Feb 28, 2023 · 20m · y-combinator
The REAL potential of generative AI · Y Combinator
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Y Combinator Podcast episode, host Ali Rowghani interviews Raza Habib, Co-Founder and CEO of Humanloop, exploring how developers build commercial applications with large language models. The conversation covers fine-tuning workflows, prompt context grounding, evolving software engineering roles, AI safety, and the roadmap toward Artificial General Intelligence.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The partners as informed peer 3.8 Guest teaching 6.3 Guest disagreement 1.2 The partners pushing back 1.5
speaking balance: gold is the partners, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 14:58 Raza challenges the general model moat hypothesis
Raza firmly disagrees with the premise that general foundation model creators have an insurmountable feedback moat, arguing feedback fine-tuning only confers advantages in narrow domains.
Hardest push from the partners ▶ 14:47 Ali challenges Raza on data network effectsAli directly challenges Raza's skepticism, pressing him on why a two-year lead and thousands of production applications wouldn't generate a dominant data flywheel.
Biggest teaching moment ▶ 4:37 Raza explains RLHF vs base modelsRaza clearly breaks down how InstructGPT and RLHF enabled smaller models to outperform 100x larger unaligned models, detailing the technical mechanisms behind ChatGPT's breakthrough.
The partners hold their own ▶ 14:47 Ali presses on developer ecosystem moatAli demonstrates domain intuition by articulating a concrete counterargument regarding accumulated application data advantages.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The partners as informed peer | Guest teaching | Guest disagreement | The partners pushing back | Why |
|---|---|---|---|---|---|---|
| Defining Large Language Models and Scaling | 3 | 7 | 1 | 1 | Ali asks introductory high-level questions about what large language models are and why they have exploded in popularity. Raza provides deep educational explanations on statistical prediction, scaling laws, and hallucination reduction via context injection. | |
| Understanding Fine-Tuning and RLHF | 4 | 7 | 1 | 1 | Ali prompts Raza to explain fine-tuning and provides conversational examples like email logs. Raza gives a thorough technical explanation of instruction tuning, RLHF, and Anthropic's scalable automated feedback methods. | |
| Production Data Capture and Humanloop Fine-Tuning Demo | 4 | 6 | 1 | 1 | Ali asks practical developer questions about fine-tuning pipelines and summarizes the key stages. Raza outlines Humanloop's workflow addressing prototyping, evaluation, and customization. | |
| How Generative AI Changes the Developer Role | 3 | 7 | 1 | 1 | Ali inquires about future developer workflows and upcoming technical breakthroughs. Raza explains how LLMs augment coding today and why developers might be among the first knowledge workers automated under AGI. | |
| AI Safety, Ethics, and Existential Risk | 6 | 6 | 3 | 5 | Ali pushes back directly against Raza's skepticism regarding OpenAI's data flywheel advantage, arguing that a two-year head start and thousands of apps provide a strong moat. Raza counters that feedback data is hard to maintain across a generalized model without performance degradation. | |
| The Startup Explosion and Building Unique AI Products | 3 | 5 | 0 | 0 | Ali asks about startup opportunities and invites Raza to pitch Humanloop's open hiring roles. The discussion is entirely collaborative and supportive. |
Statements from this episode (14)
Habib: Scaling LLM parameters and data consistently improves prediction accuracy
“As you scale the language models, both in terms of the number of parameters they have, but also in the size of the data set that they're trained on, it turns out that they continue to get better and better at this prediction task.”
Habib: Next-word prediction at scale forces LLMs to develop reasoning
“They are able to do this task extremely well. And the only way to do that is to have gotten better at under, you know, some form of reasoning and some form of knowledge.”
Habib: ChatGPT's obsequious tone drove the need for custom, use-case-specific models
“I think when ChatGPT came out, there was a lot of frustration from people who didn't like its personality. The tone was a bit obsequious and it's, you know, it'll defer. It doesn't want to give strong opinions on things. And to me that demonstrates the need fo…”
Habib: Fine-tuning is what separated ChatGPT from earlier GPT models
“If you look at what the difference is between ChatGPT or the most recent OpenAI Text DaVinci Three model, and what's been in the platform for two years and has not gotten as much attention, the difference is fine tuning. Like it's the same base model, more or …”
Habib: OpenAI's 1.3B InstructGPT beat the 100x larger GPT-3 via RLHF
“In the InstructGPT paper that OpenAI released, they compared, you know, a one or two billion parameter model with instruction tuning and RHF to the full GPT-III model and people preferred that despite the fact it was a hundred times smaller.”
Habib: Anthropic matched RLHF performance using AI-generated feedback
“Anthropic had this very exciting paper just a couple of weeks ago where actually we're able to get similar results to RLHF without the H. So just actually having a second model provide the evaluation feedback as well. And that's obviously a lot more scalable.”
Habib: User Edits and Send Actions Serve as Model Improvement Signals
“They probably edit it, so you can capture the edited text, and they maybe get a response or they don't get a response. So all of those bits of feedback are things we would capture and then use to drive improvements of the underlying model.”
Habib: LLM evaluation is harder than traditional ML due to subjectivity
“Then the use cases that people are building now tend to be a lot more subjective than you might have done with machine learning before. And so evaluation is a lot harder. You can't just calculate accuracy on a test set.”
Habib: Senior developers benefit more from GitHub Copilot than junior developers
“One thing that is surprising to me is that the people who say to me they use it the most are some of the people I consider to be better or more senior developers. You might've thought this tool would help juniors more, but I think people who are more accustome…”
Habib: Software engineers will be among the first jobs automated by AGI
“When we do get towards things that look like AGI, I suspect that developers will actually be one of the first jobs to see large fractions of their job be automated, which I think is very counterintuitive but also predicting the future is hard, so.”
Habib: Action-Taking Capabilities Shift LLMs from Text Generators to Agents
“One thing that I'm really excited about is actually augmenting large language models with the ability to take actions. And so we've seen a few examples of this at the startup called adapt AI that are doing this and a few others where you essentially let the la…”
Habib: AI models bake in creator and dataset biases
“The models bake in biases and preferences that were in the model and the data and the team that built it at the time that it was being constructed.”
Habib: Foundation model barriers are capital and talent, not secret sauce
“Like to me, the barriers to entry of training, one of these models are mostly capital and talent. Like the people needed are still very specialized and very smart and you need lots of money to pay for GPUs. But beyond that, I don't see that much secret sauce, …”