Tuesday, August 18, 2026
The Power of High Agency and Efficient Intelligence
August 18 · 15 videos
White House strategy scales.
Generative video accelerates.
Transformers shrink 100x.
High agency wins.
“The goal is to make these models 100x smaller without losing the intelligence.”
Michael Kratsios: Inside the White House's AI Strategy
Michael Kratsios · Y Combinator · 39 min
Watch on YouTube →Michael Kratsios discusses the White House strategy for maintaining an open AI ecosystem and preventing regulatory capture by large incumbents. He highlights the shift toward private sector R&D and the need for federal standards to avoid a patchwork of state laws.
- The U.S. government manages problems at a scale that exceeds the capacity of even the largest private technology companies.
- Policy decisions are coordinated across national security and economic councils to ensure a unified approach to innovation.
- Startups should advocate for federal preemption of state laws to prevent complex regulatory barriers that favor big tech.
- The Department of Energy aims to produce a scientifically relevant quantum computer by the year 2028.
- Federal R&D funding has shifted from 70 percent of the total in the 1950s to just 30 percent today.
I'm 37. If You're In Your 20s or 30s, Watch This | Birthday Stream
Alex Hormozi · Alex Hormozi · 48 min
Watch on YouTube →Alex Hormozi reflects on his 37th birthday by arguing that adulthood is defined by commitment rather than maintaining optionality. He provides a framework for choosing which life regrets are acceptable while pursuing success through high volume and statistical certainty.
- Maturity is measured by the number of people who rely on you rather than the desire for personal celebration.
- Success is achieved by increasing the frequency of attempts until a lucky outcome becomes a statistical certainty.
- The anticipatory anxiety of a difficult task is usually more significant than the actual effort required to complete it.
- Negotiators should use the phrase what would it take to force the other party to define their terms.
- There is a 0.71 correlation between the quality of a primary relationship and an individual's subjective well-being.
Britain’s Afghanistan Shame and the Jason Arday Tragedy
Rory Stewart · The Rest Is Politics · 64 min
Watch on YouTube →Rory Stewart and Alastair Campbell analyze the long term consequences of the withdrawal from Afghanistan and the rise of global instability. They examine the human cost of gender apartheid and the impact of modern media frenzies on public figures.
- The withdrawal from Afghanistan left 2.4 million girls without access to secondary education under a new gender apartheid.
- Western sanctions often fail because they punish civilian populations while the intended leadership remains largely unaffected.
- The 20 year war in Afghanistan is estimated to have cost approximately 3 trillion dollars with 457 British military deaths.
- Modern public discourse lacks a framework for forgiveness which leads to the destruction of individuals during media frenzies.
- Dictators often ignore domestic economic impacts which makes their actions difficult for Western logic to predict.
The Next Medium: Why Real-Time Interactive Video Changes Everything: Ahmed Ahres, Reactor
Ahmed Ahres · AI Engineer · 17 min
Watch on YouTube →Ahmed Ahres explains how real time interactive video represents a fundamental shift in medium from static files to programmable software. He describes world models that allow for steering and interaction during the generation process.
- Real time capability transforms video from a passive content format into a controllable environment for training and simulation.
- Interactive world models require latency of less than 100 milliseconds to be perceived as truly real time.
- Current generative AI models act like slot machines where users must wait for a static result without mid process control.
- Evaluating the consistency and fidelity of real time models remains a significant unsolved problem in AI research.
- Programmable video enables new business models such as real time advertising based on live user intent.
Infra behind Krea 2: How to train and serve at scale: Gabriel Jorge Menezes, Krea.ai
Gabriel Jorge Menezes · AI Engineer · 16 min
Watch on YouTube →Gabriel Jorge Menezes details the infrastructure behind Krea 2 and the challenges of training diffusion transformer models at scale. He emphasizes building resilient systems that can automatically recover from hardware failures and optimize GPU usage.
- Traditional GPU utilization metrics can be misleading because they do not reflect the actual efficiency of tensor cores.
- Hardware stability is critical as a single GPU exceeding 78 degrees Celsius can destabilize an entire training run.
- Krea uses a high performance filesystem capable of writing a one terabyte checkpoint in less than 30 seconds.
- The orchestration strategy involves flipping inference workloads to external providers when internal GPUs are needed for training.
- Building systems that allow for automatic reconciliation of failed states is more effective than manual debugging at scale.
Generative Video at the Speed of Light: Keegan McCallum, uRun
Keegan McCallum · AI Engineer · 8 min
Watch on YouTube →Keegan McCallum discusses the shift in generative video from photorealistic quality to real-time efficiency and interactivity. He explains how infrastructure is evolving to support continuous video generation at costs comparable to text tokens.
- Generative video costs have plummeted to approximately 3 dollars per hour making it a viable medium for continuous interaction.
- The Helios model is a 14 billion parameter distillation designed for high speed and low latency performance.
- Market focus is moving away from high cost shot generation toward steerable and interactive video experiences.
