Tuesday, October 6, 2026
Blackwell Benchmarks and the Zero Trust Agent Era
October 6 · 18 videos
Blackwell hardware hits the bench.
Speculative decoding faces the test.
Agents require zero trust security.
Evaluation platforms remain a challenge.
The agentic stack is here.
“How many credentials should your AI agent have? Zero.”
Is Speculative Decoding Worth It? Profiling vLLM on NVIDIA Blackwell
Sheilah Kirui · AI Engineer · 15 min
Watch on YouTube →Sheilah Kirui explains how speculative decoding uses small draft models to accelerate LLM inference on NVIDIA Blackwell GPUs. This technique offers significant speedups for structured tasks like code generation but introduces specific memory and hardware trade-offs.
- Speculative decoding uses a draft model 10 to 50 times smaller than the target model to predict token sequences.
- Structured tasks like JSON and SQL generation achieved performance increases of up to 1.6x on NVIDIA Blackwell hardware.
- The technique is most effective for small batch sizes and low temperature settings where token acceptance rates are high.
- Memory overhead is a significant factor as the system must host weights and KV caches for both the draft and target models.
- Performance gains are minimal in prefill heavy workloads such as long context RAG or high concurrency environments.
Why Jev Is Changing How We Build With AI
Diogo Almeida · The TWIML AI Podcast with Sam Charrington · 89 min
Watch on YouTube →Diogo Almeida introduces Jev, a model architecture designed for machine native intelligence rather than human centric text generation. It focuses on reliability and calibrated decision making to serve as a predictable software primitive.
- Jev utilizes Reinforcement Learning from Calibrated Decisions to optimize for the probability of a decision being correct.
- The architecture addresses the reliability gap in AI agents by enabling models to accurately estimate their own uncertainty.
- TypeSafe aims to move AI from conversational demos to production ready software automation through a verticalized intelligence stack.
- The model features a 32k context window and is engineered to function as a component within larger software harnesses.
- Market differentiation is shifting from pure scale to integration depth and the ability to handle multi step tasks with precision.
Move Fast and Don't Break Things: Scaling Databases for the AI Era
Ben Dicken · AI Engineer · 19 min
Watch on YouTube →Ben Dicken discusses how database infrastructure must evolve to support the massive traffic surges driven by AI agents. He argues that developer experience primitives like git like branching are essential for maintaining high availability at scale.
- AI agents can cause traffic to scale 100x overnight, requiring databases with extreme elasticity and automated sharding.
- PlanetScale implements isolation between data and control planes to ensure that management tasks do not impact core accessibility.
- Programmable interfaces and JSON based configurations allow AI agents to safely manage schema changes and prototype sharding schemes.
- Traffic Control mechanisms act as back pressure to allow databases to gracefully degrade through resource budgets during surges.
- High growth AI companies like Cursor represent a new class of customers requiring three to five nines of availability.
Fully Connected 2026 | Day 2 Full Replay
Brian Venturo · Weights & Biases · 108 min
Watch on YouTube →This keynote explores the transition of the AI industry from initial hype to a grind phase focused on durable infrastructure and massive scale agents. Leaders from CoreWeave and SemiAnalysis discuss the shift toward agent triggered web interactions and the complexities of GPU financing.
- Parag Agrawal predicts a thousand fold increase in web interaction scale as agents move the internet to a push model.
- For every 10 GPUs used for model inference, approximately 1 to 2 will be dedicated to agent based web search.
- Security testing of over 80 clouds revealed that many newer providers lack the enterprise grade rigor required for complex AI systems.
- The East Coast capital model uses debt to finance massive GPU clusters while West Coast equity drives technological innovation.
- A no blame culture is identified as a critical factor for allowing employees to find creative breakthroughs in hyper complex systems.
The Burnham & Badenoch Bounce: Is Farage Starting to Panic?
Kemi Badenoch · The Rest Is Politics · 57 min
Watch on YouTube →This political analysis examines how mainstream UK parties are attempting to reclaim momentum from populist movements through authenticity and rigorous policy. It highlights Kemi Badenoch's focus on incentive based policy and the global rise of intergenerational inequality.
- Kemi Badenoch's policy document The Right Track emphasizes judging outcomes by the incentives they create rather than stated intentions.
- Intergenerational inequality in housing and education is identified as the primary driver of social instability across Western nations.
- Authenticity has become a higher value currency than traditional polish in an attention economy dominated by outsider disruption.
- Political brands like those of Farage or Bolsonaro often survive scandals by framing themselves as enemies of a rigged establishment.
