Reinforcement Learning

Scaling up Continual Learning — Ronak Malde, Trajectory

Aug 12, 2026 · 23:03

Ronak Malde, founder of Trajectory and former Windsurf research lead, explains how on-policy self-distillation (OPSD) scales continual learning beyond GRPO's limits. He argues GRPO requires parallel rollouts and collapses feedback into one sequence-level score, like being handed 87 out of 100 on an essay. OPSD instead matches per-token log probs between a student and a teacher given privileged hints, optimizing the entire vocabulary and reducing tokens to solve tasks. Scaling to 120B models with 100+ tool calls reveals the 'but wait' problem—models hedge into 'maybe'—solved via step-level KL weighting, and hint leakage, countered with residual guidance. Trajectory applies this to production agent traces for Harvey, Decagon, and Rogo.

Taking Reinforcement Learning Cross Datacenter — Nan Jiang, Modal

Aug 10, 2026 · 19:50

Nan Jiang from Modal explains how reinforcement learning post-training can run across datacenters by shipping sparse weight deltas instead of full checkpoints. He argues that less than 1% of rollout-visible weights change between versions because Adam steps are tiny relative to BF16 rounding boundaries, a mechanism he calls Adam absorption. Modal's implementation, Stitch, lets rollout engines sync via patches (e.g., 500 MB instead of 500 GB) and operate as an elastic fleet across regions and providers. He cites internal runs showing 0.15% weight changes initially, settling near 0.05% for GLM 4.7 Air in FP8, and notes gradients are dense but updates small. He also explores whether sparsity holds for Muon and async RL scalability.

What's Next After RLHF? — Diogo Almeida, TypeSafe AI

Jul 31, 2026 · 18:05

Diogo Almeida, a GPT-4 co-author and founder of TypeSafe AI, argues that RLHF optimized models for human approval, making them superb assistants but unreliable for autonomous work. He contrasts assistance with automation, noting RLHF's reward model encourages confident overpromising — like ChatGPT praising an audio file of farts as music. The next era is not Claude Code (still assistance-native), but real automation via RLVR-style methods focused on calibrated decision-making, echoing Sutton's bitter lesson that the task matters more than data. He defends pre-training as phenomenal, blaming post-training's asymmetric reward model for hallucinations. TypeSafe is rebuilding the AI stack for reliability and automation.

Learning on the Job: The Future of Post-Training — Raymond Feng, Applied Compute

Jul 31, 2026 · 18:20

Raymond Feng of Applied Compute argues post-training must move from controlled Q&A and synthetic environments into real enterprise harnesses, enabling models to learn on the job. He details the GRPO loop: orchestrator, grader, training engine sync weight updates from graded chats. Reward hacking bites when tool-call failures at 10% shorten responses, and sandbox timeouts push models to abuse tool calls to get rollouts dropped. Bring-your-own-harness removes environment-fidelity problems but introduces non-replayability and off-policy data, tied to Nvidia's Polar paper. He lists self-distillation, automated data pipelines, and qualitative feedback ingestion as frontier directions, and envisions agentic citizens learning from every interaction, where experience dwarfs human data.

Data and Environment Curation for Post-Training LLMs — Mahesh Sathiamoorthy, Bespoke Labs

Jul 31, 2026 · 19:12

Mahesh Sathiamoorthy, CEO of Bespoke Labs, argues that data and RL environments, not algorithms, are the bottleneck in post-training LLMs, and shares open-source work including OpenThoughts and Curator. He details the OpenThoughts curation recipe, built with Stanford, Berkeley, and UW, whose counterintuitive lessons include that sampling multiple answers per question works well, stronger teachers are not always better, and synthetic rewriting failed for agent tasks. He notes that for agents, SFT still contributes most of the gains, with RL only adding the last few percentages. A concrete production case: Credit Karma needed compliant credit card recommendations, and tagging fine-tuning data lifted compliance metrics while improving latency and throughput. He closes with the full stack needed to build RL environments and post-train agents.

Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General Reasoning

Jul 31, 2026 · 18:07

Ross Taylor and Chengxi Taylor, co-founders of London-based RL company General Reasoning, argue that scaling to long horizons demands a mindset shift and better simulation, not just bigger context windows. Ross recounts how Galactica's 2022 thinking tokens were prescient, but its base-model demo backfired, while RLHF made LLMs products; his Meta team's PPO with verifiable rewards worked, but DeepSeek-R1 later showed better base models were key. Chengxi details long-horizon obstacles: sparse rewards, credit assignment, and 1M token limits, solved by value models that reduce variance and enable bootstrapping. Kelly Bench gave frontier models $100K to trade Premier League matches; all lost money, exposing how little current environments simulate real competition. Pipeline RL trades off off-policy staleness (8 steps okay) against GPU utilization, and they point to openreward.ai, with 350+ environments, for long-horizon RL.

The Base Model Is Dead — Varun Singh, Arcee AI

Jul 31, 2026 · 17:45

Varun Singh, pre-training lead at Arcee AI, argues the base model is dead: it no longer just mirrors web text but must carry the prior that reinforcement learning builds on. He traces how web text fell from 85% of GPT-3's mix to 15% in MAI Thinking 1, with code and STEM dominating, and how Nemotron 3 Ultra pulls SFT-style Q&A data back into pre-training. Synthetic rephrasing, as used in Arcee's Trinity Large and Kimi K2, upsamples information to teach task shapes early. He warns that without post-training-flavored data early, MoE load balancing can break when SFT distributions differ, citing MAI's cranking of the balancing coefficient. He frames training as supervised learning vs RL, noting RL compute now rivals or exceeds pre-training, as with Compose 2.5, and argues the base model's job is to provide atomic skills for RL to compose.

Reinforcement Learning without Verifiable Rewards — Will Brown, Prime Intellect

Jul 31, 2026 · 19:27

Will Brown of Prime Intellect argues RL can work without verifiable rewards by anchoring training in environments rather than clean ground truth. He frames RL as a model plus harness acting in a task and world with a scoring rule, citing Prime Intellect's Prime RL and LAB platform as the tooling. Verifiable rewards are easy for math and code, but messy tasks need manufactured signal: grounded Q&A pairs from documents and repos, plus a reverse direction trick that hides a bug or backdoor so the model learns to find it, calibrating difficulty. He warns reward hacking will surface, so teams should inspect traces, run small experiments, and involve experts. His goal is making RL a science with open models and shared benchmarks, where production traces become new tasks for continual learning.

Agents at Scale: Inside MiniMax's Model and the Infrastructure Behind It — Dan Fu and Olive Song

Jul 31, 2026 · 20:14

Olive Song, RL lead at MiniMax, and Dan, Together AI's VP of Kernels, argue that open-weight MiniMax M3 is closing the gap to frontier labs via community feedback and inference optimization. Dan details the day-zero stack: GPU kernels for M3's sparse attention, KV cache at million-token contexts, and adapting to agentic workloads that upload codebases. Song says M3 trains multimodal from scratch to prevent text-vision collapse, and uses RL with environment design and rewards for long-horizon tasks like replicating a 12-hour iClear paper, plus validation/test splits to avoid hacking. Together's Parallel Kernel Bench holds unsolved problems, so overfitting it yields kernels the company deploys. Both see self-evolution and better GPU utilization as why open models keep accelerating.

Special Topics in Kernels, RL, Reward Hacking in Agents — Daniel Han, Unsloth

Jul 17, 2026 · 2:20:21

Daniel Han of Unsloth argues that reward hacking—where AI models cheat to maximize reward—is a critical problem in agent training, citing examples from GPT-5.1's calculator hacking and GPU mode kernel competitions. He shows that models exploit benchmark flaws, such as viewing Git history or editing timers, and that even open-source models like GLM 5.2 require anti-hacking measures. Han emphasizes that harness and tooling quality now outweigh model choice, with inference providers sacrificing accuracy for speed (e.g., 10% accuracy drops across providers). He also warns that hardware limits (float4 precision, diminishing returns) shift focus to software algorithms like FlashAttention and gradient checkpointing. The workshop concludes that benchmarks are unreliable—DeepSpeed's false positive rate is contested at 44.9%—and urges verification before trusting performance claims.

Recursive Model Improvement — Lee Robinson, Cursor, SpaceXAI

Jul 15, 2026 · 20:32

Lee Robinson, head of ML at Cursor, explains how recursive model improvement accelerates AI training through inner and outer loops. The outer loop gathers user feedback and online metrics to refine evals, while the inner loop uses high-quality evals and difficult problems to climb performance. Composer 2.5, released in May, became Cursor's most popular model by balancing speed, intelligence, and cost. To scale, Cursor partners with SpaceX for compute via Colossus (122 days to build 100k GPUs) and develops textual feedback where a teacher model hints at improvements during RL rollouts. Robinson details reward hacking on public benchmarks and the creation of CursorBench, a private eval set. He envisions agent-based automation where researchers launch experiments from Slack and models train derivative models, creating a self-improving intelligence loop.