- Real time video applications require complex infrastructure including WebRTC integration and global GPU distribution.
- The transition from a slot machine prompting style to granular steering allows for more precise creative control.
- Design partners are currently prioritized over broad user bases to refine infrastructure for real time video to video pipelines.
How to Find Consumer Startup Ideas
Dalton Caldwell · Dalton + Michael · 30 min
Watch on YouTube →Dalton Caldwell and Michael Seibel explore how modern consumer startups can succeed by democratizing luxury experiences through AI. They argue that founders should focus on providing elite level services to the masses rather than making marginal digital improvements.
- Successful consumer products often begin as high end services for the wealthy before being scaled to the general public.
- New startups must offer a 100x improvement to compete with established giants like TikTok and Instagram.
- The Eleven Star Experience framework encourages founders to design magical interactions that go far beyond basic functional needs.
- AI should be used to provide infinite personalization rather than simply replacing human workers with lower quality chatbots.
- Founders from middle class backgrounds must actively research ultra premium services to identify opportunities for democratization.
- The era of winning by simply putting a physical service online has ended because digital competition is now ubiquitous.
Voice agents with Realtime Video: Sidney Primas, LemonSlice
Sidney Primas · AI Engineer · 26 min
Watch on YouTube →Sidney Primas describes the technical journey toward creating AI avatars that pass the visual Turing test through real-time video generation. He highlights the importance of emotional intelligence in AI and the move toward cost parity with voice models.
- LemonSlice utilizes a world model approach focused on humans to generate realistic physics and micro expressions in real time.
- The architecture distinguishes between high IQ reasoning models and high EQ emotional interaction models for better human engagement.
- Error accumulation remains a significant technical hurdle for maintaining stability in infinite video generation streams.
- A major breakthrough involved reducing denoising steps from 30 down to a single step for real time performance.
- The Roosevelt avatar demonstrated the ability to generate continuous video for 8 hours without requiring a system reset.
- Providing a visual layer for AI interactions taps into biological human preferences for better understanding and focus.
While my guitar gently speaks: Todd Fisher, Philo Ventures
Todd Fisher · AI Engineer · 18 min
Watch on YouTube →Todd Fisher demonstrates a project that uses AI to turn a guitar into a conversational interface capable of speech synthesis. He details the digital signal processing challenges and the use of local models to bridge hardware and software.
- The project integrates Whisper speech to text and local large language models with a custom JUCE plugin for real time audio processing.
- Sonority peak syllabification proved more effective than energy gap segmentation for accurately slicing continuous human speech.
- Digital Audio Workstations now function as integrated development environments for creative engineers to build complex audio tools.
- The YIN pitch detection algorithm and VocalSet dataset were used to enable the guitar to simulate singing voices.
- Recent advancements in AI have significantly shortened the development cycle for complex hardware and software integration projects.
- Building public passion projects is highlighted as an effective way for engineers to demonstrate their technical skills to the community.
The Next Game Engine Won't Have a Manual: Arturo Nunez, Nereu
Arturo Nunez · AI Engineer · 19 min
Watch on YouTube →Arturo Nunez introduces Nereu, a game engine designed to replace technical boilerplate with a conceptual vocabulary based on intent. He explains how AI can simplify the game development process by focusing on the joy of making.
- The Asset Tag System allows developers to define objects by their function rather than their technical implementation details.
- Level of Detail context management prevents LLM context window bloat by prioritizing data based on the user focus area.
- Nereu uses vision models to tag a library of over 6000 assets to make them easily accessible through natural language.
- The engine aims to lower the barrier to entry so that language becomes the primary interface for complex software creation.
- AI assistants are most effective when they help developers get unstuck rather than attempting to generate entire games in one shot.
- Simplifying toolsets by removing obscure features improves the experience for the vast majority of creative users.
Building an Agentic Video Editor for Mass Consumer
Ekaterina Deyneka · AI Engineer · 12 min
Watch on YouTube →Ekaterina Deyneka explains how agentic video editing solves the friction between capturing footage and posting by treating editing as a code generation task. The approach uses agents with specific creative skills to transform raw media into polished compositions via the Remotion framework.
- Editing real footage requires high level judgment to identify successful shots and omit unsuccessful ones compared to blank canvas generation.
- Reelful codifies creative taste into agent skills including specific rules for cuts, font pairings, and b-roll placement.
- The system generates compositions as React code using the Remotion framework to ensure precise control over the final output.
- Complex agentic workflows are hidden behind mobile first templates and verification layers to make the tool accessible to mass consumers.
- The company secured funding through the A16Z Speedrun program to scale this video as code methodology.
High Agency (she wrote the book on it)
Cate Hall · Greg Kamradt · 24 min
Watch on YouTube →Cate Hall defines agency as the learned ability to recognize and act on degrees of freedom that most people overlook due to social friction. She provides a framework for building agency through deliberate exposure to rejection and prioritizing action over contemplation.
- Agency is a cultivatable skill similar to charisma that involves looking past self imposed rules and social norms.
