Published on June 26, 2026 at 06:00 CEST (UTC+2)
Om Malik has died (598 points by minimaxir)
Om Malik, a well-known San Francisco-based writer, photographer, and investor, passed away on June 24, 2026, at Stanford Hospital after a long heart-related health journey. The article is a tribute from his family, inviting readers to share remembrances. Malik was highly regarded for his insightful commentary on technology and innovation, particularly in the early 2000s, and his photography was widely admired.
Apple to skip high-end M6 Mac chips in favor of AI-focused M7 line (135 points by scrlk)
Bloomberg reports that Apple is skipping the planned high-end M6 Mac chips and shifting focus to an M7 line that prioritizes AI capabilities. The M7 Pro, M7 Max, and M7 Ultra chips are expected to be designed specifically for AI workloads, signaling a major strategic pivot for Apple’s silicon roadmap. This move underscores the growing importance of on-device AI processing over raw CPU/GPU performance for traditional computing.
An entire Herculaneum scroll has been read for the first time (1063 points by verditelabs)
For the first time, a complete Herculaneum scroll (PHerc. 1667) has been read without physically unrolling it, using advanced virtual unwrapping and AI-based imaging techniques. The scroll, carbonized by the eruption of Mount Vesuvius in 79 AD, was digitally unrolled and its Greek text reconstructed in full. This breakthrough, achieved by the Vesuvius Challenge community, opens the door to reading hundreds of other sealed scrolls without damaging them.
No-One Escapes the Permanent Underclass (12 points by hamish-b)
The article argues that if AI reaches human-level or better performance on all cognitive and physical work, no one—including billionaires and AI lab founders—will escape the resulting permanent underclass. It contends that owning property, stocks, or even guns will not protect anyone from being replaced by cheaper, more capable machines. The author extends the argument to governments, suggesting they too would be disempowered.
Libre Barcode Project (17 points by luu)
The Libre Barcode Project provides open-source fonts that allow users to generate scannable barcodes in common formats like Code 39, Code 128, and EAN/UPC. These fonts can encode text directly into barcode images, with or without human-readable text below. The project includes an online encoder tool and downloadable font files for use in documents or labels.
Framework's 10G Ethernet module exposes USB-C's complexity (68 points by Alupis)
Jeff Geerling tests Framework’s new 10G Ethernet Expansion Card, which uses USB-C connectivity but is often bottlenecked by USB-C’s bandwidth complexity. The Realtek RTL8159 chip requires USB 3.2 Gen 2x2 (20 Gbps) to achieve full 10 Gbps speeds, but many Framework laptops only support slower USB standards, capping throughput well under 10 Gbps. The article highlights the practical challenges of USB-C’s fragmented ecosystem.
The 'papers, please' era of the internet will decimate your privacy (490 points by bilsbie)
The article warns that age-verification laws—requiring users to upload government IDs or facial scans to access social media or websites—are creating a “papers, please” internet that destroys privacy. Even if you are over 16, you may be forced to verify your identity to post or comment, and the collected data could be abused by third-party verification apps or leak in breaches. The author notes this trend is spreading globally, including in the United States, and presents serious risks for free speech and anonymity.
The Garbage Collection Handbook: The Art of Automatic Memory Management (2nd Ed) (2023) (78 points by teleforce)
The Garbage Collection Handbook (2nd Edition, 2023) is an authoritative guide covering automatic memory management techniques, from simple algorithms to advanced parallel, concurrent, and real-time collectors. The updated edition addresses new challenges from modern hardware and software environments. The book is presented as essential for programmers, especially given the near-universal adoption of garbage collection in modern languages.
Overfitted a 900KB Transformer to Compress a 100MB CSV into 7MB (38 points by spidy__)
A developer overfitted a small 900KB transformer model to memorize a single 100MB CSV file, then used arithmetic coding to compress it to 7MB (a 93% reduction). The same approach compressed a 100MB slice of enwik9 to 21MB. Although slow (about 2 hours total on an AMD 7800XT), the experiment demonstrates the potential of overfitted neural models for file-specific compression, building on earlier work by Fabrice Bellard.
