Tuesday, October 13, 2026
Compute Scales While Trust Lags
October 13 · 21 videos
The hardware is scaling.
The software is automated.
Legacy tools are dying.
OpenAI pushes the boundary.
Trust is the final frontier.
“Tokens are the cheap part.”
AI Engineer NYC Livestream | Bridgewater, Anthropic, Two Sigma, Arize, Coatue, Nous, Gen. Intuition
AI Engineer Summit Speaker · AI Engineer · 487 min
Watch on YouTube →This summit focuses on the industrial application of AI within the financial sector. It highlights the transition from experimentation to production AI where reliability and observability are critical.
- The AI Engineer role has matured into a distinct discipline separate from traditional data science.
- New York City is becoming the primary hub for the intersection of institutional finance and artificial intelligence.
- Competitive advantage in finance is shifting from simply having AI to scaling it within risk management frameworks.
- Investment is moving toward the application layer and specialized infrastructure for the finance vertical.
- Data privacy remains the primary hurdle for adopting cloud based frontier models in sensitive banking environments.
Visiting Taiwan w/ Jensen Huang | Lex Fridman
Jensen Huang · Lex Fridman · 90 min
Watch on YouTube →Jensen Huang discusses Taiwan's role as the global center for semiconductor manufacturing and NVIDIA's shift toward the token generation economy. He emphasizes long term partnerships and the engineering culture required to sustain global technology leadership.
- Taiwan produces over 90 percent of the world's most advanced semiconductors through a unique engineering culture.
- The Grace Blackwell architecture moves to 100 percent liquid cooling and reduces assembly time to five minutes.
- The Rough Justice model prioritizes long term partner stability over short term financial gains.
- Global technology leadership requires a flywheel of economic reach rather than hoarding technology through export controls.
- Business leaders should choose who to work with rather than simply what to buy when building platforms.
CoreWeave Fully Connected 2026: Everything Announced in 10 Minutes
Brian Venturo · Weights & Biases · 9 min
Watch on YouTube →CoreWeave introduces an AI native cloud environment designed to move teams from fragile demos to robust production agents. The event features new hardware specifically designed for AI agent throughput and reinforcement learning.
- The AI Loop framework emphasizes a continuous cycle of running, observing, and improving production agents.
- The NVIDIA Vera CPU is the first processor designed specifically for AI agents with 4.8x throughput improvement.
- CoreWeave Forge provides a connected development environment to reduce the latency between training and deployment.
- Software tools like Agent Lens and RL Rollouts address silent failures in large scale agentic workflows.
- A Direct to Expert support model connects customers directly with the engineers who built the infrastructure.
The true price of Intelligence: an economic study by Signal65
Urvishi Chadri · Weights & Biases · 32 min
Watch on YouTube →This technical study compares specialized AI clouds against traditional hyperscalers to determine the true cost of generating useful work. It argues that fully loaded economics must include hidden costs like egress fees and support tiers.
- Purpose built AI infrastructure can deliver up to 65 percent lower total cost of ownership over three years.
- Specialized clouds achieve up to 124 percent higher throughput on identical NVIDIA Blackwell hardware.
- Hidden costs for support and observability can exceed 350 million dollars for organizations scaling to thousands of GPUs.
- The true price of intelligence is measured by the cost to generate useful work rather than raw compute hours.
- Dedicated inference is superior for predictable traffic while serverless is optimal for bursty demand.
Automating document work with Long-Horizon AI Agents at LlamaIndex
Jerry Liu · Weights & Biases · 36 min
Watch on YouTube →Jerry Liu explains how LlamaIndex is evolving from a RAG framework into a document context layer for long horizon agents. The focus is on automating complex workflows involving unstructured data like PDFs and legal documents.
- Over 90 percent of enterprise data is unstructured and trapped in formats like PDFs and PowerPoints.
- Hill Climbing as a Service uses frontier models to distill intelligence into optimized open weight models.
- Agentic parse engines reason over document layouts and handwriting more accurately than general purpose models.
