Friday, August 28, 2026
The Shift to AI Native Engineering and Formal Verification
August 28 · 16 videos
AI native teams scale fast.
Formal verification replaces bugs.
Multi agent systems review code.
NVIDIA makes a massive purchase.
Personal branding evolves for 2026.
“Your code has bugs. Lean4 has proofs.”
I Did Live Streaming For 30 Days. Here's What I've Learnt.
Alex Hormozi · Alex Hormozi · 55 min
Watch on YouTube →Alex Hormozi shares the results of a 30 day live streaming experiment focused on forcing operational efficiency and maximizing value per second. He discusses how high production speed and business integrity are the primary drivers of long term growth.
- The 30 day experiment generated over 41.5 million total views and reduced the production cycle to under 12 hours.
- Integrity is defined as doing the right thing specifically when it is inconvenient or costly to do so.
- Disagreements between business partners are usually caused by information gaps rather than fundamental personality conflicts.
- A 4.9 star rating often converts better than a perfect 5.0 because it appears more authentic to savvy consumers.
- In house legal teams are typically unnecessary until a company reaches between 30 million and 100 million dollars in annual revenue.
- The Dolly Parton test is used to measure the speed of moving from a creative concept to a finished high production video.
AI-Native Organisations Run on Skills: How to Structure and Scale Them
Imad Touil · AI Engineer · 20 min
Watch on YouTube →Imad Touil from QuantumBlack explains why AI native organizations must transition from monolithic agents to a granular catalog of manageable skills. He argues that centralized governance is the only way to prevent technical debt and redundant development in agentic systems.
- AI skills should be treated as modular and deterministic units of know how similar to microservices.
- The Agentic Software Stack separates the inner loop of agent execution from the outer loop of business workflows.
- Centralized skill registries prevent the duplication trap where multiple teams build identical capabilities in isolation.
- Progressive disclosure ensures agents only access necessary context to reduce token costs and improve accuracy.
- Human domain experts must own specific skills across architecture and security domains to ensure long term maintenance.
- Technology alone cannot solve scaling issues without a corresponding human organizational structure for governance.
Your Code Has Bugs. Lean4 Has Proofs: Formal Verification for Engineers
Varun Pant · AI Engineer · 10 min
Watch on YouTube →Varun Pant of AWS discusses how formal verification using Lean4 can replace traditional testing to ensure AI generated code is mathematically correct. He advocates for a shift where humans define specifications while AI handles implementation and proof generation.
- Formal verification provides mathematical proof that code is correct for all possible inputs rather than just tested cases.
- The role of the human developer is shifting from writing code to owning and validating the technical specification.
- AWS uses Lean4 to verify the Cedar authorization language through 100 million differential tests performed nightly.
- An AI agent successfully rewrote zlib into Lean and generated 32,000 lines of proof in just one week.
- Project Strata aims to bring multiple programming languages into a common core for automated verification.
- Automated verification is necessary to match the high throughput of code produced by modern AI agents.
How do you diffuse AI into the real world?
Varun Shenoy · AI Engineer · 17 min
Watch on YouTube →Varun Shenoy explains why Long Lake acquires service businesses to integrate AI directly into operations instead of selling software as a vendor. This model allows for better access to off internet data and aligns incentives for deep automation.
- Long Lake has raised 3 billion dollars to acquire 35 service businesses across property management and architecture.
- Owning the company being automated removes the vendor client blame game and allows for faster iteration.
- The exceptions in a business process are often the most important tasks that require automation for true efficiency.
- Access to off internet data like physical blueprints is essential for building accurate ground truth for AI models.
- The Ladder of Autonomy describes the step by step process of moving from a Copilot to a fully autonomous AI worker.
- Extreme software service co design is necessary to change the established processes of 100 year old firms.
How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage)
Eyal Blum · AI Engineer · 17 min
Watch on YouTube →Eyal Blum from Figma outlines a strategy for adopting AI coding agents by focusing on planning and rigorous verification guardrails. He addresses the cultural challenges of integrating AI while respecting human attention and maintaining code quality.