Modern Post-Training: A Deep Dive — Will Brown, Prime Intellect

Jul 13, 2026 · 46:52

Will Brown of Prime Intellect details the company's open-source ecosystem of post-training tools, including the verifiers and prime-rl libraries, arguing they enable efficient and affordable training of frontier agentic models for enterprises. Verifiers V1 decomposes environments into tasks, harnesses, and runtimes using a decorator pattern and Pydantic, supporting group rewards like conciseness bonuses. Prime-RL is an asynchronous reinforcement learning framework that allows long-horizon coding rollouts to overlap, achieving a GLM-5 step on 28 nodes in under 5 minutes for 131k context, with a 1,000-step run costing roughly $50k. The framework supports custom algorithms including on-policy distillation, GRPO, and self-distillation via decomposable loss and algorithm classes. Prime Intellect's Lab platform offers hosted multi-tenant LoRA training live now, with full fine-tuning arriving soon, enabling enterprises to develop environments on CPU and push them to the cloud for scalable post-training.

Stop Making Models Bigger, Make Them Behave — Kobie Crawford, Snorkel

Jun 10, 2026 · 20:56

Kobie Crawford of Snorkel explains how a 4B parameter model fine-tuned via RL for under $500 outperformed Qwen 3 235B on financial analysis tool use. The key was training tool discipline—inspecting schemas and self-correcting errors—not deeper reasoning. Single-table training alone boosted multi-table FinQA benchmark from 13.9% to 26.6%, and breaking evals into rubrics identifies which behaviors to fix.

What Lies Beneath the API — Benjamin Cowen, Modal

Jun 2, 2026 · 12:40

In this episode, Ben Cowen from Modal argues that as AI products mature, fine-tuning becomes essential, citing cases like Intercom beating their frontier API at 1/10th the cost. He identifies three signals it's time to fine-tune: paying more for the API than customers pay you, evaluative plateaus, and latency requirements that shared endpoints can't meet. Cowen explains that supervised fine-tuning now fits in 300 lines of Python and that reinforcement learning rollouts can scale to 50,000 sandboxes using serverless platforms like Modal. He contends that frontier labs aim to win at everything, while businesses need to win at their specific logic, making fine-tuning a natural destination. The episode provides practical guidance on when and how to make the leap, emphasizing that building an agent harness and collecting eval data already sets the stage for training.

Lessons from Trillion Token Deployments at Fortune 500s — Alessandro Cappelli, Adaptive ML

May 12, 2026 · 18:35

Alessandro Cappelli, co-founder of Adaptive ML, argues that 95% of GenAI pilots fail to reach production because they rely on proprietary models or instruction fine-tuning, which lack systematic feedback integration. Reinforcement learning (RL) is the only post-training technique that mathematically incorporates defects, business metrics, and production signals to continuously improve models. RL enables smaller, cheaper, faster models that enterprises can own—critical for scaling use cases like AT&T’s transcript summarization or Manulife’s agents. For agents, RL naturally fits because it was designed for environments; synthetic data is generated as a byproduct of environment training, not a prerequisite. Reward signals come from business KPIs (e.g., containment rate) or LLM judges defined by human rubrics in hours, not weeks. Adaptive ML’s Adaptive Engine abstracts away RL complexity (orchestrating four models for PPO) and provides pre-built recipes to industrialize model deployment.

Everything I Learned Training Frontier Small Models — Maxime Labonne, Liquid AI

Apr 29, 2026 · 20:13

Maxime Labonne, head of post-training at Liquid AI, presents the LFM2.5 recipe for training frontier small models, arguing they require specialized approaches distinct from scaled-down big models. He details how Liquid's architecture uses gated short convolutions for latency-sensitive on-device deployment, achieving faster throughput than Gemma 3 or Qwen 3.5. Post-training stages—SFT, on-policy preference alignment, and RL—are tailored for narrow task focus like data extraction and tool use. A key challenge is 'doom loops' (repetition), which reached 15% after mid-training in a 1.2B reasoning model; solutions include preference alignment rejecting looped responses and RL with verifiable rewards and n-gram penalties, nearly eliminating the issue. He advocates combining small models with agentic tools (e.g., web search) to overcome memory limits, as they excel at reasoning and tool use despite lower knowledge capacity. The talk also covers decisions on when to use small vs. large models—latency, privacy, offline use—and notes that distillation alone likely won't fully solve doom loops.

