Apr 24, 2025 · 30m · no-priors
No Priors Ep. 112 | With OpenAI Deep Research, Isa Fulford
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 episode of No Priors, host Sarah Guo interviews OpenAI's Isa Fulford to explore the genesis, technical architecture, and design philosophy behind Deep Research. Fulford explains how reinforcement learning, trajectory evaluation, and focused information synthesis are driving the evolution of autonomous AI toward unified digital coworkers.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The hosts hold 27.1% of the talking time here. How this is scored →
The hosts as informed peer 5.6 Guest teaching 4.8 Guest disagreement 1.3 The hosts pushing back 1.3
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:07 Rejecting manual UX controls for model thinking time
Isa directly counters Sarah's suggestion for a user toggle to control research time, explaining that forcing the user to calibrate compute is poor product UX.
Hardest push from the hosts ▶ 25:56 Insisting on time-budgeted execution togglesSarah pushes for explicit user control over compute latency, arguing she wants a direct mechanism to bound deep research execution within a five-minute window.
Biggest teaching moment ▶ 7:03 Prescribing when reinforcement fine-tuning is necessaryIsa delivers a comprehensive strategic breakdown detailing exactly when startups should invest in RFT versus when base model improvements make custom RL redundant.
The host holds their own ▶ 14:50 Connecting product evolution to retrieval originsSarah demonstrates insider familiarity with OpenAI's technical roadmap by linking the new agent architecture to Isa's foundational retrieval research.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Prioritizing Read-Only Information Synthesis Over Action Agents | 5 | 4 | 1 | 1 | Sarah frames the browsing agent landscape by contrasting trivial transactional demos with ambitious synthesis tasks. Isa explains OpenAI's technical rationale for starting with read-only multi-source research before moving to write-action agents. | |
| Data Engineering and Tool Integration for Agent Training | 6 | 6 | 1 | 1 | Sarah asks when startups should invest in reinforcement fine-tuning versus relying on traditional agent orchestration. Isa offers a detailed framework distinguishing tasks requiring RFT from those that base model scaling will naturally resolve. | |
| Leveraging Human Expertise and Emergent Planning Capabilities | 5 | 5 | 1 | 1 | Sarah probes how browsing expertise is defined and whether emergent planning surprised the team. Isa explains how RL enables models to discover planning strategies and circumvent sandbox constraints without explicit supervision. | |
| Safety Architectures, Hallucination Mitigation, and Agent Oversight | 6 | 5 | 1 | 1 | Sarah asks about classic failure modes like compounding errors and safety guardrails. Isa outlines why hallucinations are more deceptive in comprehensive outputs and distinguishes read-safety from write-action oversight. | |
| Future Product Roadmap and Compounding Capabilities | 6 | 4 | 1 | 1 | Sarah and Isa discuss future capabilities including private repository search and data analysis. Sarah demonstrates domain familiarity by referencing Isa's earlier technical work on retrieval systems at OpenAI. | |
| Agent Memory, Path to AGI, and Context Management Challenges | 6 | 5 | 1 | 1 | Sarah and Isa discuss the industry consensus on RL data efficiency and the hurdles remaining for AGI coworker agents. Isa highlights context window management and high-stakes safety with private data as primary technical blockers. | |
| Query Optimization, Reasoning Latency, and Extended Execution | 6 | 5 | 3 | 3 | Sarah asks for the mental model separating deep research queries from reasoning models and advocates for a user-facing latency toggle. Isa politely rejects the UI toggle idea, arguing the model itself must learn appropriate reasoning compute allocations. | |
| The Vision for Unified Coworker Agents and Conclusion | 5 | 4 | 1 | 1 | Sarah and Isa explore the future workflow of unified agents operating like remote coworkers on Slack. Sarah highlights the management advantage of interacting with a single, highly contextual unified agent. |
Statements from this episode (16)
Fulford: Pre-existing training datasets exist for math and coding, not browsing
“With the math and coding problems that people were already training on, those data sets already exist. You know, you can have a math problem with a ground truth answer, and you can train on those. But for browsing, it's kind of more open-ended. You don't reall…”
Fulford: OpenAI Prioritized Read-Only Synthesis Over Action Agents
“Yeah, so I think before we focused on taking right actions, which those are examples of taking right actions, we wanted to get really good at synthesizing information from a large number of sources and mostly read-only tasks.”
Fulford: Information Synthesis Is a Prerequisite for Scientific AI Discovery
“Secondly, I think the overall goal for OpenAI is to create an AGI that can make new scientific discoveries, and we kind of felt that a prerequisite to that is to be able to synthesize information. You know, if you can't write a literature review, you're not go…”
Fulford: Deep Research Began as a Prompted Demo Without Model Training
“We initially had built a demo to pitch people on this idea and it was no model training involved. It was fully just prompted models with the UI pitching the vision of what this product could look like.”
Fulford: Deep Research Uses a Text Browser and Python Execution Tool
“So right now we just have the browsing tool, which is a text based browser, but it can see embedded images and like open PDFs. And then also it has access to a Python tool so it can do analysis and calculations and plot graphs and things like that.”
Fulford: Training reasoning models on math and coding generalizes to writing
“So I think in general you will always get a model better, better at a specific task if you train on that task, but we also see a lot of generalization from training on one kind of task to, you know, other domains. So you can train a reasoning model on mostly m…”
Fulford: RFT is only worthwhile for out-of-distribution or make-or-break tasks
“I think if you have a very specific task that you think is so different to anything that the model was likely trained on and you try it a bunch of times yourself and you've tried a lot of different prompts and it's just really not good at it. So maybe it's gen…”
RL models only need task and outcome definitions to learn research trajectories
“The cool thing with RL is that you don't necessarily need to
Know the whole process of how the person would do the research.
You just have to know what the task is and what the outcome should be, and the model will just learn during training how to get from th…”
OpenAI: Deep Research learned upfront planning without explicit instruction
“We didn't teach it
To plan up front, but sometimes we'll see it does end up making a plan up front before starting its research.”
Fulford: OpenAI Deep Research attempts reward hacking around tool restrictions
“Sometimes the model will do smart things and try to get around restrictions you put on it.
So you have to make sure that it's not hacking, you know, and trying to use a different search engine other than the search engine that you gave it or something like tha…”
Fulford: Deep Research Hallucinates Less Than Any Prior OpenAI Model
“While this model is hallucinates less than any model that we've ever released, it is still possible for it to hallucinate most times because it will infer something incorrectly from one of its sources.”
Fulford: Deep Research next steps include accessing internal documentation and GitHub
“As to deep research, I think obvious next steps for deep research would also be to have access to private data, like be able to do research over, you know, any internal documentation or GitHub, whatever it is.”
Fulford: OpenAI Built Deep Research by Fine-Tuning o3
“Yeah, I think also the base model, or the model that we started fine tuning from O three is just a very capable model. It's trained on many different data sets, including a lot of coding and reasoning and math tasks.”
OpenAI to release mid-tier features bridging quick search and Deep Research
“And I think that we will release things soon that people will be happy about and we'll fill that gap.”