- Exposure therapy for rejection involves setting specific goals for getting no responses to desensitize oneself to failure.
- The concept of engineering luck suggests increasing serendipity by attending unplanned meetings and engaging in public writing.
- Overthinkers should prioritize acting their way into right thinking because psychological comfort rarely precedes difficult first steps.
- High agency individuals use uncomfortability as a compass in negotiations to ensure they are not leaving value on the table.
Elon's Former Battery Chief on Making Transformers 100x Smaller
Drew Baglino · Weights & Biases · 94 min
Watch on YouTube →Former Tesla CTO Drew Baglino discusses Heron Power's mission to modernize the electrical grid using solid state electronics to support AI growth. He details how new transformer technology can significantly reduce power losses and unlock massive compute capacity in data centers.
- Heron Link transformers use Silicon Carbide and Gallium Nitride to achieve a 100x volumetric reduction compared to traditional oil filled versions.
- This solid state technology can cut grid to chip power losses by 50 percent and unlock 35MW of compute for every gigawatt data center.
- Engineering teams must decouple ego from work to allow for the immediate abandonment of months of effort if better data emerges.
- The US power grid has calcified due to decades of near zero load growth which led to domestic technology stagnation.
- Data centers are ideal utility customers because their base load profile can help stabilize the grid when integrated with energy storage.
Training Krea 2: What matters in generative model training
Sangwu Lee · AI Engineer · 21 min
Watch on YouTube →Sangwu Lee describes the data philosophy behind Krea 2 which focuses on stylistic diversity rather than the average consistency of typical generative models. The project applies LLM training paradigms to image generation to better serve creative exploration.
- Krea 2 avoids the mode collapsed consistency of models like DALL-E to provide diverse visual ideas for creative studios.
- The data pipeline processes up to 10 billion images using hierarchical deduplication based on hashes and SigLip embeddings.
- The training process mirrors LLM pipelines by moving through pre-training, mid-training, and reinforcement learning using GRPO inspired methods.
- Avoiding synthetic AI data is critical to prevent inheriting the specific aesthetic biases of other existing models.
- The team uses 30 to 40 custom in house classifiers to ensure the dataset includes high quality concepts based on Wikipedia PageRank.
Leopold's Final Portfolio Just Went Public
Josh · Limitless Podcast · 30 min
Watch on YouTube →This analysis explores a major shift in AI investment from GPUs toward infrastructure layers like memory, power, and photonics. It examines recent 13F filings including Berkshire Hathaway's significant entry into Alphabet and the risks of over leveraged positions.
- Capital allocation is rotating from simple GPU exposure into the physical plumbing of AI including interconnects and power.
- Berkshire Hathaway made a 17 billion dollar investment in Alphabet to bet on Google's vertical integration of models and silicon.
- Leopold Aschenbrenner's portfolio faced liquidation due to extreme concentration and leverage in memory stocks despite a valid thesis.
- Frontier investors are pivoting to the power and optics trade to solve data transfer bottlenecks in massive GPU clusters.
- The circular economy of AI involves NVIDIA investing in its own customers who then use that capital to purchase more hardware.
References
PeopleMichael Kratsios · Luther Lowe · Peter Thiel · David Sacks · Marco Rubio · Alex Hormozi · Leila Hormozi · David Senra · Rory Stewart · Alastair Campbell · Tony Blair · Ashraf Ghani · Jason Arday · Ahmed Ahres (x.com/Boudatw) · Gabriel Jorge Menezes (gab-menezes.github.io/) · Keegan McCallum (x.com/keeganmccallum3) · Dalton Caldwell (x.com/daltonc) · Michael Seibel (x.com/mwseibel) · Brian Chesky · Anthony Bourdain · Elon Musk · Sidney Primas · Theodore Roosevelt · Todd Fisher · Peter Frampton · Eddie Munson · Arturo Nunez (x.com/arturonereu) · Ekaterina Deyneka (x.com/katedeyneka) · Cate Hall (x.com/catehall) · Sasha Chapin · Leigh Marie Braswell (x.com/LM_Braswell) · Greg Kamradt (x.com/Greg_Kamradt) · JB Straubel · Patty Poppy · Zach Kirkhorn · Sangwu Lee (re-n-y.github.io/devlog/) · Fei-Fei Li · Leopold Aschenbrenner · Warren Buffett · Greg Abel · Gavin Baker · Brad Gerstner
ToolsKubernetes · Kueue · Virtual Kubelet · Ceph · Krea 2 · Sora · Reactor · Helios · SeeDance · Juan 2.1 · React · Python · WebRTC · ICE · TURN · Twitch · Socialcam · LemonSlice · Whisper STT · JUCE · VocalSet · Nereu · Unity · YIN algorithm · Reelful · Remotion · Heron Link · Silicon Carbide (SiC) · Gallium Nitride (GaN) · SigLip · GRPO · Alphabet · Micron · SanDisk · Coherent Corp · SpaceX
PapersNew Golden Age of American Science · Eleven Star Experience framework · You Can Just Do Things