Article URL: https://omarchy.org/news/2026/08/omacom-foundation-launches-with-8-million/ Comments URL: https://news.ycombinator.com/item?id=49390132 Points: 11 # Comments: 2
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
A document-analytics startup that tracks who reads shared PDFs found its viewer counts were being inflated by automated systems including corporate email security scanners, link-preview bots from Slack and LinkedIn, and uptime monitors. To filter these out, the team built a weighted scoring system that flags sessions as suspicious when bot-like signals — such as missing user agents, zero engagement, or sub-second page views — cross a threshold of 60 points. The process revealed unexpected edge cases, including the Android phone brand CUBOT whose device name contains the substring 'bot', causing real users to be misclassified as crawlers. The team also discovered that a 1-millisecond dwell-time floor was being triggered by server-side processes rather than genuine human interaction, leading them to raise the minimum plausible reading duration to 10 seconds. The project highlighted that adding evidence meant to exonerate real users sometimes backfired, making legitimate sessions appear even more bot-like due to inverted scoring logic.
Enterprise AI systems commonly maximize the data fed to agents, but a new technical perspective argues this approach is flawed and potentially risky. An agent's effective authority is shaped by both the information it can access and the actions it is permitted to take, meaning context control is as critical as capability control. A synthetic experiment called Token-Bleed R5 found that compact, governed context selection used up to 97.9% fewer prompt tokens than full-context stuffing while achieving higher accuracy. However, the same experiment revealed that governed selection still consumed nearly seven times more tokens than a simple lexical baseline, which itself scored zero on quality metrics, highlighting real cost trade-offs. The recommended architectural approach separates agent workflows into four planes — connection, context, capability, and evidence — treating the context layer as a formal authority control rather than a mere technical convenience.
Lightricks launched LTX-2.5 on August 11, 2026, introducing a novel architecture where the VAE decoder itself operates as a diffusion model rather than a fixed deterministic component. The system uses an aggressive 1:192 spatiotemporal compression ratio and relies on the diffusion video decoder to recover fine details — such as readable text and sharp textures — that high compression typically discards. A second feature called Diffusion Fidelity Rendering allocates compute adaptively across frames and regions based on scene complexity, enabling a 10-second 720p clip to be generated in 6.8 seconds on two NVIDIA GB200 GPUs. LTX-2.5 also replaces smaller T5 or CLIP text encoders with a fine-tuned Gemma 4 12B model paired with a custom prompt enhancer, improving coherence in complex, multi-subject prompts. In Lightricks' own benchmarks, the model achieved an artifact score of 0.28 on a 98-prompt evaluation suite, outperforming Flux 3 at 0.45 and Veo 3.1 at 1.20.
A developer created MmOrganized, an AI-powered personal memory app, after finding that mainstream tools like ChatGPT and Gemini store user data in opaque, uneditable systems. The app allows users to log personal information — such as health records and vaccine dates — through a chat interface and retrieve it later using natural language queries. Unlike existing AI memory features, MmOrganized lets users view, edit, reorder, and delete every saved memory entry at any time. Users can also import and export their entire memory store as a plain text file, ensuring full data portability and no platform lock-in. The project began as an MVP inspired by the developer's personal need to track a family member's vaccination history and has since become a tool they rely on daily.
Discussion (0)
Log in to join the discussion and vote.
Log in