- The discussion covers global shifts including school violence in France and housing crises in Spain used to trigger snap elections.
Why Building an Eval Platform Is Harder Than It Looks : Braintrust
Hossein Niazmandi · AI Engineer · 17 min
Watch on YouTube →Hossein Niazmandi explains why AI evaluation platforms are complex systems engineering challenges rather than simple UI layers. He details how Braintrust manages non-deterministic agent traces to create a production flywheel for quality.
- Development has transitioned from passing static unit tests to continuous hill climbing against target quality metrics.
- Agent traces are massive and semi-structured JSON objects that can reach hundreds of megabytes per interaction.
- Traditional cloud data warehouses often fail under the specific read and write patterns required for agentic observability.
- Braintrust developed BTQL to handle real-time ingestion and long-running analytical queries for massive trace data.
- The future of evaluation involves headless agents that proactively find failure modes and run tests without manual intervention.
- Production failures should be automatically converted into test cases to drive an iterative improvement loop.
The Self-Improving OSS Agent Stack : Marc Klingen, Langfuse
Marc Klingen · AI Engineer · 17 min
Watch on YouTube →Marc Klingen discusses the shift from writing prompts to building self-improving loops for AI agents. He presents a reference stack that bridges online tracing with offline evaluation to automate agent refinement.
- AI development is moving toward a model where high-level goals are set by humans while models abstract downstream tasks.
- Self-improving agents can now identify error patterns like jargon leakage and hill-climb against datasets automatically.
- Historical trace data is the primary training material for next-generation agents and must be retained without sampling.
- Humans must remain in the loop at the highest level to prevent systems from overfitting to random quirks or creating slop.
- The goal of automation is to reach a high-quality and low-time-investment quadrant for engineering teams.
- Agents are evolving from write-intensive systems that ingest traces to read-intensive systems that query history to learn.
Fully Connected 2026 | Day 1 Full Replay
Brian Venturo · Weights & Biases · 112 min
Watch on YouTube →This keynote from CoreWeave and Weights and Biases marks the transition to the agentic era through the launch of CoreWeave Forge. It highlights the infrastructure and evaluation frameworks needed to manage complex agentic workloads.
- Agentic workloads are significantly more complex than chat, requiring up to 142 times more input tokens and 20 times more turns.
- CoreWeave Forge executes the AI Loop framework consisting of running, observing, curating, improving, and evaluating systems.
- The NVIDIA Vera CPU delivers 1.7 times faster performance on agentic benchmarks compared to previous generations.
- Competitive moats in AI have shifted from model ownership to proprietary evaluation platforms and curated golden datasets.
- Mission Control demos showed a 2 times improvement in cost-per-token efficiency for large-scale agent deployments.
- Infrastructure providers are now managing dozens of data centers to ensure hardware is not a bottleneck for agentic scaling.
The state of the tech industry in 2026 - LDX3 New York
Gergely Orosz · The Pragmatic Engineer · 29 min
Watch on YouTube →Gergely Orosz analyzes the 2026 tech landscape where agent-generated code has become the default for software development. He explores the resulting infrastructure crises and the changing role of engineering leadership.
- Agent-generated pull requests have seen a 10-fold increase in 2026, making traditional manual code reviews nearly obsolete.
- OpenAI successfully migrated its API from Python to Rust in four months using an agentic software factory approach.
- A systemic CPU shortage has caused lead times for cloud hardware to balloon to six months.
- The value of engineering has shifted from the intelligence of building to the wisdom of knowing what to build.
- Specialized mobile engineering roles are disappearing as generalists use LLMs to work across the entire software stack.
- Engineering leaders are encouraged to return to hands-on technical work to maintain relevance in an automated environment.
Harness Engineering: How to Build a Software Factory : Dru Knox, Tessl
Dru Knox · AI Engineer · 20 min
Watch on YouTube →Dru Knox introduces the concept of Harness Engineering to help teams transition into operating a software factory. He explains how engineers should focus on building the factory while agents handle all user-facing code.
- A software factory is measured by its autonomy, automation of reviews, and overall output quality.
- The Three Loop Framework includes a fast inner loop, an exhaustive outer loop, and a meta loop for system intelligence.
- Engineers are shifting from individual contributors to internal tool builders who manage the agent-ready infrastructure stack.
- Software factories eliminate backlogs by allowing agents to perform architectural refactors that humans typically neglect.
- Success is measured by driving down manual takeovers and increasing the volume of PRs initiated without human input.
- Best practices in agentic development change so rapidly that they can become anti-patterns within a single week.
