prompt engineering

Bilingual

From MIT: AI financial advice is surprisingly good
2 months ago
- AI financial advice generally promotes good financial behaviors like saving more, diversifying investments, and reducing risk with age.
- AI advice fails to properly adjust to shocks like unemployment and allows portfolios to drift without active rebalancing.
- The quality of AI advice improves when prompts are detailed and structured, but regular users often ask vague questions.
- AI advice varies based on user gender, financial literacy, and prior AI experience, leading to wealth gaps of up to $100,000 by age 60.
- Differences in prompts and model biases contribute to variation in advice, with women and less financially literate users receiving less favorable recommendations.
- AI can serve as a low-cost supplement to human financial advisors, especially for those with limited resources.
- For businesses, AI advice may shift how consumers discover financial products, reducing reliance on traditional marketing.
Prompt Engineering is for Transactional Prompting
2 months ago
- Mitchell Hashimoto distinguishes two ways of using language models: interactive prompting (ChatGPT-style, human-driven discussions) and transactional prompting (treating the model like a function, software-driven, high volume, specific tasks).
- Prompt engineering is primarily about transactional prompting, aiming for accuracy and low cost, and its main benefit lies in transactional use cases rather than interactive ones.
- Another key dimension is objectivity vs. subjectivity: prompt engineering works well for objectively correct outputs (e.g., information extraction, classification) but is less effective for subjective tasks (e.g., creative writing, art generation).
- Code generation can be both objective (small, well-specified tasks) and subjective (larger, less specific tasks), with prompt engineering more applicable to the former.
- Prompt engineering for interactive prompting is like being an effective Googler, while for transactional prompting it resembles being a data scientist.
- Agents, though automated interactive processes, are primarily transactional in nature, so prompt engineering applies to them as well.
The famous o3 "GeoGuessr" prompt did not work
2 months ago
- The elaborate 'GeoGuessr' prompt for OpenAI's o3 model did not improve geolocation performance compared to a simple default prompt.
- A benchmark of 200 images showed the default prompt performed slightly better in median and mean distance from actual location.
- The author warns that prompt engineering can create a false sense of improvement, as models often rationalize their own mistakes.
- The benchmark was easy and cheap to run ($15, six hours), yet no one had tested the prompt's effectiveness during the initial hype.
- o3's geolocation abilities have not transferred to newer models like GPT-5.4 and GPT-5.5, which performed worse in the same benchmark.
- The author addressed concerns about images being in training data, noting that public domain images still allow useful comparisons.
Nano Banana Pro is the best AI image generator, with caveats
2 months ago
- Nano Banana and Nano Banana Pro are different models; Nano Banana Pro is not a replacement but a complementary model with higher costs and specific strengths.
- Nano Banana Pro excels at style transfer, code generation with syntax highlighting, webpage rendering, and text rendering with high accuracy.
- Nano Banana Pro offers higher resolution output (2K/4K) and uses a stronger text encoder based on Gemini 3 Pro, which includes a mandatory thinking step for prompt understanding.
- Nano Banana Pro supports Grounding with Google Search for factual accuracy in infographics, but results may still have errors and require verification.
- System prompts are functional in Nano Banana Pro, allowing style enforcement, but are less useful for per-image customization.
- Nano Banana Pro can generate multi-image grids efficiently, enabling cost-effective exploration of variations, though very large grids can lead to quality degradation.
- The model tends to push prompts toward realism due to its thinking process, which can reduce effectiveness for surreal or stylized content.
- Increased cost and generation time are major limitations for frequent use, and the mandatory thinking may hinder unconventional outputs.
Prompt Engineering vs. Blind Prompting
2 months ago
- Prompt Engineering is a systematic, experimental methodology for creating prompts to reliably extract information from language models, differing from 'Blind Prompting' which relies on crude trial-and-error.
- The process starts with defining a problem, building a demonstration set for testing, generating prompt candidates, and measuring accuracy with metrics like percentage correct and token usage.
- Prompt selection involves a cost vs. accuracy analysis, considering factors like model choice and token budget, followed by continuous verification and improvement using user feedback or parsing errors.
- The post emphasizes decomposing problems for tractability, keeping outputs simple, and normalizing results, while noting that prompting is a foundational skill for higher-order LLM integrations.
A system prompt to get AI to stop pretending to be human
2 months ago
- The user dislikes AI models mimicking human social patterns like greetings and deferential phrases, finding them dystopian and unnecessary.
- The user created a system prompt to force AI to communicate neutrally and directly, like a CLI tool, avoiding all human-like social behavior.
- The prompt instructs AI to avoid discourse markers, conversational filler, evaluative acknowledgments, and rhetorical questions, stating information directly.
- AI must refer to itself as a computational entity (e.g., 'the algorithm predicted') instead of using first-person pronouns like 'I'.
- Responses should end with factual status statements or summaries, not offers or questions soliciting user input.

prompt engineering

From MIT: AI financial advice is surprisingly good

Prompt Engineering is for Transactional Prompting

The famous o3 "GeoGuessr" prompt did not work

Nano Banana Pro is the best AI image generator, with caveats

Prompt Engineering vs. Blind Prompting

A system prompt to get AI to stop pretending to be human