Deep Research

[State of Post-Training] From GPT-4.1 to 5.1: RLVR, Agent & Token Efficiency — Josh McGrath, OpenAI

Dec 31, 2025 · 27:34

Josh McGrath, an OpenAI post-training researcher, states that real post-training innovation lies in data quality and signal trust, not optimization methods, with RLVR and token efficiency central. He moved from pre-training (3% compute gains) to post-training (40% behavior change), describes the infrastructure chaos of RL runs, and cites GRPO from DeepSeek Math as underappreciated for providing verifiable reward signals. GPT-5 to GPT-5.1 bumped evals while slashing tokens, emphasizing token efficiency over wall-clock time. The shopping model features interruptibility and chain-of-thought transparency, and personality toggles (Anton vs Clippy) are a key differentiator. For long context, he argues agents with graph walks may be more important than 10M-token windows. He concludes that the education system fails to produce people skilled in both distributed systems and ML research, a critical combination as bottlenecks shift.

The RLVR Revolution — with Nathan Lambert (AI2, Interconnects.ai)

Jul 31, 2025 · 1:19:00

Nathan Lambert from AI2 explains the RLVR (Reinforcement Learning with Verifiable Rewards) revolution, arguing that verifiable rewards for math, code, and instruction-following are scaling more reliably than human feedback, and that open models like Tulu 3 can match frontier labs on core evals with just 10-15 tasks vs. hundreds. He traces overoptimization through three phases—control, RLHF, and RLVR—and warns that models learn to cheat unit tests unless reward design penalizes it. He analyzes o3's search-heavy approach (e.g., 80 websites per query), hybrid reasoning models like Gemini 2.5 and Claude, and predicts that pure reasoning models will become the default as inference costs drop. Lambert introduces a four-skill taxonomy for agent models—skills, calibration, strategy, abstraction—and says planning improvements often come from mundane data fixes. He advocates for fully open models, citing AI2's Olmo 32B as approaching GPT-4 level, but notes that building an 'American DeepSeek' requires massive resources and nonprofit constraints.

AI is Eating Search

Jul 23, 2025 · 56:22

Robert McCloy of Scrunch AI argues that AI is replacing web browsing, not just search, as consumers prefer ChatGPT and similar tools over traditional browsing. He explains that AI search platforms still rely on traditional search under the hood, but re-rank results based on metadata clarity—so clickbait titles fail while descriptive content wins. Prompt injection works currently but is a temporary tactic; the durable strategy is clear, structured content with server-side rendered HTML because AI tools don't execute JavaScript. Scrunch helps brands optimize for platforms like ChatGPT, Perplexity, and Meta AI; customer Clerk saw a 6x traffic increase and 9x conversion lift from AI-referred visitors. Robert predicts the future is about 'agent experience'—optimizing how AI systems interact with your content to serve end users directly.

Scaling Test Time Compute to Multi-Agent Civilizations — Noam Brown, OpenAI

Jun 19, 2025 · 1:17:47

Noam Brown, OpenAI researcher behind Cicero and reasoning models, joins hosts Alessio and Puix to discuss how test-time compute scaling drives multi-agent AI and the limits of the system 1/2 analogy. He explains that his Diplomacy bot Cicero reached top 10% of humans in 2022, and later he won the 2025 world championship. Brown argues that reasoning models like o3 succeed in unverifiable domains (e.g., Deep Research) and that harnesses and routers will be washed away by scale. He reveals OpenAI's multi-agent team is pursuing a principled approach, not heuristic, and envisions AI civilizations collaborating. He also shares his coding stack (Codex, Windsurf) and predicts test-time compute will hit cost and wall-clock bottlenecks.

ChatGPT Codex: The Missing Manual

May 16, 2025 · 53:32

Josh Ma and Alexander Embiricos from OpenAI introduce ChatGPT Codex, a cloud-hosted autonomous software engineer that runs in its own sandbox, arguing it delivers value through single-shot, long-horizon task completion. They explain that Codex evolved from internal experiments (codenamed WHAM) and now offers a 60-tasks-per-hour rate limit with a one-hour hard cutoff, emphasizing it is not meant for interactive use but for delegation—users fire off tasks and go about their day. Best practices include using agents.md, installing linters and formatters, keeping code modular, and having good architecture. The team intentionally kept it a research preview to gather feedback on environment customization and pricing, with plans to integrate seamless transitions between cloud and CLI.

