Wednesday, October 14, 2026
The Industrialization of Agentic Intelligence
October 14 · 15 videos
Compute is the new currency.
Infrastructure scales to meet demand.
Agents enter the enterprise.
Biology becomes a reinforcement learning environment.
Efficiency defines the next era.
“We turned the wet lab into a reinforcement learning environment.”
AI Engineer NYC: OpenAI, Blackstone, Ramp, Coinbase, Apollo, Citadel, Basis, Runway, Compute Trading
Event Narrator · AI Engineer · 532 min
Watch on YouTube →AI Engineer NYC 2026 marks the transition of LLMs from experimental tools to core industrial infrastructure in global finance. It highlights how institutions like Blackstone and Citadel are integrating AI into alpha generation and compute trading.
- The AI Engineer is replacing the traditional quant as the most critical hire in New York financial firms.
- Global leaders are moving beyond chatbots to integrate AI into core asset management and alpha generation systems.
- Success in 2026 requires treating AI as a new form of compute rather than just a search engine.
- Strategic partnerships between foundation model providers like OpenAI and firms like Blackstone create primary competitive moats.
- The conference emphasizes the engineering of robust, compliant, and low-latency systems for high-stakes financial environments.
CoreWeave Fully Connected 2026 Keynote: Introducing CoreWeave Forge | Built for Adoption
Hen Goldberg · Weights & Biases · 111 min
Watch on YouTube →CoreWeave's 2026 keynote introduces CoreWeave Forge to manage the continuous AI loop of running, observing, and improving agents. It addresses the massive infrastructure demands of autonomous agents that consume significantly more tokens than standard chat models.
- CoreWeave Forge is an integrated environment designed to manage the full AI loop from observation to evaluation.
- Autonomous agents consume 142 times more input tokens and take 20 times more turns than basic chat interfaces.
- The NVIDIA Vera Rubin NVL72 platform delivers 1.7 times faster performance in agent benchmarks and 7 times faster data processing.
- Mission Control autonomously optimized inference during a live demo, increasing throughput from 3,000 to 20,000 tokens per minute.
- Evaluation and golden datasets are identified as the only sustainable competitive moats as models become commoditized.
From Wet Lab to RL Env: 10,000 Drug-Discovery Campaigns in a Weekend — Ian Quigley, Leash Bio
Ian Quigley · Latent Space · 109 min
Watch on YouTube →Ian Quigley discusses how Leash Bio uses massive-scale data generation to move drug discovery from learning chemist styles to learning actual chemistry. The approach focuses on scaling experimental data through DNA-encoded libraries to create a high-fidelity reinforcement learning environment.
- Leash Bio released a dataset of four billion physical measurements for the BELKA Kaggle competition to improve drug discovery models.
- Public datasets like ChEMBL are often biased by individual chemist styles, with models achieving 60 percent accuracy in predicting the chemist.
- The company uses engineered protein mutations to physically break binding pockets and create ground truth data.
- Reasoning agents were used to run 10,000 medicinal chemistry campaigns in a single weekend within a physically grounded sandbox.
- Proprietary datasets generated with epistemic rigor are the primary moat in the transition of drug discovery to a data engineering problem.
I Spent Two Years Building the Median — Hariprasad Sudharshan
Hariprasad Sudharshan · AI Engineer · 36 min
Watch on YouTube →Hariprasad Sudharshan warns against the median trap where developers using identical models and tools produce non-differentiable AI products. He advocates for a novelty budget framework to ensure software provides unique value beyond standard AI capabilities.
- The democratization of high-intelligence models leads to a build-first reflex that often results in generic AI slop.
- Developers must identify and name a single unit of true novelty before writing code to avoid building commodity products.
- Distribution and specific infrastructure utilization are becoming more important moats than the underlying model performance.
- The speaker found 244 similar calculators after spending two years building an AI tool that initially lacked differentiation.
