Friday, October 9, 2026
The Infrastructure of Agentic Intelligence
October 9 · 20 videos
Compute scaling hits the physical limit.
Agents require new command line interfaces.
Multi-cloud clusters bridge the capacity gap.
Physics AI meets NASA benchmarks.
Hardware is back in style.
“Generation is cheap, review is expensive.”
Accelerating Multi-Cloud Workloads with Google Cloud and CoreWeave
Sean Darrington · Weights & Biases · 31 min
Watch on YouTube →Sean Darrington explains how the partnership between Google Cloud and CoreWeave addresses data gravity issues in AI workloads. This collaboration enables enterprises to burst compute to specialized GPU clusters while maintaining primary data in Google Cloud.
- CoreWeave uses a blank sheet of paper approach to networking and storage to eliminate hypervisor overhead.
- The Chaos object storage system features a Load accelerator that uses node-local NVMe for high throughput.
- Google Cloud Partner Cross-Cloud Interconnect reduces networking setup time from weeks to minutes via API.
- The platform achieves up to 7GB/s per GPU read throughput with a 99.7 percent cache hit rate.
- Predictable billing models eliminate egress fees and storage operation charges to attract enterprise workloads.
- Mistral reported a 2.5x speed increase in training performance using this multi-cloud infrastructure.
Generation Is Cheap, Review Is Expensive: How to Stop Shipping AI Slop — Gabriel Martinez, G2i
Gabriel Martinez · AI Engineer · 17 min
Watch on YouTube →Gabriel Martinez discusses the epidemic of AI slop where rapid code generation outpaces human judgment and review capabilities. He argues that engineers must transition from code producers to system orchestrators to maintain software quality.
- AI slop is defined as work that appears finished before necessary thinking and trade-off analysis are complete.
- While the cost of producing code has dropped to near zero, the cost of maintenance and review remains high.
- Engineers should use diagrams as primary review surfaces to reduce cognitive overhead during the evaluation process.
- Adopting the Convention over Configuration model helps limit the decision space and simplifies system architecture.
- Performance anxiety can lead teams to merge unreviewed code, creating a bigger ball of mud over time.
- Success should be measured by system clarity and value produced rather than lines of code or tokens used.
How MasterClass Builds Multi-Agent AI That Learners Can Trust
Mandar · Weights & Biases · 24 min
Watch on YouTube →Mandar describes how MasterClass is evolving from simple chatbots to AI-native learning experiences using a multi-agent architecture. This system aims to scale personalized tutoring which has historically shown significant performance gains for students.
- The architecture employs four specialized agents including a Planner, Orchestrator, Teacher, and Memory.
- Personalization allows the AI to teach complex concepts like finance using industry-specific examples relevant to the learner.
- An autonomous AgentOps loop uses execution traces to help agents self-diagnose and root-cause failures.
- The system implements pedagogical guardrails to ensure the AI pushes learners toward active recall instead of spoon-feeding answers.
- MasterClass Executive was developed in partnership with Chicago Booth and OpenAI to provide high-stakes educational reliability.
- Moving from human-in-the-loop QA to autonomous monitoring allows for proactive identification of sub-optimal learning paths.
From Context to Memory: Your Agents Need a Real Memory Layer — Anders Swanson, Oracle
Anders Swanson · AI Engineer · 18 min
Watch on YouTube →Anders Swanson argues that context windows are insufficient for true agent intelligence and that a dedicated memory layer is required. He outlines a systems-based approach to managing stateful experiences across multiple sessions.
- Context windows are temporary inputs for current runs and do not constitute persistent experience or memory.
- A real memory layer requires a multi-modal database supporting vector, lexical, graph, and temporal queries.
- Memory is categorized into four distinct types including episodic, semantic, procedural, and working state.
- The redact first principle is essential to prevent sensitive information or secrets from entering the memory layer.
- Systems without managed state waste capital by repeatedly reading documents and performing redundant exploratory work.
- Hybrid search and feedback loops are necessary to prevent silent drift and handle poisoned data as memory pools grow.
Beam: The Great American Open Model with ReflectionAI Co-Founder and CEO Misha Laskin
Misha Laskin · No Priors: AI, Machine Learning, Tech, & Startups · 69 min
Watch on YouTube →Misha Laskin discusses the launch of Beam, a 500-billion parameter open-weight model, and the economic shift toward owned intelligence. He explains why enterprises are moving from renting APIs to owning their own AI stacks for cost and sovereignty.
