 tokens/sec — finishes a batch before Gemini-pro finishes the first one. Simon Willison's `llm` + `llm-groq` plugin gives you the same -p shape.

$ curl -s "https://hn.algolia.com/api/v1/search?query=duckdb&hitsPerPage=20" \
  | jq -r '.hits[] | "- [\(.points // 0)] \(.title)"' \
  | llm -m groq-llama-3.3-70b-versatile \
      "Three bullets: dominant claim about DuckDB on HN, one underrated post (cite title), overall sentiment."

1. [01] curlcurl … hn.algolia.com/api/v1/search?query=duckdb …
   20 highest-relevance HN stories mentioning DuckDB.
2. [02] jqjq -r '.hits[] | "- [\(.points // 0)] \(.title)"'
   Tight bullet list — points + title per line.
3. [03] llmllm -m groq-llama-3.3-70b-versatile
   Simon Willison's `llm` CLI with the `llm-groq` plugin. Same `stdin → -p prompt` shape as gemini/claude. Models: `groq-llama-3.3-70b-versatile`, `groq-llama-3.1-8b-instant` (fastest), `groq-mixtral-8x7b-32768`.
4. [04] llm"Three bullets: …"
   Last positional arg = prompt. stdin is appended automatically.

- **Dominant claim**: DuckDB has displaced Pandas for laptop-scale analytics — multiple posts cite 10x speedups on identical queries.
- **Underrated post**: "Querying NYC taxi data from a $5 VPS" (47 points) — quietly shows DuckDB's HTTP-streaming Parquet trick at scale.
- **Sentiment**: pragmatically bullish — builders, not hype.

• Groq is the speed king for first-pass classification/summarization, not deep reasoning. Use it for fan-out, hand the consolidated result to claude/gemini-pro for synthesis.
• Free tier: ~30 req/min, ~14k tokens/min — fine for personal use.