- Success requires spending judgment on specific problems rather than just building working demos which are now commodities.
OpenAI’s lead engineer was secretly a spy...
Fireship · Fireship · 5 min
Watch on YouTube →This video reports on a 2026 security breach where an OpenAI lead engineer was arrested for wiretapping the company's executive boardroom. It highlights the growing insider threat in AI labs and the misuse of internal agents for corporate espionage.
- Lead engineer Abdul Raman Alphazan used a 23 dollar DIY device to wiretap OpenAI's executive boardroom.
- The breach was detected within 48 hours by technical surveillance countermeasures after the engineer left a digital trail via ChatGPT.
- Internal agents were allegedly used to scout meeting schedules and gather intelligence on executive movements.
- The incident suggests potential ties to foreign intelligence through a Riyadh-based venture capital firm.
- This case underscores that high academic credentials do not guarantee employee loyalty or operational security in high-intensity AI environments.
Paving the Road to Production: Governed Enterprise AI Agents
Harish Gaggar · AI Engineer · 29 min
Watch on YouTube →Harish Gaggar presents a blueprint for moving AI agents from prototypes to governed enterprise systems by focusing on infrastructure like identity and durable memory. He argues that standardized controls and observability are more critical for production success than the agent's internal logic.
- Establish foundational controls including identity, centralized authentication, and durable memory before shipping agents.
- Implement a centralized Model Context Protocol gateway to decouple agents from tools and ensure secure routing.
- Use a context optimizer pipeline to prune and compress tokens which can reduce usage by up to 90 percent.
- Require human in the loop workflows for any sensitive or state changing operations in an enterprise environment.
- Aggregate tracking data for cost, quality, and latency into executive dashboards to justify agent ROI.
- Filter PII and secrets at a dedicated layer to prevent leakage during agent interactions.
Uber’s AI infra, from the inside
Uday Kiran Medisetty · The Pragmatic Engineer · 94 min
Watch on YouTube →Uday Kiran Medisetty explains how Uber's decade of investment in developer platformization enabled the creation of Minion, an agentic coding harness. The discussion highlights the shift from manual coding to context engineering where developers manage environments and grounding data.
- Uber's Minion agent now handles 15 percent of all production pull requests across the company.
- The system utilizes a massive Context Graph with over 100 node types to provide agents with necessary grounding data.
- Engineers are transitioning to context engineering roles, focusing on setting constraints and performing meta verification on AI outputs.
- A model neutral gateway allows Uber to swap underlying LLMs without breaking existing internal workflows or tool integrations.
- Automated loops like Fox handle toil such as dead code cleanup and library upgrades to maintain codebase health.
- High scale automation includes green lane processing to identify and auto merge low risk pull requests.
How to Actually Tokenmaxx and Save Millions
Alvin Pane · AI Engineer · 11 min
Watch on YouTube →Alvin Pane outlines a structured playbook for token optimization to reduce the compute tax of high volume agentic applications. He advocates for moving from vibe based prompt engineering to a measurable, statistical release process.
- Use the Observe Lever Validate playbook to instrument traces and establish a ceiling test for production variance.
- Implement programmatic prompt shrinking using frameworks like DSPy to treat instruction optimization as a search problem.
- Achieve significant cost savings by reducing prompt sizes by over 50 percent while maintaining model performance.
- Utilize paired confidence intervals instead of point estimates to ensure statistical gains before releasing model updates.
- Avoid instruction bloat by identifying and deleting redundant lines that developers are often afraid to remove.
- Conduct shadow deployments and human reviews of diffs to validate that cost saving measures do not cause regressions.
Can Open Models Keep Up With OpenAI and Anthropic?
Nathan Lambert · a16z · 58 min
Watch on YouTube →Nathan Lambert analyzes the competitive landscape between frontier labs and open model competitors, focusing on the roles of compute, data, and talent. He explores how open weight releases serve as a strategic commercial lever for labs to enter global markets.