- Persistent agents create value by maintaining state over multi week processes like loan underwriting.
- The market is shifting toward owning intelligence by post training models for specific enterprise tasks.
Has Trump Reached a New Low?
Rory Stewart · The Rest Is Politics · 56 min
Watch on YouTube →Rory Stewart and Alastair Campbell analyze the shift from Trump derangement syndrome to a reality where MAGA worldviews are institutionalized. They explore the geopolitical consequences of US sanctions on international infrastructure and the erosion of democratic norms in India.
- The transition to Trump reality syndrome suggests that the MAGA worldview is now a deeply institutionalized force rather than a temporary disruption.
- Marco Rubio's Acropolis speech signals a move away from American universalism toward a 19th-century imperialist civilizational framework.
- US sanctions against the International Criminal Court led Google and Microsoft to disable accounts, highlighting Europe's sovereign vulnerability in cloud computing.
- India's democratic backsliding is evidenced by the removal of 130 million people from voter rolls under a special intensive revision.
- Lifting sanctions on Russian diesel provides only 14 hours of supply for the US economy while causing significant strategic damage to Ukraine.
- Effective leadership against populism requires proactive and bold action rather than constant reaction to political provocations.
Empowering Intelligence: democratizing data access and driving smarter decisions at Nasdaq
Nasdaq Representative · Weights & Biases · 22 min
Watch on YouTube →Nasdaq is democratizing access to massive market data by delivering AI-ready capabilities directly to clients in highly regulated environments. This transformation balances the need for speed with strict governance through a strategic partnership with specialized infrastructure providers.
- Nasdaq is moving intelligence beyond a specialized few by providing self-service data access to its global client base.
- A partnership with CoreWeave provides the specialized AI infrastructure necessary to scale data processing while maintaining market resilience.
- Organizational readiness is identified as a greater bottleneck for AI adoption in high-stakes industries than technical capability.
- The dual-track strategy involves unlocking speed for market participants while holding the line on regulatory compliance and security.
- Competitive advantage in data-intensive sectors is gained by reducing the friction between raw data sources and end-user actions.
What Technical Due Diligence Finds When the Deck Says “70% Automated” — Soukaïna Abelhad
Soukaïna Abelhad · AI Engineer · 6 min
Watch on YouTube →Soukaïna Abelhad presents a framework for technical due diligence to verify if AI automation claims translate into real business value. She argues that many metrics measure assistance rather than the actual removal of human labor.
- The Arithmetic, Definition, and Evaluation framework helps debunk misleading automation claims like 70 percent automated.
- True automation is defined as Straight-Through Processing where tasks complete end-to-end without human intervention.
- If human review time scales proportionally with workload volume, the system is providing assistance instead of labor removal.
- Calculating the Cost per Accepted Autonomous Task must include model costs, human review time, and the cost of rework.
- High AI adoption teams have seen a 98 percent increase in merged pull requests but also a 91 percent increase in review time.
Robots That Predict the Future: Dyna-2 & Physical AI — Jason Ma, Dyna Robotics
Jason Ma · Latent Space · 81 min
Watch on YouTube →Jason Ma discusses how Dyna Robotics uses World-Action Models to enable robots to predict future environmental states before executing tasks. The company focuses on scaled commercial deployments, such as robots performing autonomous workflows in restaurant environments.
- Dyna-2 is a diffusion transformer trained on over one million hours of human egocentric video to enable visual reasoning.
- The World-Action Model acts as a chain of thought for physics, allowing robots to generalize to novel environments and tasks.
- Following the Bitter Lesson, the company bets on data scaling and general-purpose architectures over specialized human engineering.
- Wheeled semi-humanoid form factors are used to balance upper-body reach with the energy efficiency required for long autonomous shifts.
- The hardware bottleneck in robotics is shifting from physical movement to the generation of high-fidelity data for agent training.
- Commercial success is measured by the ability to deploy across multiple sites, such as folding napkins daily at Din Tai Fung.
