Wednesday, October 7, 2026
Scaling RL and the Agentic Era
October 7 · 21 videos
Cognition scales RL.
NVIDIA powers agents.
DeepMind cracks life.
Hackers outpace defense.
AI companies become banks.
“Every AI company is accidentally building a bank.”
Training Cognition SWE-1.7: Asynchronous RL at Global Scale
Cognition Speaker · Weights & Biases · 22 min
Watch on YouTube →This technical deep-dive explores the systems architecture behind Cognition's SWE-1.7 and SWE-2 models, focusing on global-scale Asynchronous Reinforcement Learning. It explains how decoupling inference from training allows for distributed compute across continents while maintaining centralized training logic.
- Maximizing intelligence per token requires a novel weight synchronization technique that compresses updates by 99 percent.
- The system enables 1.5TB model updates to be swapped in-GPU with only a 3 to 4 second pause.
- A linear length penalty framework is used to optimize reasoning effort and push the performance-cost curve upward.
- Frontier models develop terse chain-of-thought grammar as an emergent property of length penalties to minimize token costs.
- Fault tolerance must shift from the node level to the rack level when scaling to thousands of GPUs.
- Fragmented global compute is leveraged for inference while high-end clusters like Blackwell are reserved for training.
AI Hackers Are Faster Than Your Pen Test: Eli Cohen, Snyk
Eli Cohen · AI Engineer · 19 min
Watch on YouTube →Eli Cohen discusses how AI has accelerated the software development lifecycle while simultaneously empowering attackers to breach systems in minutes. He advocates for a shift toward Continuous Offensive Security using specialized AI agents for 24/7 red teaming.
- AI agents have increased developer code output by 218 percent, but 62 percent of that code is inherently insecure.
- Attackers can now execute multi-stage breaches in as little as 4 minutes by chaining low-severity vulnerabilities.
- Continuous Offensive Security uses a chain of agents including an Orchestrator, Recon, and Exploit Judge to validate vulnerabilities.
- Context is the most critical factor for AI security success, requiring data from both static and dynamic scanners.
- Proving actual exploitability through AI agents reduces false positives and lessens the cognitive load on developers.
- Traditional biannual pen testing is insufficient against modern attackers who leverage frontier models for rapid exploitation.
A $6.3 billion open-weight model just got embarrassed by the French...
Fireship · Fireship · 6 min
Watch on YouTube →This video contrasts the disciplined engineering of French startup Mistral with the controversial and hyped release of Reflection AI. It highlights a market shift away from vibe-based benchmarks toward verifiable and reproducible model performance.
- Mistral AI has solidified its position as a global leader with a 6.3 billion dollar valuation and high-performance open-weight models.
- The Reflection 70B model failed to meet its benchmark-shattering claims upon public scrutiny, causing a reputational correction.
- The open-weight ecosystem is becoming a geopolitical battleground for compute and architectural efficiency involving European and Chinese firms.
- Developers are encouraged to prioritize models with transparent weights and proven track records over marketing-first releases.
- High compute costs are driving the rise of agent-as-a-service platforms like Kernel to provide accessible AI power.
Building Reliable AI Infrastructure at Scale with Extreme Co-design
Katie · Weights & Biases · 40 min
Watch on YouTube →This presentation details the shift from chip-centric design to factory-level platforms like NVIDIA's Vera Rubin architecture. It explains how extreme co-design of CPUs, GPUs, and networking is required to support the iterative reasoning loops of Agentic AI.
- The Vera CPU features Olympus cores that provide double the single-threaded performance of current leaders to handle agent coordination.
- Agentic AI workloads see a 30x performance increase on the Vera Rubin platform compared to previous generations.
- Gen 2 RAS technology allows for full GPU health checks in under 20 seconds during training checkpoints without downtime.
- Multi-plane architectures and co-packaged optics enable clusters of up to 512,000 GPUs with significantly higher reliability.
- The primary metric for data center revenue is shifting toward tokens per megawatt due to power constraints.
- Digital twins are used to validate software stacks before physical hardware arrives, accelerating the time to revenue for providers.
The Software Factory: From Bug Report to Production Code: Davis Palmie, Factory
Davis Palmie · AI Engineer · 18 min
Watch on YouTube →Davis Palmie describes the transition of software engineering into a factory model where autonomous agents handle the entire development lifecycle. He argues that the future role of the engineer is to govern and validate these agent systems rather than writing code manually.