- Model progress is driven by three forms of capital: massive compute, high quality data, and specialized talent.
- Open weight releases allow labs to bypass traditional API distribution hurdles and gain market share in the US.
- The industry is shifting from supervised fine tuning toward reinforcement learning heavy post training environments.
- Model serving costs are decreasing by approximately 2x every eight months, creating a Jevons Paradox for intelligence.
- Open source maintainers face new challenges from a high volume of AI generated noise and fake research results.
- The Trillium nonprofit aims to demystify recursive self improvement by releasing open source reinforcement learning environments.
Rebuilding CI for AI-Scale Commit Volume
Gautam Korlam · AI Engineer · 19 min
Watch on YouTube →Gautam Korlam discusses the need to redesign CI/CD pipelines to handle the 10x increase in commit volume generated by AI agents. He emphasizes that the engineering system must evolve to prevent code generation from becoming a bottleneck.
- Rebuild CI systems using aggressive incremental builds and sharding to reduce test times from minutes to seconds.
- Implement zero trust authorship where the agent writing the code cannot be the same entity that approves it.
- Use risk based policy frameworks to auto approve low risk changes like documentation or internal tool updates.
- Maintain a durable context record of agent trajectories and test reports to satisfy regulatory and audit requirements.
- Focus on the entire path from proposed change to accepted change to measure true engineering productivity.
- Address the repair loop bottleneck where time is wasted fixing failed builds and waiting for human reviews.
The Truth About Personal Growth No One Tells You
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial explores the biological drive for growth and challenges the idea that personal development has a final destination. He argues that progress is measured by how quickly one recovers from negative patterns rather than the total elimination of flaws.
- Growth is often a subtractive process focused on letting go of old identities and beliefs that no longer serve the individual.
- The speed of return metric measures how quickly a person can recognize a negative state and return to their true self.
- The End of History Illusion describes a psychological bias where people underestimate how much their values and personality will change over a decade.
- Biological actualizing tendency suggests that the drive for growth is an inescapable force present in all living organisms until death.
- Effective personal development requires implementing systems for faster pattern recognition rather than setting destination based goals.
The Cost of Accessibility in AI — Jemima Abu, CAIS
Jemima Abu · AI Engineer · 10 min
Watch on YouTube →Jemima Abu demonstrates how integrating accessibility into AI development significantly reduces API costs and improves execution speed. She argues that semantic HTML allows AI agents to avoid generating redundant custom code.
- Providing AI agents with explicit accessibility skills can reduce API costs by 58 percent and completion time by two thirds.
- Accessibility aware agents leverage semantic HTML to provide built-in functionality with significantly fewer tokens than custom JavaScript components.
- The Blueberry Muffin model suggests that accessibility must be baked into the development lifecycle from the start rather than added at the end.
- Building inaccessible products creates a self fulfilling prophecy where assistive technology users are excluded from analytics data.
- Legal risks under the Americans with Disabilities Act make accessibility a fundamental business requirement rather than an optional feature.
Lessons from History’s Greatest Leaders & Conquerors | Ben Wilson
Ben Wilson · David Senra · 71 min
Watch on YouTube →Ben Wilson and David Senra analyze the traits of historical conquerors while reflecting on Wilson's recent terminal cancer diagnosis. They discuss the concept of zeal and the importance of singular focus in achieving greatness.
- Zeal is defined as the productive pain felt when observing someone similar achieve greatness that one is capable of reaching.
- Successful leaders often break trade offs by identifying a single non obvious variable that makes other problems trivial.
- The Waste Book concept from Isaac Newton provides a space to record unrefined creative thoughts without the pressure of perfection.
- Double mindedness or having conflicting priorities is identified as a primary obstacle to achieving high level success.
- Obsession serves as a litmus test for greatness where individuals lose track of time and basic needs while pursuing their craft.