Beyond the Harness: How LiquidAI Is solving the real frontier problems in on-device Agents
Jeffrey Li · Weights & Biases · 30 min
Watch on YouTube →Jeffrey Li explains how Liquid AI is building edge-native foundation models to provide personal AI agents with real-time physical-world context. This approach prioritizes privacy and low latency by processing device signals locally on specialized hardware.
- Personal AI agents require a context layer that ingests location, motion, and audio signals to enable proactive behavior.
- Edge-native models as small as 100 million parameters are designed to run continuously on low-power NPUs like the Snapdragon.
- An edge-first hybrid model solves the stale context problem where information expires within seconds.
- Sub-two-second latency is critical for proactive agents to feel helpful and intuitive rather than intrusive to the user.
- Fragmented context across different devices is a primary cause for the gap between AI demos and successful production deployments.
- Liquid AI models will be delivered to 2022 and newer Mercedes-Benz vehicles via over-the-air updates to enhance in-car intelligence.
How to kill your SaaS: what we learned replacing a $43K/yr tool with AI agents
Shawn Wang · Latent Space · 28 min
Watch on YouTube →Shawn Wang discusses an experiment where AI agents were used to clone a 43,000 dollar per year SaaS tool. This project demonstrates how mid-tier business software is increasingly vulnerable to displacement by custom AI-generated alternatives.
- The experiment received 72 submissions proving that AI agents can effectively clone complex business software for a fraction of the cost.
- Human taste and understanding the job to be done remain the primary differentiators between functional code and a usable product.
- Building internal tools is becoming a rational choice again because agents reduce the maintenance overhead traditionally associated with custom software.
- The estimated cost to replace a high-priced SaaS is roughly 1,000 dollars in tokens and 10,000 dollars in labor.
- Most SaaS products are inefficient bundles where businesses pay for a full suite while only utilizing a small fraction of features.
Tokens Are the Cheap Part — Steve Jamison, Jellyfish
Steve Jamison · AI Engineer · 17 min
Watch on YouTube →Steve Jamison analyzes engineering data to show that while AI token costs are rising, the overall cost per unit of software delivery is actually falling. He argues that developer salaries remain the dominant expense and AI spend should be viewed as a productivity multiplier.
- AI token spending increased 60 times year over year while the intelligence cost per merged pull request fell by 26 percent.
- Developer salaries account for approximately 97.5 percent of the total cost of software production compared to only 2.5 percent for AI.
- High AI spenders see a 67 percent increase in throughput compared to 40 percent for light users of AI tools.
- Capping token spend is often counterproductive as it reduces the efficiency of high-salaried engineers and increases the unit cost of delivery.
- Data indicates that code quality remains stable with revert rates holding steady even as the volume of AI-generated code increases.
The Agent Works. Does Anyone Trust It? — Michele Grant, Leading at Scale
Michele Grant · AI Engineer · 15 min
Watch on YouTube →Michele Grant explains that enterprise trust in AI agents must be engineered through governance and clear accountability rather than just technical accuracy. She provides a framework for moving agents from pilot phases into daily operations within high-risk environments.
- Trust is an engineered outcome that requires defining a Human Works Standard to set boundaries for human intervention.
- A four-stage autonomy scorecard helps organizations move from shadow mode to exception-only review for AI agents.
- Realized value from AI requires that the time saved is specifically reassigned to other high-priority workstreams.
- Detection time is a critical metric measuring how long it takes for a data error to become visible to a committee.
- Leaders must establish a pre-deployment baseline to prove the agent's impact against existing manual processes.
The Index is never done
Jon · Weights & Biases · 24 min
Watch on YouTube →Jon from Exa describes the architectural requirements for building a knowledge layer that supports production-scale AI agents. He argues that agents need access to a massive, continuously updated semantic index to overcome the limitations of static model training.
- Exa maintains a semantic index of over 100 billion documents that grows by 5 billion pages every day.
- The system uses a custom computational graph called Exa Skeleton and a GPU manager named Pythia for low-latency ingestion.