- Beam is a massive open-weight reasoning model trained using 10,000 GB300 GPUs over a four-week period.
- Reinforcement learning with AI-in-the-loop can accelerate research progress by up to 30x compared to human-only efforts.
- The hardware lottery effect means that once an architecture like Transformers succeeds, hardware is optimized specifically for it.
- Enterprises spending over 100 million dollars annually on tokens are likely to transition to open-weight models for cost optimization.
- Open-source strategies can act as a subsidy for global infrastructure while driving demand for specialized inference services.
- Laskin advocates for Linus' Law, suggesting that open models are more secure because more eyes can identify vulnerabilities.
Designing CLIs for Agents, Not Humans
Pedro Lopez · AI Engineer · 18 min
Watch on YouTube →Pedro Lopez explains how Airbyte is evolving into a data and action layer for AI agents by building specialized interfaces like MCP servers and agent-centric CLIs. This shift focuses on machine-readable outputs and predictable command structures rather than human-friendly interactive prompts.
- Users expect Model Context Protocol sessions to function like mobile apps with long-lived authentication rather than frequent web-style re-auth.
- Standard specifications like MCP are often implemented inconsistently across clients which requires custom fallbacks for features like URL elicitation.
- A strict noun-verb command structure allows agents to successfully guess the correct syntax even when documentation is incomplete.
- Airbyte positions its infrastructure as a data and action layer to capture the emerging market for autonomous AI agents.
- The platform optimizes data indices for search specifically to support agentic querying instead of standard analytics.
- Claude imposes a 2 kilobyte limit on MCP server and tool call descriptions which necessitates concise documentation.
High-Performance Engineering: Forward Deployed at Aston Martin F1
Agata Murachik · Weights & Biases · 24 min
Watch on YouTube →This video explores the partnership between CoreWeave and Aston Martin F1 where forward-deployed engineers use AI to solve high-stakes racing challenges. It highlights how transformer models and fine-tuned audio processing provide a competitive edge in the zero-margin environment of Formula 1.
- Deep domain expertise in physics must be paired with AI knowledge to create features that accurately represent racing reality.
- The forward-deployed model identifies unsaid pain points by placing engineers directly in the factory and at the track during high-stress windows.
- A relatively small Transformer architecture outperformed traditional boosting models for physics-based regression despite limited data sets.
- The Radio Insights project uses an audio-to-text model fine-tuned on 3,000 hand-annotated samples to filter 40 concurrent radio channels.
- Competitive intelligence is extracted from high-volume noise to identify rival pitting strategies and tire degradation in real-time.
- Project delivery is dictated by an immovable race calendar rather than flexible quarterly corporate goals.
Beyond Faster Simulation: Hitting NASA's 2030 CFD Benchmark and the Infrastructure for Physics AI
Max · Weights & Biases · 21 min
Watch on YouTube →Max discusses a breakthrough in computational fluid dynamics where nTop and CoreWeave completed 10,000 simulations in 32 hours to meet NASA's 2030 goals early. This shift transforms simulation from a validation tool into a massive training engine for autonomous physics-based AI.
- The team executed 10,000 Large Eddy Simulations in 32.2 hours by utilizing GPU-native solvers and implicit geometry.
- Transitioning to Physics AI requires a closed loop where both geometry and solvers operate without traditional boundary representation.
- Engineers must shift focus from the mechanics of running jobs to interpreting results and directing autonomous AI across design spaces.
- The project generated 14 terabytes of structured 3D flow data to serve as a primary asset for training physical intelligence models.
- CoreWeave infrastructure provided near 100 percent computational saturation for embarrassingly parallel simulation tasks using 280 GPUs.
- Competitive advantage in engineering is moving from solver quality to the possession of high-quality training data for surrogate models.
Building on the Codex Harness
Dominik Kundel · AI Engineer · 18 min
Watch on YouTube →Dominik Kundel introduces the Codex App Server, an open-source protocol that allows developers to build deep agent integrations using OpenAI's core harness. He argues for extending existing battle-tested tools rather than building custom agent architectures from scratch.
- The Codex App Server uses a JSON-RPC protocol to expose over 120 client messages for deep IDE and application integration.
- Implementing Sign in with ChatGPT allows developers to offer premium AI tools while users cover token costs through their own subscriptions.
- Developers should prioritize using developer instructions and deferred dynamic tools to avoid polluting the model context window.
- Shipping a specific version of the Codex harness binary inside an application bundle prevents breaking changes from upstream updates.