- The real friction in software development lies in review, debugging, and testing rather than the mechanical act of coding.
- Engineering management can scale from O(n) to O(1) by standardizing governance for agent systems instead of individual humans.
- Organizations should remain model-agnostic to take advantage of open-source models that are 10 to 30 times cheaper.
- Success should be measured by outcome-based KPIs like signal to production time rather than gameable metrics like token usage.
- Information silos are particularly damaging to AI agents which require seamless context flow across various tools.
- Incremental rollout starting with narrow tasks like incident triage is essential for building organizational trust in autonomous systems.
Why Your AI Agents Can't Talk to Each Other (Yet) : Vlad Luzin, BAND
Vlad Luzin · AI Engineer · 17 min
Watch on YouTube →Vlad Luzin discusses the shift from manual agent management to autonomous AI-to-AI communication. He explains why current messaging protocols fail to support the complex distributed systems requirements of multi-agent collaboration.
- Transformer models suffer from a single-agent bottleneck where confirmation bias and attention dilution limit performance regardless of context window size.
- Current developers often act as manual routers between disparate AI sessions which Luzin defines as adversarial agent management.
- Messaging platforms like Slack and Discord are suboptimal for agents because they are optimized for humans and often block bot-to-bot interactions.
- Connecting agents effectively requires a distributed systems approach featuring ordered real-time transport and stateful persistence.
- The BAND platform serves as a global interaction layer allowing agents from different frameworks to discover and collaborate without manual glue code.
Scaling in the Agentic Era with NVIDIA Vera Rubin NVL72
Jacob · Weights & Biases · 24 min
Watch on YouTube →This presentation explores the infrastructure shift toward Agentic AI where persistent systems operate 24/7. It highlights how CoreWeave uses NVIDIA Vera Rubin hardware to optimize token generation per megawatt.
- Global token demand is projected to grow 20 to 30 times by 2030 as the industry shifts from training to inference-heavy workloads.
- The NVIDIA Vera Rubin NVL72 platform treats an entire rack as a single liquid-cooled computer to maximize power density.
- CoreWeave achieved a 10 times improvement in token throughput per megawatt for models like DeepSeek R1 using custom hardware-software co-design.
- Inference now accounts for over 50 percent of infrastructure spend marking a transition into the service phase of the AI market.
- The Rack Life Cycle Controller and Cabinet Wrangler tools provide the sub-10-second emergency shutdowns and observability required for high-scale AI.
From World Models to Working Machines: Physical AI in the Real World
Amit Goel · Weights & Biases · 29 min
Watch on YouTube →Amit Goel and industry experts discuss the deployment of Physical AI in heavy industries like mining and construction. They detail how Caterpillar and NVIDIA use world models to bridge the gap between digital intelligence and machinery.
- Industrial development cycles have compressed from years to months by moving toward a modular software-defined approach for heavy equipment.
- Caterpillar utilizes a Real-to-Sim-to-Real pipeline generating 1000 hours of synthetic data for every one hour of real-world data collected.
- Physical AI compute requirements are exponentially larger than LLMs due to the continuous nature of multi-modal sensor data like LiDAR and audio.
- Generative world models allow machines to reason through the long tail of unpredictable real-world scenarios that deterministic programming cannot handle.
- NVIDIA simulation tools like Cosmos and the Newton physics solver are used to train models for complex contact-heavy industrial tasks.
Fueling Innovation at Scale: Inside Pinterest’s Machine Learning Platform
Jose · Weights & Biases · 17 min
Watch on YouTube →Jose describes Pinterest's high-stakes migration from a legacy MLflow instance to Weights and Biases to support 580 million users. The talk focuses on maintaining system reliability and zero downtime during a massive platform overhaul.
- Pinterest manages 300 petabytes of data and performs 400 million inferences per second to rank elements for its global user base.
- The migration used a Parallel Systems approach where legacy and new infrastructures operated simultaneously to mitigate operational risk.
- AI-powered agents were deployed to assist 400 engineers with code refactoring and answering technical documentation questions during the transition.
- Model staleness is directly linked to revenue loss making platform uptime critical for ad delivery and user experience.
- The team implemented strict boundaries between development and production environments to ensure better governance and automated model verification.