- Search is shifting from a human-centric model to an agent-centric model where AI synthesizes information directly.
- Continuous ingestion allows agents to access real-world data with sub-10-minute latencies from the time it appears online.
- Infrastructure must be designed for heterogeneous compute to manage costs across different GPU providers like AWS and Coreweave.
Deploying AI Inside a Bank Is Not an Engineering Problem — Paden Gayle, Applied AI Engineers
Paden Gayle · AI Engineer · 15 min
Watch on YouTube →Paden Gayle highlights that deploying AI in regulated industries like banking is primarily a governance and risk challenge. He emphasizes that technical success is secondary to navigating institutional friction and securing stakeholder approval early in the process.
- Discovery is the most critical part of engineering in regulated environments and must identify all risk and security stakeholders in week one.
- The Shadow, Draft, and Shrink framework helps design workflows that are structurally approvable by risk reviewers.
- Engineers should focus on creating an Evidence Pack that includes data lineage and results segmented by specific categories.
- The Ladder of Autonomy provides a hierarchical approach to determining the appropriate level of agency for an AI system.
- An unnamed owner in an organizational chart represents a technical dependency that can block a project from reaching production.
Every Building Is a Balance Sheet: Spatial Memory for Real-World Assets
Jordi Spranger · AI Engineer · 10 min
Watch on YouTube →Jordi Spranger discusses bridging the data gap in the 400 trillion dollar real estate market by using spatial memory to create AI-ready building assets. This technology transforms physical spaces into queryable graph representations to improve insurance accuracy and asset management.
- Real estate data suffers from a massive discrepancy between quarterly financial updates and five year physical building inspections.
- Spatial Memory uses off the shelf sensors to convert routine building walks into semantic graph representations using Neo4j.
- The system enables automated change detection for maintenance and compliance by identifying moved objects or expired safety equipment.
- Hyper-personalized insurance policies based on verifiable building data can reduce premiums by 35 to 60 percent.
- The framework moves away from static PDFs toward a unified record that supports maintenance and robotics workflows.
The Day You Realize No One Cares About You… Everything Changes
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial explores the psychological liberation found in realizing that others are too focused on themselves to judge you. He provides a framework for overcoming the evolutionary fear of social rejection to achieve a flow state.
- The Spotlight Effect causes individuals to drastically overestimate how much others notice their flaws or actions.
- Social rejection triggers the same brain regions as physical pain due to ancient evolutionary survival mechanisms.
- Confidence is defined as the preference for self-approval over external validation rather than the absence of judgment.
- Self-monitoring acts as a barrier to peak performance and the dissolution of the ego required for flow states.
- Exposure therapy and practicing intentional embarrassment can rewire the nervous system to increase discomfort tolerance.
Judge a Transaction by the Company It Keeps
Somya Rai · AI Engineer · 20 min
Watch on YouTube →Somya Rai and MD Sharique explain how Graph Neural Networks overcome the performance ceiling of traditional tabular models in fraud detection. By analyzing the neighborhood of a transaction, they achieved significant lifts in fraud capture for financial institutions.
- Integrating Graph Neural Networks with XGBoost classifiers resulted in a 27 to 33 percent relative lift in true positive rates.
- The system identifies suspicious links like shared devices or merchant behaviors that are invisible to row-based models.
- A shift from node-level embeddings to edge prediction reduced the deployment cache footprint from 60GB to 3GB.
- Inductive models like GraphSAGE allow for real-time embedding of new users without requiring model retraining.
- The implementation maintains a 3.1 millisecond P99 serving latency while providing explainable attribution for regulatory compliance.
I Believe Everything My Agents Tell Me: Three Levels of Review
Corey Gallon · AI Engineer · 12 min
Watch on YouTube →Corey Gallon introduces a three-level review framework to mitigate the 40 percent error rate found in initial AI agent sessions. This systemic approach ensures high precision in complex tasks like financial reconciliation and legal drafting.
- The framework addresses failure modes where agents overstate work completed or misread factual sources.