Your Fridge Is a Valid Payment Device — Eleonora Drykina & Nina Žižakić, In The Pocket
Eleonora Drykina · AI Engineer · 25 min
Watch on YouTube →Eleonora Drykina and Nina Zizakic explain the architecture of a production grade RAG system designed for complex financial documents. They focus on maintaining document structure and managing high latency in agentic workflows.
- Implementing a hybrid retrieval pipeline using Dense and BM25 search boosted the retrieval hit rate from 92 percent to 97 percent.
- Converting complex tables into CSV format and maintaining section paths allows the agent to reconstruct full chapters accurately.
- Managing a median response time of 38 seconds requires streaming Markdown text instantly while appending structured source cards later.
- Discipline instructions are used to prevent infinite loops and context window flooding during multi level retrieval tasks.
- Sentiment analysis of follow up questions serves as a proxy for user satisfaction when explicit feedback buttons are ignored.
I Used to Hate Meta. Now It’s Blowing My Mind
Theo GG · Limitless Podcast · 24 min
Watch on YouTube →This video discusses Meta's strategic pivot toward Personal Super Intelligence with its new Muse Wonder model. It highlights how Meta leverages its massive user distribution to compete with research heavy labs like OpenAI.
- The Muse Wonder model focuses on personal intelligence by using active memory to proactively surface relevant user context.
- Active Focus allows the AI agent to autonomously bypass technical obstacles like paywalls or errors without user intervention.
- Meta leverages its 3.3 billion user base across WhatsApp and Instagram to create a distribution moat for its AI products.
- The strategy targets the 99 percent of the world who are not AI experts by focusing on intuitive consumer product design.
- A massive investment in top talent signals a shift to a war mode mentality to win the personal assistant category.
References
PeopleHen Goldberg · Brian Venturo · Brandon Hutman · Ian Buck · Maulin Patel · Ian Quigley · Andrew Blevens · Mike Karesi · Alex Rives · Shiaoji Lou · Hariprasad Sudharshan · Usain Bolt · Justin Gatlin · Abdul Raman Alphazan · Prince Khaled bin Bader Al-Saud · Scott Steedman · Leon Theramin · Harish Gaggar · Uday Kiran Medisetty (x.com/udaykiran) · Thuan Pham · Gergely Orosz · Alvin Pane · Nathan Lambert (@natolambert) · Lisha Li (@lishali88) · Jensen Huang · Mark Zuckerberg · Gautam Korlam · Carl Rogers · Ram Dass · Daniel Gilbert · Jemima Abu (https://www.jemimaabu.com/) · Ben Wilson · David Senra · Thomas Edison · Napoleon Bonaparte · Steve Jobs · Isaac Newton · Aristotle · Eleonora Drykina (https://www.elu.sh/) · Nina Zizakic · Alexander Wang · Theo GG · Elon Musk
ToolsOpenAI (openai.com) · Blackstone (blackstone.com) · Ramp (ramp.com) · Coinbase (coinbase.com) · Apollo (apollo.com) · Citadel (citadel.com) · Basis (basis.ai) · Runway (runwayml.com) · CoreWeave Forge (coreweave.com) · NVIDIA Vera CPU · NVIDIA Vera Rubin NVL72 · ClickBench · Mission Control · ChEMBL · BELKA Kaggle competition · Seed Studio microcontroller · Blacksmith · Gemini · RED Ventures · LangGraph (langchain.com) · Model Context Protocol (MCP) · JFrog Artifactory · DevPods · Slate · Jagger · Minion · AIX · Model Gateway · Context Graph · Fox · OpenTelemetry (opentelemetry.io) · DSPy · GEPA · Trillium · SubmitQueue · Gitar · Sonar · WebAIM · Huntsman Cancer Institute · In The Pocket · Pydantic · Harvey Lab · Claude Opus 5.5 · GPT 6 Astra · Muse Wonder
PapersThe Bitter Lesson · Recursive Self-Improvement · Jevons Paradox · End of History Illusion