- The protocol supports creative UI/UX experiments such as running Codex inside the game Doom to manipulate state in real-time.
- Leveraging a battle-tested harness significantly reduces time-to-market for complex agentic features in technical products.
Keep It Secret, Keep It Safe: How IBM Deploys Sensitive Data and Workloads on CoreWeave
Brian Belgadier · Weights & Biases · 35 min
Watch on YouTube →IBM Research and CoreWeave discuss the security and infrastructure challenges of deploying over 4,000 GPUs for sensitive AI workloads. They emphasize the need for cryptographic trust boundaries and federation-first identity management in high-performance environments.
- IBM transitioned from a DIY cluster to renting over 4,000 NVIDIA GPUs to bypass the long-term costs of data center refitting.
- Trust is evolving toward cryptographically defined boundaries where users do not have to trust the infrastructure provider with their keys.
- The rise of autonomous coding agents necessitates secure sandboxed execution to prevent accidental data deletion or malicious code injection.
- Automated user provisioning and short-lived keys are used to maintain strict control over researcher access across disparate clusters.
- A 91 percent scaling efficiency was achieved during multi-rack runs through tight collaboration between IBM, CoreWeave, and NVIDIA.
- Transparency in billing and performance is maintained through shared telemetry dashboards that reduce friction with auditing teams.
Former Intel CEO: Why This is the Best Time to Build Hardware
Pat Gelsinger · a16z · 53 min
Watch on YouTube →Pat Gelsinger discusses the hardware renaissance where physical constraints and energy capacity have replaced design complexity as the primary industry bottlenecks. He explains why the current explosion of AI chip startups will eventually consolidate due to massive capital requirements and software needs.
- AI tools have compressed chip design cycles from years to three months while manufacturing lead times remain fixed at nine months.
- Energy capacity is now the ultimate proxy for economic capacity as AI expansion hits limits on power delivery and gas turbine lead times.
- The memory industry added 2.5 trillion dollars in market cap over four years because AI is fundamentally a memory-compute workload.
- Current chip designs lose 40 percent of their power to guard banding which represents a significant opportunity for efficiency gains.
- Future infrastructure will require a VMware for agents to abstract complex hardware environments for autonomous AI entities.
- The industry is moving toward 800V DC power delivery and optical networking to sustain the growth of large scale AI clusters.
Stop Prompting — Greg Pstrucha, Sentry
Greg Pstrucha · AI Engineer · 15 min
Watch on YouTube →Greg Pstrucha argues that relying on manual prompting for coding agents is ineffective because agents lack long term memory and suffer from context compaction. He advocates for codifying engineering standards through deterministic checks like custom linters and automated policies.
- Context compaction causes agents to forget verbal corrections within five minutes of a shift in the conversation flow.
- Developers should use AI to write custom lint rules which were historically expensive but are now cheap to generate and maintain.
- Sentry uses automated policies to keep API endpoints in sync with OpenAPI schemas and prevent hallucination drift in agent skills.
- The Garfield review loop uses a multi-agent system to evaluate code against written engineering policies for qualitative assessments.
- Quantitative metrics like test coverage are easily gamed by agents and should not be trusted without qualitative human review.
- Raising the quality floor through automated reviews reduces the time human engineers spend on repetitive elementary feedback.
How Gratitude Changes the Way You See Your Life
Rob Dial · The Mindset Mentor Podcast · 17 min
Watch on YouTube →Rob Dial explains how to use cognitive overrides to combat the brain's natural negativity bias and hedonic adaptation. He provides practical frameworks for shifting focus from daily stressors to the abundance already present in one's life.
- The human brain is biologically hardwired with a negativity bias designed for survival which often overlooks familiar blessings.
- Using the four word trigger How lucky am I can redirect the brain to find evidence of abundance during stressful moments.
- Mental subtraction involves imagining the loss of current blessings to gain a fresh perspective on their actual value.
- High level success does not eliminate stress but instead upgrades the quality of problems from survival to management.
- Specific gratitude for small moments is significantly more effective for mental training than general or vague statements.
- Looking under a complaint often reveals that the source of stress is actually a goal that was once prayed for.
Post-Training at the Frontier: How You.com, NVIDIA, and CoreWeave Improved Nemotron’s Agentic Search
Sahil · Weights & Biases · 24 min
Watch on YouTube →This video details a three week collaboration between NVIDIA, CoreWeave, and You.com to improve the Nemotron-3.5-Lightning model for agentic search. The team focused on tool proficiency to achieve a significant increase in accuracy on the BrowseComp benchmark.