Un-0: Generating Images with Coupled Oscillators (124 points by babelfish)
Unconventional AI introduces Un-0, an image generator that uses a simulated system of coupled oscillators instead of a conventional deep neural network. Running on a physics-inspired computer, it achieves an FID of 6.74 on ImageNet 64×64, matching earlier state-of-the-art generative models. The approach aims to be 1,000x more energy-efficient than GPU-based AI by letting physics do the computation. All code and weights are open-source.
AI Hardware Specialization Accelerates
Apple’s decision to skip the M6 high-end Mac chips in favor of an AI-focused M7 line reflects a broader industry shift toward custom silicon tailored for machine learning workloads. This trend will likely accelerate as companies prioritize on-device inference, training efficiency, and energy efficiency over general-purpose CPU/GPU performance. Implication: Developers and cloud providers should prepare for more heterogeneous hardware ecosystems and invest in AI-optimized software stacks.
Physics-Based Computing as a Disruptive Alternative
Un-0’s coupled-oscillator image generator demonstrates that non-GPU, physics-driven computation can match conventional AI performance at a fraction of the energy cost. This signals a potential paradigm shift in AI hardware, moving beyond von Neumann architectures to analog or physics-based substrates. Implication: Startups and researchers should explore energy-efficient physical computing for training and inference, as it could drastically reduce the carbon footprint of large-scale AI.
AI-Powered Archaeology and Digital Preservation
The complete reading of a Herculaneum scroll via virtual unwrapping and AI shows how machine learning can unlock inaccessible historical and scientific data. This technique can be applied to other damaged or sealed artifacts, from ancient manuscripts to fossils. Implication: Cross-disciplinary collaborations between ML researchers and domain experts (archaeology, paleontology, materials science) will yield high-impact discoveries and potentially new funding streams for AI.
Overfitted Models for Compression: A Niche but Promising Use Case
The 900KB transformer that compressed a 100MB CSV to 7MB illustrates that extreme overfitting can be exploited for compression, especially for large, repetitive files. While impractical for real-time use, this approach could be valuable for archiving, storage, or transmission of large datasets where decompression speed is not critical. Implication: Further research into neural-compression hybrids (e.g., combining overfitted transformers with arithmetic coding) could lead to specialized compressors for medical imaging, scientific data, or log files.
Societal Fears About AI-Driven Economic Inequality Are Intensifying
The “permanent underclass” article captures a growing anxiety that automation will render nearly all human labor obsolete, leaving even the wealthy and powerful disempowered. This narrative is gaining traction on platforms like Hacker News, indicating that the public debate is shifting from “will AI take jobs?” to “what happens when no job is safe?” Implication: Policymakers and tech leaders must proactively address universal basic income, wealth redistribution, and governance of AGI to avoid a backlash that could stifle innovation.
Age Verification Laws Threaten Anonymity and Privacy in AI-Mediated Spaces
The “papers, please” article highlights the tension between child safety regulations and the erosion of online privacy. As governments mandate identity checks for social media and forums, AI systems will increasingly be used for facial recognition, document verification, and behavioral profiling. Implication: AI researchers and companies should advocate for privacy-preserving age verification techniques (e.g., zero-knowledge proofs, on-device verification) and anticipate stricter data protection laws that could limit AI training on user-uploaded IDs.
Garbage Collection Remains a Core Concern for AI/ML Systems
The updated Garbage Collection Handbook, while not directly about ML, is highly relevant because memory management is a bottleneck in large-scale AI training and inference (e.g., transformer models, GPU memory). As models grow, efficient automatic memory management (e.g., concurrent, real-time, and region-based collectors) becomes critical for performance and stability. Implication: ML engineers should invest in understanding GC trade-offs—especially for long-running training jobs and latency-sensitive inference—and consider languages/frameworks with advanced memory management (Rust, Julia, or custom allocators).
Analysis generated by deepseek-reasoner