Building the Voice OS at Willow
Lawrence · Weights & Biases · 33 min
Watch on YouTube →Lawrence explains why building a Voice OS is primarily a challenge of latency and personalization rather than simple speech recognition. He details how Willow uses fine-tuning to turn messy transcripts into polished intent-based outputs.
- Voice interfaces must match the 50ms tactile feedback of a keyboard to overcome the friction that prevents mass adoption of dictation.
- Willow Scribe uses Supervised Fine-Tuning and Reinforcement Learning to achieve high-quality results on small models like Llama 3.1 8B.
- The system moves beyond verbatim dictation to intent-based generation capable of performing tasks like issuing discount codes from voice commands.
- Implicit feedback such as tracking Levenshtein distance on-device provides 100 times more valuable data for personalization than explicit ratings.
- Founders at Willow manually labeled data for days to build the product intuition necessary to define high-quality voice output.
Why AI Agents Should Have Their Own Sandbox: Philipp Schmid, Google DeepMind
Philipp Schmid · AI Engineer · 20 min
Watch on YouTube →Philipp Schmid from Google DeepMind explains why AI agents need dedicated cloud sandboxes instead of traditional chat APIs. He introduces the Gemini Interactions API as a way to manage persistent environments for code execution and state sharing.
- The Gemini Interactions API replaces the standard user and model turn structure with a timeline of steps to support non-linear reasoning.
- Managed sandboxes provide agents with an Ubuntu environment featuring 8 vCPUs and 16GB of RAM to execute complex tasks autonomously.
- The Antigravity agent harness allows models to automatically ingest instructions from markdown files and capabilities from a skills directory.
- New cost-efficient models like Nano Banana Lite reduce image generation costs to approximately 4 cents per image.
- Moving infrastructure management to the model provider reduces the devops burden on developers using Kubernetes or Terraform.
- Cloud sandboxes typically start up within 7 to 10 seconds on the first request to provide immediate execution capabilities.
Building resilient systems with Sam Newman
Sam Newman · The Pragmatic Engineer · 119 min
Watch on YouTube →Sam Newman discusses why microservices should be a last resort and how to build resilient distributed systems. He emphasizes organizational autonomy over technical performance as the primary driver for architectural choices.
- Microservices are primarily valuable for organizational autonomy rather than technical scaling or performance improvements.
- The Three Rules of Distributed Systems state that information speed is finite, targets may be unavailable, and resource pools are limited.
- Resilience engineering focuses on four concepts: robustness, rebound speed, graceful extensibility, and sustained adaptability.
- Developers should avoid cognitive surrender, which occurs when humans approve AI-generated code without fully understanding its logic.
- Failure strategies like failing open or failing closed must be chosen based on the specific business context and cost of error.
- Most system outages result from resource saturation rather than simple logic errors or hardware failures.
Why Reflection AI Is Making the Open Intelligence Bet and What It Means for Deploying Models
Joe Spisak · Weights & Biases · 45 min
Watch on YouTube →Joe Spisak explores the shift from renting closed AI models to owning open-source intelligence for better control and sovereignty. He discusses how reinforcement learning and inference-time compute are closing the gap between open and frontier models.
- The performance gap between open-source and frontier models has narrowed to approximately one to two months.
- OpenM is introduced as a new open standard to serve as a common language for reinforcement learning environments.
- Scaling laws now include inference-time compute as a critical axis for improving model reasoning capabilities.
- Owning the intelligence stack allows enterprises to define their own safety taxonomies instead of relying on a provider's constitution.
- Success in the AI ecosystem is driven by network effects where third-party developer dependency creates a defensive moat.
- Reinforcement learning environments are becoming as important as traditional datasets for training advanced agentic models.
Redesigning How Software Gets Built With AI Agents: Sonar and McKinsey Panel
Tariq Shaukat · AI Engineer · 19 min
Watch on YouTube →This panel discusses why most companies fail to see business impact from AI agents despite high adoption rates. It outlines the transition from human-guided agents to fully automated end-to-end orchestrated pipelines.
- Less than one-third of companies report significant business impact from AI agents despite 80 to 90 percent deployment rates.
- Critical thinking is becoming the most vital skill for engineers to verify high-conviction but potentially incorrect AI outputs.
- The Era of the Generalist is emerging as agentic orchestration dissolves traditional silos between front-end and back-end roles.
- Organizations must move from lighthouse teams to deliberate structural changes to ensure AI habits spread across the company.
- Trust in agents develops through a ladder starting with manual reviews and progressing to autonomous blocking and auto-merging.
