GPT-4o

Why AI Agents Don't Actually Understand You — Danielle Perszyk, Amazon AGI Lab

Jul 11, 2026 · 48:54

Danielle Perszyk of Amazon's AGI Lab argues that AI agents fail because they don't truly understand users — they model tasks but not the human mind, so reliability must shift from clicking correctly to aligning representations. She explains how Amazon's AGI Lab, seeded by the former Adept team, builds perception agents that perceive digital environments like humans and interact in real time, moving beyond chatbots and coding agents. The lab is exploring new architectures for episodic memory, social world models, and multi-agent systems where agents fluidly negotiate meaning, inspired by human collective intelligence. Perszyk warns that current AI homogenizes thinking, reducing human agency, and advocates for a diverse society of AIs with different biases to augment rather than replace human cognition. She discusses how aligning AI's goals with inferring and matching human representations could unlock generalization and make agents genuinely collaborative, from automating digital drudgery to transforming education with Socratic tutors.

How Zyphra went all-in on AMD + Why Devs feel faster with AI but are slower — with Quentin Anthony

Nov 3, 2025 · 1:00:02

Quentin Anthony, head of model training at Zyphra and advisor at EleutherAI, explains why his company moved all training to AMD MI300X GPUs, outperforming Nvidia H100s on certain workloads thanks to 192GB VRAM and higher memory bandwidth, and describes his kernel development approach of writing directly in ROCm or GPU assembly rather than using Triton. He also shares his experience in the METR study on AI coding productivity, where he was one of the few developers with measurable speedup, and offers tips: timebox AI use, avoid the slot machine effect, maintain context hygiene, and use direct API over tools like Cursor. Additionally, he argues that open source AI research benefits from siloed focused teams with guaranteed funding over grand collaborations, and notes that kernel datasets alone won't solve GPU programming due to evaluation challenges.

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.

Claude Plays Pokémon Hackathon: Escape from Mt. Moon!

Apr 5, 2025 · 1:15:00

David Hershey (creator of Claude Plays Pokémon), Andrew (a developer building a Pokémon-playing virtual streamer), and Jesse Han (Morph Labs CEO) detail how to build agents that escape Mt. Moon in Pokémon FireRed using Anthropic's Claude and Morph Cloud's Infinibranch technology. Hershey explains that Claude's vision is often unreliable—it spent 8 hours pressing A on a doormat thinking it was a dialogue box—and that spatial reasoning remains a core bottleneck, best mitigated by touchscreen controls rather than prompt engineering. Andrew reveals he relied on reading raw RAM and A* pathfinding rather than computer vision, calling it “cheating” but pragmatically useful: his biggest breakthrough was a single prompt line telling the agent to try something else if it fails to grab a starter Pokémon. Jesse introduces Morph's EVA agent framework, which uses low-overhead snapshotting for test-time search, and the hackathon's judging criteria—escape Mt. Moon in the fewest agent turns, with a $1,000 prize for the coolest use of Infinibranch branching. The episode serves as both a technical walkthrough and a challenge to build general-purpose agents that can handle complex, open-ended tasks…

Raycast: Your AI Automation Assistant

Feb 26, 2025 · 20:19

Raycast CEO Thomas introduces AI Extensions, letting users @-mention apps like Slack or Google Calendar and perform actions (e.g., set status, book meetings) via natural language, powered by a fine-tuned model called Ray 1 based on GPT-4o for faster, more accurate function calling. He shares how Raycast evolved from a Spotlight replacement to an extension platform, now adding AI chat with tool-aware reasoning. Challenges include building evals for tool calls (harder than standard LLM evals) and supporting third-party extension authors with a JSDoc-based TypeScript interface. Thomas envisions Raycast becoming an AI-native layer for macOS, eventually enabling background automations triggered by time or file events.

The Unreasonable Effectiveness of Reasoning Distillation: using DeepSeek R1 to beat OpenAI o1

Jan 24, 2025 · 23:35

Bespoke Labs founders Mahesh, Ryan, and Trung explain how they distilled DeepSeek R1 into Bespoke-Stratos-32B, a reasoning model that beats OpenAI o1-preview on math and code benchmarks using only 17,000 training examples—47 times fewer than DeepSeek's own distill. They achieved this in a 48-hour sprint leveraging their data curation library Curator. Unlike Sky-T1, which required rewriting unreliable QWQ traces, R1's coherent reasoning allowed them to skip re-annotation and simply filter for correctness. The team argues data quality matters more than quantity, and that smaller models (like their 7B variant) can also improve with better teacher data. They see this as evidence that reasoning can emerge purely from supervised fine-tuning on high-quality traces, without complex search algorithms.

