LLMs reward expertise

16 hours ago

LLMs make everyone a generalist, but domain expertise remains crucial for effective prompting.
Terence Tao's conversation with ChatGPT shows expertise enables concise, targeted queries and critical evaluation of responses.
Experts can push LLMs harder by suggesting alternatives, identifying flaws, and steering toward specific solutions.
Domain knowledge is more valuable than generic principles for system design and extracting value from LLMs.
Human expertise remains a bottleneck as models improve; skilled humans are needed to communicate exact requirements.

LLMs make everyone a generalist by enabling tasks like writing CSS without deep expertise.
Domain expertise is the most critical skill for effective prompting, not generic prompting techniques.
Terence Tao's conversation with ChatGPT shows expert prompting: short messages, concise outputs, pushback, and self-directed leaps.

The traditional advice to specialize is being challenged by a new era where generalists, empowered by AI, can solve problems across multiple disciplines.
AI enables one person to carry a problem further by gaining competence in adjacent areas, reducing the need for handoffs between specialists.
Useful generalists are those who can stay with a problem, understand it deeply, and test solutions, rather than just knowing a little about many things.

The author prefers consulting experienced individuals over relying solely on LLMs for complex, nuanced questions.
When seeking expert advice, responses often redirect to asking Claude, even after thorough LLM research.
Personal experience and subjective judgment are valued over generic or consensus-based information.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

<antirez>7 days ago

The author, an experienced programmer, uses LLMs like ChatGPT to accelerate coding by reducing time spent on documentation, learning complex APIs, and creating disposable programs.
LLMs are often described as a 'stupid but all-knowing savant'; they excel at providing knowledge but struggle with complex reasoning, making them most useful for high-level tasks in Python rather than system programming in C.
Examples include using GPT4 to write Python scripts for data analysis, convert between machine learning frameworks (e.g., Keras to PyTorch), and create quick utilities like CSV report analyzers.

LLMs make everyone a generalist by enabling tasks like writing CSS without deep expertise.
Domain expertise is the most critical skill for effective prompting, not generic prompting techniques.
Terence Tao's conversation with ChatGPT shows expert prompting: short messages, concise outputs, pushback, and self-directed leaps.

The traditional advice to specialize is being challenged by a new era where generalists, empowered by AI, can solve problems across multiple disciplines.
AI enables one person to carry a problem further by gaining competence in adjacent areas, reducing the need for handoffs between specialists.
Useful generalists are those who can stay with a problem, understand it deeply, and test solutions, rather than just knowing a little about many things.

The author prefers consulting experienced individuals over relying solely on LLMs for complex, nuanced questions.
When seeking expert advice, responses often redirect to asking Claude, even after thorough LLM research.
Personal experience and subjective judgment are valued over generic or consensus-based information.

Invest in upfront alignment by making implicit assumptions explicit in the initial prompt to avoid unexpected AI outputs.
Restart with better context rather than steering AI when initial outputs are far off; avoid compounding errors through path dependence.
Equip AI agents with the same tools you use (CLI, MCP, browser control) to handle setup, testing, and instrumentation tasks.

<antirez>7 days ago

The author, an experienced programmer, uses LLMs like ChatGPT to accelerate coding by reducing time spent on documentation, learning complex APIs, and creating disposable programs.
LLMs are often described as a 'stupid but all-knowing savant'; they excel at providing knowledge but struggle with complex reasoning, making them most useful for high-level tasks in Python rather than system programming in C.
Examples include using GPT4 to write Python scripts for data analysis, convert between machine learning frameworks (e.g., Keras to PyTorch), and create quick utilities like CSV report analyzers.

Hasty Briefsbeta