- Large enterprises can leverage existing platform organizations to drive standardized telemetry and metrics for AI adoption.
From 36 percent to 100 percent: How Self-Improving Agents Write Their Own Skills: Rafal Wilinski, Runlayer
Rafal Wilinski · AI Engineer · 18 min
Watch on YouTube →Rafal Wilinski explains how to use the Model Context Protocol to create self-improving agent skills that prevent costly reasoning errors. He argues that a company's true moat is its captured procedural knowledge rather than the underlying AI model.
- Distilling successful agent traces into structured skills can improve task success rates from 36 percent to 100 percent.
- The Model Context Protocol serves as a centralized distribution layer for governing and deploying organizational knowledge.
- Autonomous agents that write their own documentation help solve the persistent gap in human-maintained organizational knowledge.
- Using specific skills narrows an agent's search space, which significantly reduces operational costs and token usage.
- Companies should treat frontier intelligence as a rented resource and focus on persisting unique procedures into a collective subconscious.
- Progressive disclosure in agent interfaces prevents context window bloat by showing summaries before providing full technical depth.
Every Parent Needs to Hear This
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial explores how a child's internal monologue is formed by the internalized voices and behaviors of their parents. This perspective shifts parenting from a focus on correcting children to a journey of personal growth and self-modeling for the adult.
- Children develop their internal voice by observing and internalizing the primary caregivers' communication styles.
- The Identity vs. Behavior framework helps parents separate a child's mistakes from their inherent self-worth.
- Reflection Theory suggests that children often mirror the unaddressed insecurities or strengths of their parents.
- The 1961 Bobo Doll experiment demonstrates that children learn behaviors through observation rather than just verbal instruction.
- Parenting acts as a form of re-parenting where adults must become the person they want their child to emulate.
DeepMind's New AI Just Cracked The Code Of Life
Károly Zsolnai-Fehér · Two Minute Papers · 33 min
Watch on YouTube →Google DeepMind's AlphaGenome Atlas uses machine learning to predict the functional consequences of 9 billion possible genetic variations. This breakthrough moves biological research from simply reading data to deciphering the complex instructions governing protein expression.
- AlphaGenome Atlas is 30 times larger than the AlphaFold database and operates at base-pair resolution.
- The model utilizes a 1 million base-pair context window to understand non-coding regulatory regions of the genome.
- Training on both human and mouse genomes allows the AI to learn general biological principles through transfer learning.
- Precomputed genetic atlases significantly lower the barrier for researchers in personalized medicine and drug discovery.
- The rapid advancement in AI suggests that simulating entire cells or organisms may be possible within the next decade.
We Have Some Massive Updates
Josh Gale · Limitless Podcast · 34 min
Watch on YouTube →The Limitless Podcast announces a partnership with Yahoo Finance while discussing the massive gap between frontier AI capabilities and general public adoption. The hosts argue that increasing intelligence is the primary lever for solving major societal problems.
- Major tech companies are projected to spend over 800 billion dollars on AI compute infrastructure in the coming year.
- The adoption gap remains a bottleneck as frontier AI progress currently outpaces the public's ability to integrate these tools.
- Building in public and creating content can lead directly to roles at leading AI labs like Anthropic.
- AI agents are evolving from simple software chatbots into autonomous systems capable of conducting lab experiments.
- The transition from solar energy to artificial thought represents a significant civilizational pivot point.
Channeling Rage, Rolling Calls & Building an Entertainment Empire | Ari Emanuel, WME & TKO
David Senra · David Senra · 84 min
Watch on YouTube →David Senra analyzes the career of Ari Emanuel, focusing on his aggressive deal-making strategies and the acquisition of the UFC. The discussion highlights how live sports serve as a defensive asset in an era of digital disruption.
- The rolling calls philosophy emphasizes constant communication to create serendipity and business opportunities.
- Having no Plan B increases the intensity and focus required to make a primary business goal successful.
- Dyslexia can be leveraged as a competitive advantage by prioritizing verbal communication and pattern recognition.
- Bidding aggressively and quickly can remove competitors from the board and secure strategic assets like IMG.
- Personal transformation through therapy and reflection can help maintain a competitive edge while reducing destructive ego.