⚡️Open Questions in Agentic RL — Will Brown (Prime Intellect)

May 9, 2025 · 17:13

Will Brown of Prime Intellect maps the frontier of open-source agentic reinforcement learning, arguing that training models like o3 requires multi-turn tool use, intermediate reward verification, and efficient context management. He outlines challenges such as scaling tool calls to hundreds (e.g., Deep Research's 100 calls), assigning credit across long chains, and avoiding context blow-up from websites or images. Key solutions include async RL pipelines to hide compute inefficiencies, offloading sub-tasks to small models as tool calls, and using reasoning reward models for step-level verification. Brown highlights progress in model merging for decentralized skill acquisition, where specialized models trained on code, math, or other tasks can be weight-averaged effectively. The talk concludes that while hard, these problems are solvable via open-source recipes, infrastructure, and community collaboration.

Building Manus AI (first ever Manus Meetup)

Mar 27, 2025 · 48:59

Tao Zhang, co-founder of Manus, explains how Manus gives LLMs 'hands' to take real-world actions, inspired by MIT's 'mens et manus' (mind and hand). Manus assigns each task its own cloud virtual machine via E2B, provides pre-paid data APIs (stock, social media), and a knowledge system to remember user preferences. Their previous product, Monica.im, grew to 20M monthly active users and $50M ARR before they pivoted from a failed AI browser project to Manus. On the GAIA benchmark, Manus achieves a per-task cost of ~$2, far cheaper than OpenAI's Deep Research (~$20) and prior SOTA. Zhang emphasizes a 'less structure, more intelligence' philosophy, opting not to predefine workflows but to give the model rich context and tools. He also discusses plans to partner with Cloudflare, pay paywalls on behalf of users, and integrate more APIs based on usage patterns, while ruling out building their own foundation model.

Why is everyone cloning Deep Research?

Feb 18, 2025 · 1:00:10

In this episode, hosts Alessio and Swyx interview Aarush Selvan and Mukund Sridhar, the lead PM and tech lead for Gemini Deep Research, about how Google created the deep research agent category and the technical decisions behind it. They finetuned a special version of Gemini 1.5 Pro for the feature, which autonomously browses the web for about five minutes to produce a research report. The team built an asynchronous platform to handle long-running tasks and developed an ontology of research behaviors—from broad shallow exploration to deep focus—to evaluate performance. They found that users value the visible browsing process even when it takes minutes, contrary to typical latency norms. The discussion covers iterative planning, trade-offs between model memory and web grounding, and future directions including thinking models, multimodality, and personalization.

smol agents are all you need

Feb 13, 2025 · 22:31

This episode features Aymeric Roucher of Hugging Face discussing smol agents, a lightweight agent framework, and argues that code-based agents outperform JSON-based agents. Roucher explains that agents are defined by LLM control over execution flow, and smol agents simplifies this with a core file under 1000 lines. He describes the origin from transformers.agents and the security-focused custom Python interpreter. The library achieved third place on the GAIA benchmark, using o1 and Sonnet models, and Roucher projects 90% accuracy by 2026 based on a sigmoid trend line. Hugging Face is launching an agent course and plans to build GUI agents for computer use, while also fine-tuning R1 for agentic tasks.

The AI Architect: Bret Taylor

Feb 11, 2025 · 1:35:59

Bret Taylor, CEO of Sierra and chairman of OpenAI, recounts engineering AI's future across three eras — rewriting Google Maps' frontend in a weekend (cutting un-gzipped code to 20K), building Sierra's in-house agent platform for consumer brands like Sonos and SiriusXM, and navigating OpenAI's leadership crisis. He argues that AI agents' core abstractions are still pre-React, compares today's agents to jQuery, and predicts that outcome-based pricing will dominate as agents complete jobs rather than assist humans. He urges software engineers to move up the stack toward domain-specific agents and to embrace Rust-like languages for AI-generated code, emphasizing formal verification and safety. Taylor also shares how OpenAI prioritizes mission over product features, and why the relationship with Microsoft remains OpenAI's most important partnership.