- The best engineers are often the slowest to adopt agents because they recognize the risks to system stability.
- Figma uses a Testing Pyramid for Agent Review where deterministic analysis handles the majority of verification.
- A workflow of one week of human planning followed by one night of agent implementation can yield a 5x productivity boost.
- Attention Aware Communication requires marking AI generated content and providing a hand typed human summary.
- AI adoption follows a three act journey from early wins to a failure of trust and finally to mature verification.
- Senior skeptics should be tasked with creating the roadmap for making agents safe based on their specific concerns.
How to Win The Game of Life
Rob Dial · The Mindset Mentor Podcast · 18 min
Watch on YouTube →Rob Dial explores the gap between one's current self and highest potential self to avoid the tragedy of being successful at a life you never wanted. He provides frameworks to unlearn external conditioning and align internal identity with external habits.
- Identity acts as a governor for habits because individuals rarely act out of alignment with their self-perception.
- The 10-Minute Weekly Life Review serves as a tool to audit life choices and prevent common end-of-life regrets.
- Reframing challenges as a Soul Curriculum shifts the mindset from being a victim to being a student of growth.
- Success without fulfillment is defined as hitting professional metrics that do not align with a leader's core identity.
- The Wrong Scorecard trap occurs when professionals optimize for industry or peer standards instead of personal KPIs.
From AI-Assisted to AI-Native: Building a Frontier Development Team
Clare Liguori · AI Engineer · 20 min
Watch on YouTube →Clare Liguori discusses the transition from AI-assisted coding to AI-native frontier development where engineers delegate the majority of code production to autonomous agents. She highlights how high-leverage habits and intentional adoption lead to productivity gains exceeding 10x.
- Frontier developers shift from interactive chat to hands-off orchestration where agents run autonomously for extended periods.
- Early career engineers often struggle with high cognitive load when reviewing AI-generated code due to underdeveloped review skills.
- Flow Mat describes a specific type of burnout caused by engineers obsessing over perfect prompts for overnight agent execution.
- Teams should expect a temporary productivity dip of one to two months while preparing codebases for AI-native workflows.
- Decision-making speed becomes the primary bottleneck when code generation timelines shrink from months to days.
How to avoid disaster when vibe-coding a billing engine
Andrew Garvin · AI Engineer · 17 min
Watch on YouTube →Andrew Garvin outlines the risks and strategies for building billing engines in an era of agentic commerce. He argues that while AI can accelerate development, high-stakes systems require human-in-the-loop oversight and robust sandboxing.
- High-velocity development requires orchestrators capable of provisioning complete business environments including backends and billing in minutes.
- Critical business logic and financial movements must maintain human oversight to prevent catastrophic failures from autonomous agents.
- The rise of agents as primary users is driving a shift from seat-based pricing to usage-based credit models.
- Skills Files act as portable context providers that enable agents to navigate and interact with complex APIs effectively.
- Agents are increasingly acting as buyers by procuring their own infrastructure and services through command-line interfaces.
Productionizing LLM Gateways: Architecture, Tradeoffs and Hard Lessons
Kanish Manuja · AI Engineer · 16 min
Watch on YouTube →Kanish Manuja examines the architectural trade-offs of LLM gateways, focusing on availability, latency, and cost management. He explains why traditional reliability patterns like blind retries can be counterproductive for expensive AI APIs.
- Traditional microservice patterns like exponential backoff must be adapted for the high cost and slow response times of LLM calls.
- Reasoning models exhibit extreme latency variance, making tail hedging strategies necessary to maintain performance.
- Streaming improves user experience but removes the gateway's ability to perform mid-request fallbacks to alternative providers.
- Organizations should distinguish between centralized governance and centralized traffic to avoid creating single points of failure.
- Guardrail placement requires a strategic decision between failing open or failing closed based on specific application risks.
AI Evals for Cross-Functional Teams
Nachiket Paranjape · AI Engineer · 16 min
Watch on YouTube →Nachiket Paranjape and Swaroop Chitlur Haridas describe how DoorDash turned AI evaluation into a cross-functional process involving engineering, operations, and product management. They demonstrate how an API-first architecture allows non-technical staff to build their own annotation tools.