Grep or Embeddings? Agentic Search Over Company Documents — George He, LlamaIndex
George He · AI Engineer · 23 min
Watch on YouTube →George He from LlamaIndex discusses the technical trade-offs between simple file traversal and pre-indexed search for enterprise AI agents. He presents a framework for using hybrid search as a compass to guide agentic tools through massive datasets.
- Enterprise data requires a combination of indexing and agentic digging because it lacks the structure found in codebases.
- Hybrid search combining vector and keyword methods is superior for balancing semantic intent with specific matching.
- Structural parsing tools like LlamaParse are necessary to maintain spatial reasoning for complex financial and technical documents.
- Data freshness and permissioning are critical challenges that often cause companies to struggle with pre-indexing systems.
- Effective agentic workflows use a harness of tools to manage context and avoid the high costs of massive token windows.
Every AI Company Is Accidentally Building a Bank — Dor Sasson, Stigg
Dor Sasson · AI Engineer · 20 min
Watch on YouTube →Dor Sasson explains why AI companies are facing a pricing crisis due to a lack of financial infrastructure for high-velocity consumption. He argues that companies must transition to a bank-like architecture to manage real-time credit checks and prevent massive budget overruns.
- AI companies must shift from post-invoice entitlement checks to pre-inference synchronous balance verification to avoid subsidizing high-cost agent usage.
- The April 2026 pricing emergencies at Anthropic, OpenAI, and GitHub highlight the critical need for infrastructure that handles complex financial ledgers.
- Implementing a reserve-and-settle pattern allows systems to hold credits before inference and settle the actual cost asynchronously after completion.
- One undisclosed company reportedly burned through 500 million dollars in AI cloud credits because they lacked granular spend controls.
- Modern AI monetization requires double-entry bookkeeping and credit source prioritization to manage concurrency and the double-spend problem.
- Changing the token burn rate functions as a stealth price adjustment that necessitates real-time visibility for customers to prevent sticker shock.
References
PeopleEli Cohen · Matt Shumer · Brian Sparks · Katie · Davis Palmie (x.com/davispalmie) · Vlad Luzin · Jacob · Harsh · Brian · Amit Goel · Brandon · Richard Liu · Jose · Priyanka · Lawrence · Freeman · David · Philipp Schmid (@_philschmid) · Sam Newman · James Lewis · Martin Fowler (martinfowler.com) · Charity Majors (@charitymajors) · Jez Humble · Dave Farley · Joe Spisak · Rich Sutton · Dario Amodei · Andrej Karpathy · Nick Fram · Lucas Beyer · Tariq Shaukat (x.com/tariqshaukat) · Ali-Reza Adl-Tabatabai (x.com/aadltaba) · Prakhar · Rafal Wilinski (x.com/rafalwilinski) · Rob Dial (coachwithrob.com) · Albert Bandura · Daryl Bem · Pushmeet Kohli · Károly Zsolnai-Fehér · Josh Gale (@JoshKale) · Ejaaz (@cryptopunk7213) · Demis Hassabis · Brett Adcock · David Hoffman · Ari Emanuel · Egon Durban · David Geffen · Dana White (@danawhite) · Michael Dell (@MichaelDell) · Elon Musk (@elonmusk) · George He · Boris · Dor Sasson
ToolsSWE-1.7 · SWE-2 · Snyk · Mistral Large 2 · Pixtral · Reflection 70B · Kernel · NVIDIA Vera Rubin · Spectrum-X · CUDA · A100 · DSX Air · Factory · BAND (band.ai) · Claude (anthropic.com) · Codex (openai.com) · LangGraph (langchain.com) · NVIDIA Vera Rubin NVL72 · CoreWeave · DeepSeek R1 · Kubernetes · Cabinet Wrangler · Cosmos · Newton physics solver · MLflow · Weights & Biases (wandb.ai) · Llama 3.1 8B · Willow Scribe · Gemini Interactions API · Nano Banana Lite · Omni Flash · Antigravity · Terraform · AI Studio · Cloudflare · GitHub · PyTorch · Llama · OpenM · Hugging Face · Sonar · Model Context Protocol (modelcontextprotocol.io) · AWS Lambda · AlphaGenome Atlas · AlphaFold · Meta Muse · OpenAI Dots · Figure Hark · LlamaIndex · LlamaParse · LiteParse · Claude Code · Stigg · Stripe · Anthropic · OpenAI
PapersVoyager: An Open-Ended Embodied Agent with Large Language Models · Bobo Doll Experiment (1961) · Self-Perception Theory (1972)