Chronological feed of everything captured from John Carmack.
John Carmack is funding the donation of science fiction book sets to libraries through the LFS Prometheus Awards Collection. This initiative expands access to notable works within the genre, as evidenced by his repeated orders for book batches to fulfill demand.
John Carmack proposes that many evaluation and judging tasks, including literary awards and employee performance reviews, could be significantly improved by utilizing aggregated partial orderings and pairwise comparisons. He suggests an ELO ranking system, similar to those used in chess, to facilitate more efficient and robust assessments, particularly when comprehensive review of all items is impractical.
John Carmack suggests that GPU power draw, obtainable from tools like nvidia-smi, serves as a superior indicator of true GPU utilization compared to traditional metrics like job scheduling or "GPU busy" states. He proposes visualizing this data as heatmaps across data centers to identify inefficiencies. Carmack speculates that while inference and frontier training in major labs are highly efficient, significant GPU capacity might be underutilized due to scheduling issues and suboptimal research code.
Ablative heat shields, historically problematic for the X-15 program due to maintenance demands, are being re-evaluated in the context of modern spaceflight. While Starship’s rapid turnaround goals generally preclude ablatives, their potential for daily flight, given technological advancements, warrants consideration. This suggests a potential trade-off between maintenance intensity and reusability frequency for specific use cases.
Initial efforts to enhance value function generalization using SIGReg on prefinal activations and pre-trained LeWM have not yielded positive results. Both approaches, while aiming to induce independent Gaussian distributions or improve estimation, have demonstrably hurt generalization or resulted in worse value function estimation in early tests. Further investigation is needed as these techniques have not yet provided a performance advantage.
The provided content is a single X post from John Carmack, containing only a URL. No further information or context is available within the provided text. Therefore, a meaningful synthesis or extraction of claims is not possible without accessing the linked content.
LeWorldModel applies the LeJEPA framework to world models, demonstrating stable end-to-end joint-embedding predictive architecture from pixels. It utilizes a ViT-Tiny vision transformer for latent generation and a ViT-S backbone for prediction. The approach demonstrates effective planning at test time using the Cross-Entropy Method and shows that latent spaces capture physically meaningful quantities, offering an alternative to traditional policy-based learning in imagination.
Elon Musk exhibits a unique combination of profound technical capability, often associated with traits on the autism spectrum, and a genuine enjoyment of collaborative work. This contrasts with a prevalent pattern where highly skilled technical individuals tend towards solitude, which can hinder their overall impact. His conscientiousness further amplifies his ability to execute and build, distinguishing him from others who might share similar cognitive profiles but lack the drive or social inclination.
The content is a direct query from John Carmack asking for the source of unspecified data. No further information or context is provided, making it impossible to extract any substantive technical insights or claims related to the data itself or its implications. The content solely represents a request for clarification of a data source.
John Carmack acknowledges Palmer Luckey's superior ability to delegate, contrasting it with his own limitations in this area. This suggests Luckey possesses a crucial leadership quality that Carmack believes he lacks, particularly in brand stewardship. The statement offers insight into the self-awareness of a prominent figure in the tech industry regarding their operational strengths and weaknesses.
John Carmack, a prominent figure in the tech industry, expresses a desire for a shift in the editorial stance of leading technology publications like WIRED and MIT Technology Review. He advocates for a 'Technological Triumphalism' perspective, suggesting a more optimistic and celebratory tone regarding technological advancements. This implies a critique of current tech media for potentially lacking sufficient enthusiasm for innovation.
Bluetooth audio streaming introduces significant quality degradation due to multiple stages of lossy compression and decompression. Audio is initially stream-compressed with Vorbis or AAC, then decoded, and subsequently re-compressed with SBC or AAC for Bluetooth transmission. This multi-layered compression process, highlighted by John Carmack, suggests an inefficient system architecture ripe for optimization breakthroughs.
Traditional corporate advisory boards often serve as superficial "vibe checks" rather than providing substantive guidance. This model is inefficient as it primarily involves presentations and general discussions without directly influencing company decisions. A more effective approach involves engaging advisors on an ad-hoc basis for specific, contentious issues, leveraging their expertise as unbiased external consultants rather than as a standing board.
John Carmack, a prominent figure in open-source, views AI training on open-source code as an amplification of its inherent value, aligning with his original intent of open-source as a "gift to the world." He acknowledges the overlap between open-source and anti-AI sentiments but struggles to reconcile them given his perspective on AI's ability to magnify the utility of open-source contributions.
John Carmack critiques the current platform subsidy and 30% revenue share model, identifying it as inefficient due to "wasteful churn." He advocates for Epic's tiered revenue model, where the first $1M in annual revenue is exempt from platform fees, as a cleaner solution that directly rewards increased economic activity. Carmack suggests a "negative rate 'earned income tax credit'" for initial revenue as a strong incentive, provided exploitation by self-purchase can be prevented.
John Carmack criticizes the "Deep learning and Ethics" chapter in Prince's "Understanding Deep Learning" for its superficial treatment of bias. He highlights the distinction between "illegitimate" factors (societal choices) and "irrelevant" factors (data-driven priors), arguing that the book conflates them. Carmack proposes an alternative approach to exploring bias in ML by swapping neutral feature labels with socially charged ones to analyze generalization credibility.
John Carmack's Keen Technologies is pursuing AGI through reinforcement learning on suites of Atari games, explicitly rejecting LLMs as a sufficient path to general intelligence. His core thesis is that fundamental problems — sequential multitask learning, transfer learning, sparse rewards, and continuous forgetting — remain scientifically unsolved, and that standard RL benchmarks mask this by ignoring real-world constraints like latency and turn-based assumptions. His physical Atari robot demo exposed that state-of-the-art agents like BBF collapse under even modest added latency, and that reward extraction from raw pixels (score reading) remains a surprisingly hard systems problem. He is pushing for a new community benchmark that sequences agents through multiple Atari games and tests for genuine transfer, rather than single-game mastery.
John Carmack and Rich Sutton have partnered at Keen Technologies to focus on developing Artificial General Intelligence (AGI) based on reinforcement learning and neural networks. They aim to achieve a prototype AGI by 2030, believing a small, focused team can contribute significantly despite the dominance of larger organizations. Their approach is contrarian, emphasizing real-time, continuous learning and distributed systems over large language models and monolithic architectures.
John Carmack discusses the Meta Quest Pro, emphasizing engineering and design choices. He highlights the device's advanced display and tracking, while also addressing challenges in software optimization, particularly regarding performance, scalability, and user experience within Meta's VR ecosystem. Carmack underscores the importance of balancing quality with accessibility and the necessity of improved core system functionality for mass-market adoption.
John Carmack, a legendary programmer, discusses his journey and insights into software development, from the early days of personal computing to modern AI and VR. He emphasizes the importance of user value, highlights the trade-offs between optimization and development speed, and reflects on the evolution of programming languages and tools. Carmack also delves into the genesis of id Software and the intensive development culture that shaped groundbreaking games like Doom and Quake.
John Carmack discusses the evolution and future of VR, emphasizing its potential as a universal screen replacement beyond gaming. While acknowledging current limitations in resolution, comfort, and direct physical interaction, he highlights progress in standalone VR (Oculus Quest) and the vision for more flexible, immersive experiences. The conversation also touches on challenges like simulator sickness, safety in room-scale VR, and the long-term prospects of AR and brain-computer interfaces.