- The team achieved an 8.5 percent absolute increase in BrowseComp accuracy by focusing on tool proficiency rather than just model intelligence.
- Targeted data strategies using only 5,000 SFT traces and 1,500 RL tasks proved more effective than massive scale training.
- CoreWeave infrastructure accelerated reinforcement learning training by 15 times through the use of GPU hotloading features.
- Specialized post-training reduced the average number of tool calls per task from 168 down to 115 for the same results.
- Search infrastructure for AI agents requires a different approach than human search because agents are becoming the primary web users.
- The collaboration used a proprietary web graph to generate high quality Q&A pairs from the middle band of web popularity.
Google just surprised everyone (Gemini 4)
Demis Hassabis · Limitless Podcast · 22 min
Watch on YouTube →The video analyzes Google's release of Gemini 4 Argon as an attempt to reclaim its position as an AI leader following a period of perceived stagnation. It explores the tension between Google's massive infrastructure advantages and its challenges with shipping speed and cultural identity.
- Gemini 4 Argon benchmarks competitively in agentic coding and knowledge work against top models from OpenAI and Anthropic.
- Google processes 22 billion tokens per minute across its products and has seen 82 percent growth in its cloud division.
- Leading AI labs are using recursive self-improvement to build next generation models using current versions of their own AI.
- Google faces a 2.5 year window to solidify its AI product strategy before faster moving labs reach revenue and distribution parity.
- The company must leverage its vertical integration of TPUs and flagship models to move beyond being just an infrastructure provider.
- Public perception of tech leadership can shift rapidly if a company fails to match the iteration speed of competitors like Claude Code.
Optimize capacity, control cost, and scale economically
Samantha Fam · Weights & Biases · 40 min
Watch on YouTube →Samantha Fam and Jesse Miller discuss the shift from binary cloud pricing to hybrid models like Flex Reservations. This approach helps AI teams balance guaranteed peak capacity with cost savings during idle periods.
- CoreWeave introduced Flex Reservations which use a baseline holding fee combined with pay as you go usage rates.
- The Capacity Bundling strategy combines reserved instances for training, flex for inference, and spot instances for experiments.
- Mission Control provides an AI driven interface that uses natural language queries to identify underutilized clusters and orphan workloads.
- CoreWeave adopted the FOCUS 1.2 standard for their API to enable programmatic cost analysis across multi cloud environments.
- Spot instances on the platform feature a 7 minute preemption window for workload migration.
- The AI Loop framework matches specific infrastructure characteristics to the phases of running, observing, and evaluating models.
Data pipeline at Petabyte scale: Production Video Curation with Ray on CoreWeave
Unknown Speaker · Weights & Biases · 34 min
Watch on YouTube →This session details the engineering process of scaling a video curation pipeline to 6 petabytes using Ray and NVIDIA Blackwell GPUs. It provides a reference architecture for maximizing GPU utilization during massive multimodal data processing.
- The team processed 70 million video clips and generated captions in 90 minutes using 1,600 NVIDIA Blackwell B40 GPUs.
- Throughput reached 12,000 captions per second by re-partitioning data into 87,000 blocks and using node shaped bundles.
- The architecture utilizes Ray Data for streaming and CoreWeave AI Object Storage for high performance caching.
- GPU utilization was increased from 10 percent to near 100 percent saturation by moving to a streaming over staging mindset.
- Software fault tolerance is critical at this scale because hardware failure is a statistical certainty when running over 1,500 GPUs.
- Using managed services like Anyscale reduced the engineering overhead required for Kubernetes scaling and orchestration.
Quota Isn't Enough: Topology Aware Scheduling on Modern GPU Clusters
JJ · Weights & Biases · 31 min
Watch on YouTube →JJ and Amy explain why standard Kubernetes scheduling is insufficient for modern rack scale systems like the NVIDIA NVL72. They argue that topology must be a core requirement to avoid massive performance penalties between interconnects.
- There is an 18x performance difference between NVLink at 900 GB/s and Infiniband at 50 GB/s interconnects.
- Topology Aware Scheduling in Kueue allows teams to define physical hierarchies and specify required job shapes.
- Large scale AI success requires PodSet slicing to manage 64 GPU blocks within 72 GPU racks effectively.
- Waiting for correct topology placement is often more cost effective than starting a job immediately on scattered hardware.
- Automated node hot swapping using Kubernetes taints helps manage silent GPU failures that standard checks might miss.