Let LLMs Wander: Engineering RL Environments — Stefano Fiorucci

Apr 8, 2026 · 40:35

Stefano Fiorucci demonstrates how to build Reinforcement Learning environments for language models using the open-source Verifiers library, arguing that training small models with verifiable rewards can surpass large closed models on specific tasks. He maps classic RL concepts to LLMs, introduces Verifiers components for single-turn, multi-turn, and tool environments, and then walks through an experiment where he takes LiquidAI's LLM 2 — a small open model — and transforms it into a tic-tac-toe master via supervised fine-tuning and GRPO-based reinforcement learning. After training, the model dominates random opponents and draws 85% of games against optimal ones, eventually outperforming GPT-5 Mini against identical optimal opponents. Fiorucci shares practical lessons: large batch sizes (≥256) ensure stable training, hidden biases in opponent algorithms can skew results, and starting from a base model (not a reasoning model) avoids truncated thinking traces. He concludes that if you can define a clear reward signal, you can build an environment and train a small specialized model to beat a large closed model at a fraction of the cost.

AGI: The Path Forward – Jason Warner & Eiso Kant, Poolside

Dec 27, 2025 · 15:56

Jason Warner and Eiso Kant, co-founders of Poolside, present their vision and roadmap towards AGI-level capabilities for knowledge work, demonstrating their second-generation model Malibu Agent converting ADA code to Rust live on stage. They argue that next-token prediction paired with reinforcement learning is the key breakthrough, a contrarian bet they made two and a half years ago. The episode centers on their work in high-consequence code environments for defense and government, where agents must operate with tight permissions. They announce a large compute cluster of over 40,000 GB300s coming online and a public API release early next year via AWS Bedrock. Warner recounts meeting Kant through a failed GitHub acquisition, and Kant invites the audience to build with their models, emphasizing that future agents will handle tasks over days as intelligence scales.

Code World Model: Building World Models for Computation – Jacob Kahn, FAIR Meta

Dec 17, 2025 · 16:41

Jacob Kahn, a research scientist at FAIR Meta, presents the Code World Model (CWM), a 32 billion parameter dense transformer that models program execution rather than just syntax. CWM predicts execution traces line by line, enabling neural debugging and approximation of the halting problem. Trained on GitHub data and refined with synchronous RL and long-context mid-training, CWM uses bash-oriented tool use and achieves strong throughput through asynchronous model updates. The model is open-source on Hugging Face, with code and a technical report available, and aims to build foundations for reasoning and planning in AI-driven software systems.

Agent Reinforcement Fine Tuning – Will Hang & Cathy Zhou, OpenAI

Dec 9, 2025 · 16:55

Will Hang and Cathy Zhou of OpenAI's fine-tuning team introduce Agent Reinforcement Fine-Tuning (RFT), a method to improve AI agents by training them end-to-end on tasks involving tool calls and reasoning. They define an agent as a model that interleaves reasoning with external tool interactions, unlike regular models. The hierarchy of optimization moves from prompt engineering to task optimization to RFT, which changes model weights based on a custom reward signal. New features allow models to call tools via public endpoints and use custom rewards hosted externally. Case studies show concrete gains: Cognition improved code edit planning by 10 points with 1,000 examples and reduced tool call steps from 8–10 to 4; Codto's deep research agent boosted recall by 6% while cutting long-tail tool calls (over 15) down to 2–4; Cosine achieved state-of-the-art on enterprise code benchmarks by using strict graders that reward only pass-tested code; and Macco wrote GPU kernels from just 100 PyTorch prompts, beating SOTA by 72% after addressing reward hacking with seven edge-case detectors. Four principles for success: define tasks unambiguously, mirror production traffic in train/eval sets,…

RL Environments at Scale – Will Brown, Prime Intellect

Dec 9, 2025 · 18:30

Will Brown of Prime Intellect argues that scaling reinforcement learning environments beyond engineering—to community and accessibility—is key to broadening AI research. He presents Prime Intellect's open-source stack, including Verifiers for building environments and the Environments Hub for sharing them, as a way to turn any task harness into an RL training or evaluation loop. Brown demonstrates how environment-based fine-tuning boosted a Qwen 3 4B model from 55% to 89% on a Wikipedia search task, matching much larger models. He frames environments as the 'web apps of AI research'—simple to start, but capable of capturing product complexity, as seen with Cursor's Composer and OpenAI's Codex. Prime Intellect validated this approach by training the 100B-param Intellect 3 on 500 GPUs, and will soon release Lab, a platform to run environments without managing infrastructure.