Building Cyber Defense for the Agentic Era
Kevin Mandia · a16z · 47 min
Watch on YouTube →Kevin Mandia discusses the shift from human-led cyber attacks to autonomous drone swarms that require fully automated defense systems. He explains how AI changes the economics of security by finding logic flaws at microsecond speeds.
- Armadin uses a Hyper-attack framework to conduct continuous offensive operations that map a network's metadata twin.
- The system discovered over 90 zero-day vulnerabilities in Fortune 500 production environments during 2026.
- AI-led attacks operate as autonomous drone swarms capable of probing thousands of paths simultaneously rather than single sniper rounds.
- Armadin Blue creates autonomous tourniquet compensating controls through EDR and firewalls before adversaries can strike.
- Security expertise must be paired with AI research because models alone often underestimate how adversaries bypass guardrails.
- The platform utilizes up to 25,000 AI agents working in concert during a single hyper-attack session.
Your Best Defense Against Brain Rot
Rob Dial · The Mindset Mentor Podcast · 16 min
Watch on YouTube →Rob Dial explores how the loss of hobbies leads to identity compression and cognitive atrophy in adults. He argues that diverse interests are essential for neurological health and protecting against existential crises.
- Identity compression occurs when an individual's sense of self narrows into a single dimension like work or parenting.
- Approximately 95 percent of an average person's daily thoughts are repetitive, leading to functional fixedness and mental stagnation.
- Engaging in hobbies activates play circuits and directed attention restoration to lower cortisol and stimulate neuroplasticity.
- Purpose is discovered through action and curiosity rather than waiting for a single moment of inspiration.
- Choosing engaging behaviors over numbing activities like scrolling helps regulate the nervous system while building new skills.
- Innovation often stems from a mode of play which allows the brain to form new neural connections outside of work grinding.
Cooking with Codex — Charlie Guo & Gabriel Chua, OpenAI
Charlie Guo · AI Engineer · 67 min
Watch on YouTube →OpenAI engineers demonstrate how to transition from simple chatbots to autonomous software factories that execute long-running engineering tasks. They showcase agents building complex applications and reconstructing packages from research papers over several hours.
- The software factory model shifts human developers into management roles where they unblock agents and set high-level goals.
- A live macOS translation app was built in just over four minutes using a Slack screenshot as the primary context.
- Compaction techniques allow agents to manage context over ultra-long threads without losing logic or performance.
- Thread-to-thread handoff enables cloud-based agents to trigger local browser testing on a user's physical machine.
- One demonstrated agentic session ran for over 15 hours to achieve a single autonomous goal.
- Developers should focus on building the tools and plugins an agent needs to unblock itself rather than writing every line of code.
How AI Agents Pay: Checkout in ChatGPT and Google AI Mode — Sam Parsons, PayPal
Sam Parsons · AI Engineer · 14 min
Watch on YouTube →Sam Parsons explains the technical paths for agentic commerce where AI agents handle transactions directly for users. He details how PayPal bridges the gap between AI models and merchant checkout systems.
- Agentic commerce allows user intent to move directly from discovery to checkout within a single AI interface.
- PayPal supports three primary paths including OpenAI's Agentic Commerce Protocol and Google's Universal Commerce Protocol.
- The Model Context Protocol enables custom agent applications like Claude to interact with payment endpoints.
- Agent-initiated payments are tracked in the PayPal dashboard to provide merchants with specific attribution data.
- Settlement and refund processes remain consistent for enterprises even when the transaction is initiated by an autonomous agent.
- There is a technical trade-off between seamless in-agent checkout and the trust provided by external handoff methods.
Advanced Workshop: Mastering AI Observability — Doug Guthrie, Braintrust
Doug Guthrie · AI Engineer · 111 min
Watch on YouTube →Doug Guthrie presents a framework for active AI observability where production data automatically informs development through a continuous flywheel. He demonstrates how to use automated scoring and topic modeling to identify failures in agentic systems.
- The active flywheel framework uses production data to drive automated regression testing and agent improvements.
- BrainStore is a purpose-built database designed to handle sub-second queries over massive trace datasets that challenge traditional solutions.
- Automated topic modeling identifies unknown unknowns in production logs at a cost of six cents per million tokens.
- A 100 percent pass rate on evaluations often indicates that test cases are not sufficiently diverse or difficult.
- Coding agents can be equipped with observability skills to query logs, modify code, and submit pull requests with evidence.
- Hybrid deployment models allow enterprises to keep data planes within their own secure cloud environments while using observability tools.
How Many Credentials Should Your AI Agent Have? Zero.