- A single NVL72 rack costs approximately 18,000 dollars per day on demand, making efficiency gains financially significant.
AI Writes More PRs. Who Validates Them?
Ali-Reza Adl-Tabatabai · AI Engineer · 11 min
Watch on YouTube →Ali-Reza Adl-Tabatabai explores the bottleneck created when AI agents generate more pull requests than human reviewers can handle. He proposes Agentic Validation to automate code review and maintain engineering velocity.
- Agentic Validation combines LLM based agents with classical algorithmic program analysis for precision and security.
- The Trust Trajectory model allows teams to move from passive AI reviews to full auto merge capabilities over time.
- Unvalidated AI code shifts organizational risk to the right, increasing the burden on verification and production teams.
- Categorizing pull requests into features, chores, and refactors helps leaders track where engineering effort is spent.
- AI agents provide necessary leverage for platform teams to maintain complex CI configurations without increasing headcount.
- Automated systems can identify flaky tests and generate autofixes to reduce the duration of outer loop failures.
Why 80% Reliability Isn't Good Enough
Felipe Blanes · AI Engineer · 17 min
Watch on YouTube →Felipe Blanes from Amazon AGI Lab discusses why static benchmarks fail to predict real world AI agent performance. He introduces the eval flywheel as a method for continuous improvement based on customer signals.
- The Trust Cliff occurs at 80 percent reliability because users must monitor every step, creating more work than manual tasks.
- True autonomous hand off only begins when an agent reaches a reliability threshold of 90 to 92 percent.
- Developers should use production derived evaluations rather than imagination based evals to reflect actual customer behavior.
- Transparency about what is out of scope for a product builds more trust than inflating benchmark scores.
- Evaluations must be updated weekly to prevent them from becoming stale and failing to capture new failure modes.
- Direct conversations with customers are necessary to understand the reasons behind failures that instrumentation might miss.
References
PeopleSean Darrington · Engin · Xander · Gabriel Martinez (x.com/g2i_ai) · Edsger W. Dijkstra · D. Richard Hipp · Mandar · Priya · Professor Kumar · Mark Cuban · Anders Swanson (blogs.oracle.com/authors/anders-swanson/) · Misha Laskin · Giannis · Reed Hoffman · Linus Torvalds · Pedro Lopez (pedroslopez.me) · Agata Murachik · Eric Erns · Morgan Magguire · Max · Dominik Kundel (x.com/dkundel) · Roman · Maddie · Brian Belgadier · Andy Manoski · Ian Buck · Peter Salganik · Pat Gelsinger (x.com/PGelsinger) · Raghu Raghuram (x.com/RaghuRaghuram) · Guido Appenzeller (x.com/appenz) · John Hennessy · Nick McKeown · Greg Pstrucha · Rob Dial · Philip Brickman · Tony Robbins · Sahil · David · Kyle · Eddie · Demis Hassabis · Jeff Dean · John Jumper · Sundar Pichai · Samantha Fam · Jesse Miller · JJ · Amy · Ali-Reza Adl-Tabatabai (x.com/aadltaba) · Felipe Blanes
ToolsCoreWeave (coreweave.com) · Google Cloud (cloud.google.com) · Chaos (CoreWeave AI Object Storage) · Load (Local Object Transport Accelerator) · Partner Cross-Cloud Interconnect · Ruby on Rails (rubyonrails.org) · MasterClass Executive · OpenAI (openai.com) · Model Context Protocol (MCP) · Beam (ReflectionAI) · GB300 GPUs · Airbyte · OpenAPI · Claude · CoreWeave · Weights & Biases · Aston Martin Aramco F1 · Transformer · Radio Insights · nTop · Lattice Boltzmann Method (LBM) · NVIDIA RTX PRO 6000 · Codex App Server · JSON-RPC · ChatGPT · Doom · IBM Research · NVIDIA GB200 · Red Hat Directory Server · SCIM · StepCA · MLPerf · VMware (vmware.com) · OpenAPI (openapis.org) · Nemotron-3.5-Lightning · You.com (you.com) · Gemini 4 Argon · GPT-6 Astra · Claude Fable 5.1 · Claude Code · Ray · Anyscale · CoreWeave AI Object Storage (CAIOS) · NVIDIA Blackwell B40 · NVIDIA NVL72 · Kueue · Kubernetes · Gitar · Amazon AGI Lab
PapersNASA CFD Vision 2030 · Brickman study on lottery winners (1978) · FOCUS 1.2 (FinOps Open Cost & Usage Specification)