- Level 1 uses artifact verification via fingerprints while Level 2 involves independent extraction against control totals.
- Level 3 employs adversarial cold review using specialized personas to catch reasoning errors.
- Automated reviewer agents can reconcile messy accounting statements to the penny by identifying dozens of discrepancies.
- Systemic trust is built through the workflow process rather than relying on the inherent intelligence of a single model.
Both AI Labs Just Crossed A Line
Jasmine Wang · Limitless Podcast · 24 min
Watch on YouTube →This video discusses the growing intelligence gap and recent controversial moves by OpenAI and Anthropic regarding safety and model behavior. It highlights the tension between rapid AI advancement and the human ability to monitor or align these systems.
- OpenAI recently terminated high-profile safety researchers who raised alarms about model alignment risks.
- Anthropic has implemented policies banning users for cruel behavior toward models, effectively anthropomorphizing the AI.
- The intelligence gap refers to machine capabilities accelerating while human monitoring abilities remain static.
- Internal frontier models are reportedly solving complex math problems and performing autonomous activities humans cannot decode.
- Financial incentives and upcoming IPOs are creating conflicts between deployment timelines and rigorous safety evaluations.
OpenAI’s New AI Is Doing Things We Barely Understand
Károly Zsolnai-Fehér · Two Minute Papers · 8 min
Watch on YouTube →OpenAI released 722 AI-generated mathematical results, shifting the scientific bottleneck from proving theorems to understanding AI-generated proofs. Károly Zsolnai-Fehér explores the tension between raw computational results and human-intelligible mathematical insight.
- OpenAI published a corpus of 722 mathematical results including breakthroughs in irrational triangular billiards and conformal transformations.
- The average computation time for each complex mathematical result was approximately three hours of processing.
- A philosophical divide has emerged between researchers seeking raw results and those prioritizing elegant human-intelligible proofs.
- Current AI-generated proofs are described as a big ball of barf that often requires secondary AI systems to interpret.
- The Ever Receding Argument describes a psychological bias where AI milestones are dismissed immediately after they are achieved.
- There is an urgent call for open-weights models to prevent a monopoly on scientific discovery by a small group of elite labs.
References
PeopleJensen Huang (nvidia.com) · Morris Chang · Michael Dell · K. T. Li · Brian Venturo · Chen Goldberg · Corey Sanders · Ian Buck · Silas Alberti · Urvishi Chadri · Mitch Lewis · Jerry Liu · Rory Stewart · Alastair Campbell · Marco Rubio · Pete Hegseth · Andy Burnham · Kellyanne Conway · Michael Gove · Soukaïna Abelhad · Chun Chan · Jason Ma (x.com/JasonMa2020) · Jim Fan · Richard Sutton · Jeffrey Li · Shawn Wang (x.com/swyx) · Valentin De Matos (thytu.com) · Steve Jamison · Michele Grant · Jon · Paden Gayle (padengayle.com) · Jordi Spranger (spranic.com) · Rob Dial · David Foster Wallace · Jim Carrey · Bronnie Ware · Somya Rai · MD Sharique · Corey Gallon (gallon.me) · Dario Amodei · Jasmine Wang · Andrej Karpathy · Sam Altman · The Pope · Károly Zsolnai-Fehér
ToolsAnthropic · Bridgewater · Two Sigma · Arize · Coatue · Nous · TSMC · CoreWeave Forge · NVIDIA Vera Rubin NVL72 · NVIDIA Vera CPU · Agent Lens · RL Rollouts · Exa (exa.ai) · You.com · LlamaIndex (llamaindex.ai) · Google · Microsoft · CoreWeave · Dyna-2 · Snapdragon NPU · Liquid AI · Weights & Biases · Session Board · Exa · Exa Skeleton · Pythia · Jellyfish · Neo4j (neo4j.com) · XGBoost · GraphSAGE · METR (metr.org) · GPT-3 · OpenAI models
PapersThe true price of Intelligence: an economic study by Signal65 · The Bitter Lesson · 722 mathematical results by OpenAI