Jim Clark · AI Engineer · 18 min
Watch on YouTube →Jim Clark discusses the transition of AI agents from supervised tasks to long running unsupervised processes. He argues that security should focus on the execution sandbox rather than the model loop itself.
- The ideal security state for an agent sandbox is to contain zero long lived credentials.
- Cross App Access (XAA) allows agents to use existing corporate SSO and OAuth providers for identity claims.
- Progressive Disclosure ensures agents only receive specific tools and resources required for their current sub task.
- Model Context Protocol (MCP) Gateways create a single control point that makes agent harnesses interchangeable.
- Decomposing complex tasks into role separated sandboxes prevents mixing untrusted inputs with dangerous tools.
- Coding agents often spend only two minutes using signing keys for every hour of actual work.
CoreWeave Forge: Demo from Fully Connected 2026, with Corey Sanders
Corey Sanders · Weights & Biases · 11 min
Watch on YouTube →Corey Sanders demonstrates CoreWeave Forge as a development layer designed to close the gap between production failure and model improvement. The presentation focuses on a five stage loop to refine an IT help desk agent.
- The AI Loop framework consists of five stages: Run, Observe, Curate, Improve, and Evaluate.
- CoreWeave Forge uses serverless Reinforcement Learning to automate the path from a first agent to a best agent.
- AI research agents like ARIA and Agent Lens identify production patterns that human developers might miss.
- Improving AI quality may require accepting performance trade offs such as a one second increase in latency.
- Self service infrastructure removes the friction of hardware commitments and speeds up iteration cycles.
- The demo successfully fixed a 1.6 percent misprioritization error in an automated help desk system.
What Is an Inference Engine, Anyway?
Charles Frye · AI Engineer · 59 min
Watch on YouTube →Charles Frye explains the architecture and engineering trade offs of modern LLM inference engines like SGLang and vLLM. He highlights how inference serves as the primary revenue center for AI businesses.
- Inference workloads are split into high arithmetic intensity prefill and memory bandwidth bottlenecked decode phases.
- Speculative decoding can provide speedups ranging from 2x to 8x by transforming sequential decodes into parallel evaluations.
- The host side scheduler acts as a single threaded gatekeeper to the parallel processing power of the GPU.
- Large models require loading approximately 100 gigabytes of data for every single token generated.
- Custom inference stacks allow for optimizations like prefix reuse that generic APIs cannot efficiently handle.
- Performance engineering shifts from kernel optimization to host side orchestration as GPUs become faster.
References
PeopleSheilah Kirui · Diogo Almeida · Sam Charrington (x.com/twimlai) · Rich Sutton · Ben Dicken (x.com/BenjDicken) · Brian Venturo · Parag Agrawal · Dylan Patel · Jenson Button · Mike Intrator · Kemi Badenoch · Andy Burnham · Pedro Sánchez · Thomas Sowell · Shabana Mahmood · Hossein Niazmandi · Marc Klingen (x.com/marcklingen) · Chen Goldberg · Brandon Hutman · Ian Buck · Corey Sanders · Gergely Orosz · Boris Cherny · Peter Mattis · Dima Zats · Solomon Chachere · Charity Majors · Dru Knox (x.com/drufball) · Kevin Mandia · David George (x.com/DavidGeorge83) · David Slater · Travis Lam · Evan Pñena · Rob Dial (coachwithrob.com) · Tony Robbins · Andrew Huberman · Charlie Guo (x.com/charlierguo) · Gabriel Chua (x.com/gabrielchua) · Sam Parsons (sjparsons.com) · Doug Guthrie · Ankur Goyal · Jim Clark · Tushar Jain · Hen · Charles Frye (x.com/charles_irl) · Sean Wang · Alexa Gordic
ToolsvLLM · NVIDIA Blackwell · Jev · TypeSafe · PlanetScale · Cursor · CoreWeave · SemiAnalysis · Weights & Biases · Braintrust · BTQL · Langfuse · CoreWeave Forge · NVIDIA Vera Rubin NVL72 · NVIDIA Vera CPU · Tessl · Armadin Red · Armadin Blue · EDR · Whiz · ChatGPT (openai.com) · Google AI Mode · Claude (anthropic.com) · Agentic Commerce Protocol · Universal Commerce Protocol · Model Context Protocol · Braintrust (braintrustdata.com) · BrainStore · ClickHouse · Docker · Anthropic · Okta · MCP (Model Context Protocol) · ARIA · Agent Lens · SGLang · Modal · CUDA
PapersThe Right Track · The Bitter Lesson · The AI Loop · The Three Loop Framework