- AI evaluation is treated as a team sport where domain experts interact directly with technical primitives to improve quality.
- Clear organizational ownership assigns goals to strategy teams, rubrics to product managers, and labeling to operations.
- The self-serve judge calibration UI uses the JPEA library to optimize prompts against human-verified golden sets.
- Reducing friction in the evaluation loop correlates directly with increased shipping velocity and lower operational costs.
- Decentralizing UI development allows the platform to support diverse use cases like restaurant menu grading and image reviews.
Building uReview, Uber’s Multi-Agent Code Review Engine
Will Bond and Ameya Ketkar · AI Engineer · 15 min
Watch on YouTube →Will Bond and Ameya Ketkar explain how Uber built an agentic code review system to handle a tripling of developer wait times. The system uses a multi-agent architecture to process thousands of comments weekly while maintaining high developer acceptance rates.
- Uber developed uReview to address wait times for first code reviews that grew from three to nine hours over two years.
- The system processes 25,000 comments weekly and achieves a 67 percent addressal rate for its automated suggestions.
- Engineers implemented a post-processing layer to rate and deduplicate comments which reduced operational costs by 60 percent.
- The platform tracks agent trajectory to monitor the thinking process of LLMs and mitigate their inherent overconfidence.
- The project supports legacy Phabricator workflows alongside modern GitHub environments during a multi-year migration.
- The ultimate goal is an expanded outer loop where humans focus on high-level architecture while agents handle implementation details.
NVIDIA Just Made Its Biggest Purchase Ever
Josh · Limitless Podcast · 41 min
Watch on YouTube →This episode discusses NVIDIA's reported 13 billion dollar acquisition of Hugging Face and its strategic move into the software and inference layers. It also covers the rapid advancement of Chinese AI hardware and the massive cost reductions in autonomous vehicle sensors.
- NVIDIA is reportedly acquiring Hugging Face for 13 billion dollars to secure its position in the open weights market.
- The company generates approximately 1 billion dollars in cash daily and projects 70 percent growth through 2028.
- China's Z.ai GLM 5.3 Flash model successfully processed 100 trillion tokens using only Huawei hardware.
- Waymo has reduced its sensor suite costs from 125,000 dollars to 25,000 dollars while increasing compute power by 20 times.
- NVIDIA is pursuing vertical integration by providing chips, models like Nemotron, and cloud inference services.
- The circular virtuous loop involves acquiring major customers to guarantee a market for proprietary hardware.
Building the Engine While Flying the Plane: Launching the Figma MCP Server
Jesse Lumarie · AI Engineer · 16 min
Watch on YouTube →Jesse Lumarie details how a grassroots side project at Figma evolved into a critical Model Context Protocol server for the design community. The tool allows AI agents to interact directly with Figma scene graphs to bridge the gap between design and code.
- The Figma MCP server was developed as a 20 percent project by a single engineer to solve context gaps for non-designers.
- The team chose React and Tailwind representations for design nodes because LLMs are heavily trained on web code.
- Integration with Code Connect allows the system to point to existing production components instead of generating raw markup.
- The project moved from manual spreadsheet evaluations to automated LLM-judged systems running hundreds of times per week.
- A local-first deployment strategy was used to satisfy enterprise security requirements and avoid early authentication hurdles.
- The development process required navigating a volatile technical landscape where the MCP specification changed mid-build.
Why Top Founders Are Racing Into AI Infrastructure
Raghu Raghuram · a16z · 53 min
Watch on YouTube →Raghu Raghuram and the Andreessen Horowitz team discuss the launch of the Machine Age Fund to address the massive infrastructure bottleneck in AI. They argue that the industry is shifting from software constraints to capital-intensive physical resource constraints.
- Hyperscaler capital expenditure is projected to reach 1 trillion dollars annually as the industry faces a massive supply crunch.
- Token demand is growing at 1,000 percent per year as AI moves from simple chat to complex reasoning and agents.