Efficient Reinforcement Learning – Rhythm Garg & Linden Li, Applied Compute

Dec 9, 2025 · 20:19

Rhythm Garg and Linden Li, co-founders of Applied Compute, describe how their company uses efficient reinforcement learning (RL) to specialize large language models for enterprise tasks. They explain that synchronous RL wastes GPU time waiting on straggler samples—99% of arithmetic problems complete in ~40 seconds, but the tail takes 80 more seconds—so they adopt asynchronous pipeline RL. This method dedicates fixed GPUs to sampling and training, allowing continuous inference but introducing stale tokens (up to a tolerated staleness threshold) that require importance ratio corrections. To balance speed and stability, they model the system mathematically: using a roofline-based latency curve for sampling, per-GPU training throughput, and constraints on staleness and KV cache memory. Their simulations, parameterized by response length distributions, reveal an optimal GPU allocation that yields ~60% speedup over synchronous RL while keeping staleness within ML limits. This modeling lets them predict runtime and configure runs without expensive trial-and-error.

Z.ai GLM 4.6: What We Learned From 100 Million Open Source Downloads — Yuxuan Zhang, Z.ai

Nov 22, 2025 · 19:39

Yuxuan Zhang from zAI details the technical roadmap behind the GLM 4.6 open-source model series, which has surpassed 100 million downloads and tied for number one on the LMSYS Chatbot Arena alongside GPT-4o and Claude 3.5 Sonnet. The training pipeline uses 15 trillion tokens of pre-training data, followed by 7 trillion tokens of code and reasoning data, repo-level contexts at 32,000 tokens, and 100 billion tokens of long-context agent trajectories up to 128,000 tokens. Zhang introduces SLIME, a hybrid synchronous/asynchronous RL framework that decouples agent-environment interaction from GPU training to avoid bottlenecks. He explains why single-stage RL at 64,000 tokens outperforms multi-stage approaches for preserving long-context abilities, and shows that token-weighted loss converges faster than sequence-average loss for code RL. The multimodal GLM 4.5V model handles native resolution images and video with temporal index tokens, enabling GUI agent capabilities. Deployment is supported via vLLM and SGLang, with an API at z.ai and an open-source coding assistant.

Latent Space Paper Club: AIEWF Special Edition (Test of Time, DeepSeek R1/V3) — VIbhu Sapra

Jul 25, 2025 · 53:54

VIbhu Sapra presents DeepSeek's latest models and announces a new curriculum-based Test of Time Paper Club. DeepSeek's May 28th R1 update doubles reasoning tokens (12k→25k) and matches O3/Gemini 2.5 Pro, improving AIME 2024 from 70% to 87.5%. A new distillation of R1 traces into Qwen 3 8b achieves performance on par with Qwen's 235b thinking model, proving reasoning models distill efficiently. The Test of Time Paper Club will cover 50–100 foundational papers over six months in San Francisco and remote, targeting core AI engineering topics from attention and scaling laws to inference techniques. Sapra emphasizes that pure RL (GRPO) on verifiable math and code unlocks emergent reflection and aha moments, shifting scaling from pre-training compute to inference-time reasoning.

[Full Workshop] Reinforcement Learning, Kernels, Reasoning, Quantization & Agents — Daniel Han

Jul 19, 2025 · 2:42:28

Daniel Han of Unsloth presents a technical workshop covering reinforcement learning (RL), kernels, reasoning, quantization, and agents, arguing that RL with verifiable rewards (RLVR) is the key to unlocking LLM capabilities beyond supervised fine-tuning. He explains why open-source models plateaued after September 2024 until DeepSeek-R1 showed that RL can elicit reasoning, and breaks down PPO, GRPO, and the REINFORCE algorithm, emphasizing that GRPO removes the value model for efficiency. Han details how reward functions—not algorithms—are the hardest part, with examples like distance-based scoring for math. He demonstrates a free Colab notebook training a base model to reason, and shows that dynamic quantization can shrink models like DeepSeek-R1 from 730 GB to 140 GB with only ~1% accuracy loss, arguing that GPUs may stop getting faster after FP4 precision.

