Amber Teng

AI application developer and analyst of large language models

Mentions

Appeared in

Discussed in

Key positions and views

Developing applications with LLM APIs like OpenAI's is incredibly fast, with a functional prototype achievable in hours, not weeks.

Effective prompt engineering is a nuanced skill; prompts must be carefully balanced, as being too simple or too specific leads to model failure.

Fine-tuning model parameters such as temperature, frequency penalty, and presence penalty is critical for controlling the creativity, coherence, and quality of AI-generated text.

Out-of-the-box LLMs are unreliable for production use without human oversight, frequently fabricating information and producing low-quality content that requires significant manual correction.

There are significant ethical concerns in the AI space, including the practice of overcharging for simple AI wrappers and the potential for AI-generated content to flood and degrade systems like the job market.

Podcast consensus on Teng

Points of consensus

▶LLM APIs, such as OpenAI's, facilitate extremely rapid application development, allowing a functional prototype to be built and deployed in a matter of hours.May 2026

▶Achieving high-quality, reliable outputs from models like GPT-3 is non-trivial, requiring nuanced prompt engineering and careful tuning of parameters like temperature, frequency penalty, and presence penalty.May 2026

▶Out-of-the-box LLMs exhibit significant flaws, including frequent factual inaccuracies (hallucinations), generation of incoherent text, and a general need for human oversight and manual editing to produce usable results.

▶The specific choice of LLM (e.g., DaVinci vs. Curie) directly impacts the output's quality, format, and length, due to fundamental differences in token limits and the recency of their training data.May 2026

Points of debate

▶There is a tension between the remarkable speed of LLM application development (a project completed in five hours) and the poor quality of the resulting output (only 1 in 10 results were usable after manual tweaking).

▶A conflict exists between the extremely low operational cost of running an AI application (less than $2 for a month of public use) and the ethical concern about companies charging high prices for similar simple AI wrappers.May 2026

▶Teng's experience shows a contrast between the model's powerful generative capabilities and its unreliability, as it can either create novel text or fall into failure modes like fabricating user details or copying input text verbatim.May 2026

▶The model's creativity, controlled by the 'temperature' parameter, is presented as both a desirable feature for generating novel ideas and a risk factor that contributes to a 'danger zone' of incoherent outputs.May 2026

Key themes

▶Rapid Prototyping and Accessibility of LLMs

Teng's experience building a cover letter generator in just five hours highlights the dramatically reduced barrier to entry for creating AI-powered applications. The availability of tools like the OpenAI Playground and simple APIs enables developers to move from idea to a deployed prototype with unprecedented speed.

This accessibility suggests a future proliferation of niche, single-purpose AI tools, but also indicates that sustainable competitive advantage will likely come from proprietary data, unique user experience, or sophisticated backend processing rather than simply wrapping a public API.

▶The Nuance and Necessity of Prompt EngineeringMay 2026

Teng details the delicate balance required for effective prompting, where overly simple prompts lead to generic falsehoods and overly specific ones cause the model to copy input verbatim. Her work underscores that interacting with LLMs is a nuanced skill involving iterative testing and precise parameter tuning to guide the model's output.

Prompt engineering is emerging as a critical human-in-the-loop skill, meaning the value of AI tools is heavily dependent on the operator's ability to effectively communicate context and constraints to the model. This creates a new layer of technical expertise that companies must cultivate.

▶Practical Limitations and Unreliability of LLMsMay 2026

Despite the technology's power, Teng's project revealed significant flaws in GPT-3. The model frequently produced factual errors, assigned incorrect attributes to users, and generated incoherent text, with a low success rate of only one in ten outputs being 'sendable' even after manual edits.

Analysts should remain skeptical of claims of full automation, as the 'last mile' problem of ensuring factual accuracy and quality control remains a major hurdle. The high rate of failure suggests that for many use cases, LLMs are better suited as assistants that augment human work rather than fully replacing it.

▶Emerging Business Ethics and Market DisruptionMay 2026

Teng raises ethical questions about the business models built on LLMs, specifically criticizing companies that charge high prices for simple API wrappers. She also speculates on the disruptive potential of the technology to flood systems like the job market with low-quality, AI-generated applications, creating new challenges for recruiters.

The low marginal cost of AI generation will likely lead to both market saturation and malicious use cases. This will force platforms and industries to develop new methods for detecting automated content and will create a market demand for services that can verify authenticity and quality.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Project Inception

Teng initiated her first LLM project, a resume cover letter generator, using OpenAI's GPT-3 API and the $18 in free credits provided to new users.

Initial Development

She experienced a surprisingly quick ramp-up time, creating a functional application with only a few lines of code, utilizing the OpenAI Playground for code-free prompt testing.

Model Selection and Prompt Engineering

Teng selected the DaVinci model for its larger token limit and more recent training data. She iteratively refined her prompts, discovering that both overly simplistic and overly specific inputs resulted in poor outputs.

Parameter Tuning

She adjusted parameters like 'temperature', 'frequency penalty', and 'presence penalty' to improve output quality, identifying a 'danger zone' where a high temperature combined with a long text request led to incoherent results.

Deployment and Performance Review

The application was deployed on Streamlit, handling 50-100 daily queries for under $2 in API costs over a month. Teng concluded that only 1 in 10 generated letters were 'sendable' after manual tweaking.

Broader Reflection

Following the project, Teng articulated ethical concerns about overpriced AI wrapper apps and the potential for AI-generated spam to disrupt the job market.

Suggested prompts

How does Teng's experience with the 1-in-10 'sendable' output rate inform the viability of fully automated content creation tools for professional settings? &nearr;Based on Teng's analysis of prompt engineering, what are the key trade-offs between prompt simplicity, specificity, and the use of parameters like 'temperature' in achieving desired AI outputs? &nearr;Teng raises concerns about AI-generated job application spam. What technical or policy-based solutions could mitigate this potential negative externality of accessible AI? &nearr;Given the low operational cost (<$2/month) of Teng's application, how does this challenge traditional SaaS business models and what new monetization strategies might emerge for simple AI 'wrapper' applications? &nearr;

Key concepts

OpenAI 1 ep GPT-3 1 ep Cover Letter Generator 1 ep Prompt Engineering 1 ep Model Parameters (Temperature, Penalties) 1 ep DaVinci Model 1 ep AI Ethics 1 ep Development Time & Cost 1 ep Curie Model 1 ep

Notable quotes

“I think like five hours is it's incredible.”

Amber Teng · Amber Teng - Building apps with a new generation of language models

“if it was too specific, it would take the text verbatim in my in my prompt to print that out. And I was like, OK, well, the whole reason I want this cover letter generator is so I don't have to write it.”

Amber Teng · Amber Teng - Building apps with a new generation of language models

“So would you say it's like the combination of high temperature plus a longer bit of text that's generated? That's the danger zone? Okay. Yeah, I think so, at least for me.”

Amber Teng · Amber Teng - Building apps with a new generation of language models

“Common errors made by GPT-3's DaVinci model in the cover letter generator included assigning an incorrect gender to the user or fabricating user experiences, such as attending MIT.”

Amber Teng · Amber Teng - Building apps with a new generation of language models

Report last updated: May 5, 2026

Get started free

Back to Entities Intelligence Report