Agents @ Work: Dust.tt — with Stanislas Polu

Nov 11, 2024 · 58:53

Stanislas Polu, co-founder of Dust.tt, argues that horizontal agent platforms—not vertical, single-use-case agents—will win the enterprise by letting non-technical employees build their own automation. A former OpenAI researcher who worked directly with Ilya Sutskever on formal mathematics and saw GPT-4 internally before ChatGPT, Polu left in 2022 to start Dust, now an open-source infrastructure layer that connects company data (Slack, Notion, GitHub) to LLMs and lets anyone create simple, instruction-driven agents. He contrasts Dust’s API-first, integration-heavy approach with LangChain’s open-source orchestration and Adept’s browser automation bet, insisting that most useful agent workflows require only a few tool calls and that function-calling quality (where GPT-4 Turbo still leads) matters more than model size. Dust achieves 88% daily active user penetration at some enterprise customers, and Polu predicts the first billion-dollar company run by a single person will emerge once platforms like Dust mature enough to let tiny teams scale to massive output.

In the Arena: How LMSys changed LLM Benchmarking Forever

Nov 1, 2024 · 41:02

Anastasios and Weilin from LM-Sys explain how ChatBot Arena became the de facto standard for LLM benchmarking by replacing static benchmarks with dynamic human preference evaluations. They trace the origin from Vicuna fine-tuning to the viral launch in April 2023, where anonymous side-by-side battles let the community decide which model is better. To address biases like length preference, they developed Style Control, a logistic regression method that adjusts for confounders such as response length and markdown formatting. They address the controversy around labs testing multiple private models, arguing that selection bias is empirically small and the live benchmark self-corrects over time. They also discuss RouteLLM for cost-performance routing, the graduation of ChatBot Arena from LMSys to support new projects, and call for community help with red teaming, vision modalities, and implementing a REPL for coding evaluations.

Building AGI in Real Time (OpenAI Dev Day 2024)

Oct 4, 2024 · 2:09:35

OpenAI's DevDay 2024 unveils the Realtime API with WebSocket-based voice and function calling, Vision Finetuning, Prompt Caching, and Model Distillation. Product leads Olivier Godement, Romain Huet, and Michelle Pokrass detail design choices like zero-code caching and WebSockets for human-level latency. Alistair Pullen of Cosine explains how fine-tuning GPT-4o on custom reasoning traces beat o1 on SWE-bench Verified, while Sam Altman and Kevin Weil discuss AGI levels (o1 as level 2 reasoner), the coming agent era (level 3), safety through iterative deployment, and the importance of building for the frontier. The episode also covers moderation policy relaxation, plans for o1 system prompts by year-end, and the competitive admiration for NotebookLM.

Building AGI with OpenAI's Structured Outputs API

Sep 17, 2024 · 1:12:23

Michelle, OpenAI's API tech lead, explains how the newly launched Structured Outputs feature achieves near-100% reliability for schema-following outputs by combining constrained decoding with model training, moving beyond the limitations of JSON mode and Function Calling. She details the refusal field that allows the model to decline harmful requests while maintaining parseable output, and discusses use cases like dynamic UI generation and mathematical chain-of-thought. The roadmap includes parallel function calling and custom grammars. The episode also covers the GA of GPT-4o fine-tuning, Batch API's 50% cost savings, Vision API integration with structured outputs, the Whisper API's surprising lack of diarization, and the shift toward speech-to-speech APIs. Michelle shares insights from her Waterloo co-op background and what qualities succeed at OpenAI: low ego, user-focused, and driven.

Is finetuning GPT4o worth it?

Aug 22, 2024 · 1:01:16

Ali Pullen, founder of Cosine, discusses Genie, an AI software engineering colleague that achieved 30% on SWE-Bench and 43.8% on SWE-Bench Verified via fine-tuning GPT-4o on billions of tokens of synthetic data. Cosine built Genie by training on the process of software engineering—not just working code—using synthetic runtime errors and a self-improvement loop. The agent follows a four-stage workflow: code retrieval (66% accuracy), planning, writing diffs, and running CI tests. Pullen reveals that model performance degrades linearly beyond 60K tokens in the context window, and that OpenAI dynamically sizes LoRA adapters to handle Cosine's massive dataset. He explains why Genie isn't on the SWE-Bench leaderboard (refusing to publish trajectories to prevent distillation) and outlines plans to scale the dataset and fine-tune on customer codebases for personalized performance.

The 10,000x Yolo Researcher Metagame — with Yi Tay of Reka

Jul 5, 2024 · 2:18:43

Yi Tay, chief scientist at Reka and former Google Brain architecture co-lead on PaLM 2, recounts his journey from an NTU PhD to co-founding Reka, where a team of just 5 people pre-trained Reka Core on $60M of GPU compute—debating at #7 on the LMsys leaderboard. He explains why he believes the Noam transformer (with SwiGLU, GQA, RoPE) remains the strongest baseline, how encoder-decoder architectures offer 'free sparsity' (a 2× flop-efficiency over decoder-only), and why Chinchilla scaling laws are often misunderstood (training past the optimal compute frontier is routine). He argues that long context will eventually outperform RAG for complex reasoning, and that open-source models like Llama 3 are catching up only because Meta invested in a top-tier training stack, not due to grassroots innovation. The episode also covers his productivity habits (working backwards from a paper title, camping arXiv) and the cultural shock of moving from Singapore academia to Google's impact-driven research environment.