- Founder interest in hardware and infrastructure has increased from 5 percent to nearly 30 percent of top-tier deals.
- The supply chain for critical components like chips and high-density cooling is currently booked out through 2028.
- The term Machine Intelligence is proposed as more accurate than AI because systems learn from existing human output.
- Competitive advantage is shifting toward the ability to convert massive capital into intelligence via hardware clusters.
This Free AI Just Caught The Billion Dollar Giants
Dr. Koa Eher · Two Minute Papers · 4 min
Watch on YouTube →Dr. Koa Eher reviews the Qwen 3.8-Flash-Next model, which brings high-end performance to open-weight, locally runnable software. The model uses architectural innovations to match the capabilities of much larger closed-source systems.
- Qwen 3.8-Flash-Next achieves 38 tokens per second on DGX hardware and can run on consumer-grade RTX 3090 GPUs.
- The model utilizes Quen Sparse Attention to solve the quadratic complexity issues associated with long-context memory.
- A gated residual architecture with four branches prevents layer-overwriting and improves model efficiency.
- Engram embedding allows for fast lookup of complex token combinations to improve processing speed.
- The rapid advancement of open-weight models allows companies to replace expensive SaaS subscriptions with local deployments.
- Architectural ingenuity is becoming more significant than sheer parameter count in the current phase of AI development.
If I Started A Personal Brand In 2026, I'd Do This - Dolly Parton
Alex Hormozi · Alex Hormozi · 13 min
Watch on YouTube →Alex Hormozi explains how the X-Factor serves as the ultimate competitive advantage for personal brands in an AI saturated market. He uses Dolly Parton as a case study for combining rare skills and contradictory traits to capture outsized attention.
- The X-Factor is defined as the intersection of multiple rare skills or physical traits that do not typically coexist in one person.
- Dolly Parton maintains market dominance by blending a bombshell aesthetic with business acumen and extreme humility across six decades.
- In an era where AI commoditizes content, the messenger becomes more important than the message for building a competitive advantage.
- Building a unique brand requires identifying and leaning into personal contradictions that break standard human pattern recognition.
- True exceptionalism often requires sacrificing the desire to fit in and accepting judgment for being different.
- Hormozi emphasizes that while competence is the baseline requirement, the X-Factor acts as the multiplier for brand growth.
References
PeopleAlex Hormozi · Leila Hormozi · Dolly Parton · Imad Touil · Varun Pant (x.com/varun_pant_) · Varun Shenoy (x.com/varunshenoy_) · Elad Gil · Jensen Huang · Eyal Blum · Rob Dial (coachwithrob.com) · Bronnie Ware · Ken Sheldon · Tony Robbins · Clare Liguori (x.com/clare_liguori) · Andrew Garvin · Kanish Manuja · Nachiket Paranjape (x.com/nmparanjape) · Swaroop Chitlur Haridas · Andy Fang · Raghav · Will Bond (x.com/wbond) · Ameya Ketkar · Uday · Adam · Leopold Aschenbrenner · Sam Altman · John Ternus · Josh · Jesse Lumarie (x.com/jesselumarie) · Ben Horowitz (x.com/bhorowitz) · Martin Casado (x.com/martin_casado) · Raghu Raghuram (x.com/RaghuRaghuram) · Mark Andreessen · Dr. Koa Eher · Brad · Mark · Joe Rogan · Donald Trump · Bernie Sanders
ToolsLean4 (lean-lang.org) · Cedar (cedarpolicy.com) · Strata · zlib · Rust · Slack · Figma · Stripe CLI · Metronome · JPEA library · AWS · HubSpot · SAP · uReview · Phabricator · GitHub · Claude · Hugging Face · Waymo · Z.ai · Huawei · Nemotron · MCP (Model Context Protocol) · Cursor · VS Code · React · Tailwind · Code Connect · Machine Age Fund · Qwen 3.8-Flash-Next · DGX · RTX 3090 · DeepSeek 4 Pro · Acquisition.com
PapersAnthropic: AI Skills Article · Top 5 Regrets of the Dying