Saturday — June 28, 2025
Facebook's AI eyes private photo training, Echo Chamber Attack bypasses LLM guardrails with 90% success, and a 15-year-old develops Gofer for remote terminal control via Telegram.
News
Facebook is starting to feed its AI with private, unpublished photos
Facebook is starting to ask users to opt-in to a feature called "cloud processing" that allows the company to access and analyze private, unpublished photos from their camera roll, potentially using them to train its AI models. While Meta claims it is not currently training its AI on these photos, its terms and conditions do not provide clarity on whether unpublished photos are exempt from being used as training data, raising concerns about user privacy.
Show HN: I'm an airline pilot – I built interactive graphs/globes of my flights
The flight statistics show a total distance of 1,876,784 nautical miles, equivalent to 9 trips to the Moon or 86.9 laps of the Earth, accumulated over 5,813 hours or 242.2 days in the air. The flight log includes 2,155 flights, 1,436 landings, and 232 night landings, with the last logged flight being from Denver to London on June 20, 2025.
Judge rejects Meta's claim that torrenting is “irrelevant” in AI copyright case
A US judge has ruled that Meta's torrenting of over 80 terabytes of books to train its AI models may be relevant to a copyright case, as it could suggest bad faith and potentially profited pirate libraries. However, the authors may struggle to win this part of the case due to a lack of evidence, and the judge noted that the outcome could ultimately lead to publishers making it easier to license authors' works for AI training.
Echo Chamber: A Context-Poisoning Jailbreak That Bypasses LLM Guardrails
Researchers at Neural Trust have discovered a novel jailbreak technique called the Echo Chamber Attack, which can manipulate Large Language Models (LLMs) into generating harmful content without using explicit prompts. The attack uses a multi-stage approach, leveraging context poisoning and indirect references to guide the model into producing policy-violating responses, achieving a success rate of over 90% on several leading models, including GPT-4.1-nano and Gemini-2.0-flash.
As job losses loom, Anthropic launches program to track AI's economic fallout
Anthropic has launched its Economic Futures Program to support research and policy development on the economic impacts of AI, including potential job loss and GDP growth. The program will provide grants for research, host symposia to develop policy proposals, and build datasets to track AI's economic usage and impact, with the goal of understanding and preparing for the shifts AI will bring to the labor market and global economy.
Research
Theoretical Analysis of Positional Encodings in Transformer Models
This paper analyzes the impact of various positional encoding methods on transformer models, examining their expressiveness, generalization ability, and extrapolation to longer sequences. The study proposes new encoding methods based on orthogonal functions and finds that they outperform traditional sinusoidal encodings in experiments, providing insights for design choices in transformer applications such as natural language processing and computer vision.
LLM-Net Democratizing LLMs-as-a-Service Through Blockchain-Based Expert Networks
The development of Large Language Models (LLMs) is hindered by centralization, limited access to high-quality training data, and the complexity of maintaining expertise across various domains. The proposed LLM-Net framework addresses these challenges by utilizing a blockchain-based, decentralized network of specialized LLM providers, allowing for collective knowledge growth and sustained service quality through collaborative mechanisms and transparent transaction validation.
A Guide to Failure in Machine Learning
ML models can fail unexpectedly due to lack of reliability or robustness, and understanding the causes of failure is crucial for their adoption. This work provides a guide for practitioners to comprehend and address ML model failures by discussing key concepts, techniques, and real-world examples related to reliability and robustness.
Whois using AI to code? Global diffusion and impact of generative AI
The adoption of AI-generated coding tools is increasing, with an estimated 30.1% of Python functions from US contributors written by AI by December 2024, and this adoption is leading to significant productivity gains, with a 2.4% increase in quarterly commits for developers who use AI for 30% of their work. The value of AI-assisted coding in the US is estimated to be between $9.6-$14.4 billion annually, with potential for much higher gains if productivity effects are greater, and AI usage is also driving learning and innovation among programmers.
A Remarkably Luminous Galaxy at z=14.44
The James Webb Space Telescope (JWST) has discovered a luminous galaxy, MoM-z14, at a redshift of 14.44, just 280 million years after the Big Bang, challenging previous models of galaxy formation. This galaxy's properties, including its compact size, steep UV slope, and unique abundance pattern, suggest a rising star-formation history and may be connected to the formation of ancient stars in the Milky Way, providing insights into galaxy evolution across cosmic time.
Code
Show HN: PILF, The ultimate solution to catastrophic oblivion on AI models
There is no text to summarize. The provided message appears to be an error notification indicating that a README file could not be retrieved.
Show HN: I'm 15 and built Gofer, an AI that gets actual terminal work done
Gofer is an AI agent that runs on your computer, allowing you to control it remotely via Telegram, monitor your desktop, and execute shell commands with built-in safety protections. To use Gofer, you'll need to install it on a Linux system with Node.js, set up a Telegram bot, and configure environment variables, after which you can use various commands to interact with your computer remotely or through a local REPL interface.
Show HN: Dungeon Master in Your Console
Dungen is a generative dungeon explorer that uses AI models to dynamically create a world of mystery and peril, where the player's choices shape the story and uncover secrets. The game can be played in the console or in a web browser, and it offers various settings and experimental features, including a map generator and the option to run the AI model locally or on a serverless endpoint.
Show HN: AI-SDK-Cpp – Unified C++ SDK for OpenAI, Anthropic, and More
The AI SDK CPP is a modern C++ toolkit that enables developers to build AI-powered applications with popular model providers like OpenAI and Anthropic, providing a unified and easy-to-use API. It supports various features such as text generation, streaming, multi-turn conversations, error handling, and tool calling, with plans to add additional providers, text embeddings, and image generation support in the future.
Show HN: I built an AI-Powered semantic search for Mac
There is no text to summarize. The provided message appears to be an error notification indicating that a README file could not be retrieved.