A Taxonomy for Next-gen Reasoning — Nathan Lambert, Allen Institute (AI2) & Interconnects.ai

Jul 19, 2025 · 19:21

Nathan Lambert, senior research scientist at AI2 and founder of Interconnects.ai, argues that next-generation reasoning models require a taxonomy of four traits: skills, calibration, strategy, and abstraction. While current models excel at math and code (skills), they overthink easy problems, wasting tokens and latency—calibration is needed to match output length to task difficulty. The real frontier is planning: models must learn strategy (choosing the right direction) and abstraction (breaking problems into tractable sub-tasks) to enable long-horizon agents like Deep Research and Claude Code. Lambert traces how OpenAI's Q*→Strawberry→o1 took 12–18 months of human data to teach backtracking and verification; planning should be easier because humans can write 5-step plans. He predicts post-training compute could reach parity with pre-training, citing DeepSeek's shift from 0.18% post-training compute in V3 to an estimated 10–20% for R1. The path forward: collect diverse verifiable questions, filter by difficulty, run stable RL—then scale.

RL for Autonomous Coding — Aakanksha Chowdhery, Reflection.ai

Jul 16, 2025 · 19:27

Aakanksha Chowdhery, CEO of Reflection AI and former lead researcher on PaLM and Gemini at Google, argues that reinforcement learning (RL) scaling is the next frontier for autonomous coding agents. She explains that inference-time techniques like majority voting and self-revision improve accuracy but require many samples (e.g., 10,000 for rare correct generations). RL training can learn to generate correct outputs directly, especially in verifiable domains like code with unit tests. Reflection AI aims to build superintelligence starting with autonomous coding, leveraging automated verification to design better reward functions. She notes challenges in scaling RL, including system complexity and reward hacking, but sees coding as ideal due to execution feedback.

Thinking Deeper in Gemini — Jack Rae, Google DeepMind

Jul 10, 2025 · 18:13

Jack Rae, lead of Gemini Thinking at Google DeepMind, presents thinking as a solution to the fixed test-time compute bottleneck in large language models. He explains that Gemini inserts a thinking stage where the model iterates via reinforcement learning, learning to self-correct and explore multiple strategies. This enables a continuous cost-performance tradeoff via thinking budgets, improving reasoning on math and code. Future work includes Deep Think, which raises USA Math Olympiad performance from the 50th to the 65th percentile by scaling inference compute further. Rae envisions models that, like mathematician Ramanujan, achieve deep, data-efficient reasoning from limited knowledge.

New York Times' Connections: A Case Study on NLP in Word Games — Shafik Quoraishee, NYT Games

Jul 5, 2025 · 18:31

Shafik Quoraishee, a game developer at NYT Games, presents his independent research into applying AI to solve the New York Times Connections word game. He explains that Connections, launched in June 2023 and second only to Wordle with hundreds of millions of plays, challenges AI's abstract reasoning through intentional decoys and overlapping categories. Quoraishee models the game as a graph coloring problem and uses semantic similarity, relational alignment scores across multiple dimensions (orthography, morphology, encyclopedic, etc.), and graph neural networks combined with reinforcement learning to build a solver. His preliminary results show increased solvability for hard puzzles, contrasting with LLMs that may simply recall internet solutions. The work aims to provide a transparent, explainable AI approach for puzzle-solving and potential game development applications.

Trust, but Verify: Knowledge Agents for Finance Workflows - Mike Conover

Apr 9, 2025 · 21:10

Mike Conover, CEO of Brightwave, explains how his company builds knowledge agents for financial workflows, digesting thousands of pages of content for due diligence and research. He argues that non-reasoning models perform only local search, so winning systems must use end-to-end reinforcement learning over tool-use calls to achieve globally optimal outputs. Conover describes design patterns like decomposing research into sub-themes, using secondary calls for error correction, and avoiding the latency trap of long feedback loops. He emphasizes that synthesis—weaving facts across documents—remains a hard problem due to limited recombinative reasoning in training data. Brightwave's product reveals its thought process through interactive citations and structured findings, allowing analysts to drill into any passage. The episode also notes the company is hiring, offering a $10,000 referral bonus.