Saturday — June 21, 2025
Phoenix.new enhances real-time collaborative coding for Elixir, ZenStack V3 evolves the TypeScript ORM landscape with a Prisma-free approach, and language models face challenges identifying missing information.
News
Phoenix.new – Remote AI Runtime for Phoenix
Chris McCord, the creator of Elixir's Phoenix framework, has introduced Phoenix.new, a fully-online coding agent tailored to Elixir and Phoenix that allows for collaborative, real-time application development. Phoenix.new provides a root shell in an ephemeral virtual machine, enabling agents to install packages, run programs, and interact with web applications in a real browser, making it a powerful tool for building and deploying applications quickly and efficiently.
Meta announces Oakley smart glasses
Meta has announced a new line of smart glasses in partnership with Oakley, featuring a front-facing camera, open-ear speakers, and microphones, with the ability to capture 3K video and translate languages. The glasses, which start at $399, are geared towards athletes and offer double the battery life of Meta's previous smart glasses, with a limited-edition model available for $499.
How malicious AI swarms can threaten democracy
The emergence of malicious AI swarms poses a significant threat to democracy, as these systems can coordinate covertly, spread disinformation, and evade detection, leading to fabricated consensus, harassment, and erosion of institutional trust. To counter this threat, a three-pronged response is proposed, including platform-side defenses, model-side safeguards, and system-level oversight, such as a UN-backed AI Influence Observatory, to protect democratic processes from the potential harm caused by malicious AI swarms.
AI is going to hack Jira
The current approach to managing engineering teams, known as Big Agile, focuses on measuring productivity through metrics like new features and deployment velocity, but this approach is flawed as it ignores the critical work of maintaining and evolving complex systems. The increasing use of AI in engineering will exacerbate this problem, as AI excels at producing surface-level features but lacks the expertise and understanding to manage complex systems, leading to potential disasters when human engineers are replaced or marginalized.
Jürgen Schmidhuber:the Father of Generative AI Without Turing Award
Professor Jürgen Schmidhuber, a pioneer in artificial intelligence, has made significant contributions to the field, including the development of Long Short-Term Memory (LSTM) networks and laying the foundations for Generative AI, yet he has not received a Turing Award, sparking a discourse about his overlooked achievements. Schmidhuber's work has had a profound impact on the AI industry, and he continues to advocate for scientific integrity and recognition of underappreciated contributors to the field, while also sharing his vision for the future of AI, including the potential for self-replicating, self-improving machine civilizations.
Research
Approximating Language Model Training Data from Weights
Modern language models can have their weights openly available, but the training data used to create them remains closed. Researchers have developed a gradient-based approach to approximate the original training data from the model weights, which can recover useful data and achieve performance close to the original model, even when using only publicly available text corpora.
Who is 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.
AbsenceBench: Language models can't tell what's missing
Large language models struggle to identify missing information in documents, achieving only 69.6% accuracy on the AbsenceBench test, despite excelling at recalling surprising information. This poor performance is attributed to a fundamental limitation in Transformer attention mechanisms, which cannot easily attend to "gaps" in documents, highlighting an unexpected breakdown in model capabilities.
Emergence of Diffusion Models from Associative Memory
Hopfield networks, a type of associative memory system, exhibit a phenomenon where excessive training data leads to the emergence of "spurious states" or unintended stable points, resulting in incorrect recall. Similarly, diffusion models, used in generative modeling, also exhibit a transition from a memorization phase to a generalization phase as the training data increases, with "spurious states" emerging at the boundary of this transition, corresponding to new attractor states not present in the training set.
From Bytes to Ideas: Language Modeling with Autoregressive U-Nets
The traditional tokenization approach, such as Byte Pair Encoding (BPE), imposes a fixed granularity on input text, limiting a language model's ability to operate on data. An autoregressive U-Net is introduced, which learns to embed its own tokens as it trains, allowing for a multi-scale view of the sequence and enabling the model to handle character-level tasks and low-resource languages.
Code
Show HN: Nxtscape – an open-source agentic browser
Nxtscape is an open-source browser that utilizes AI locally, providing a privacy-first alternative to other browsers, and allows users to automate tasks with AI agents while keeping their data secure on their computer. The browser is community-driven, compatible with Chrome extensions, and offers features such as a built-in AI ad blocker and support for local AI models, with the goal of reimagining the browsing experience with AI-powered automation.
Show HN: Ts-SSH – SSH over Tailscale without running the daemon
Ts-ssh is a powerful command-line SSH client and SCP utility that connects to Tailscale networks, offering features like multi-host operations, batch command execution, and real tmux integration. It provides a streamlined and efficient way for DevOps teams to access and manage their Tailscale infrastructure, with support for various authentication methods, secure host key verification, and direct SCP transfers.
Show HN: SnapQL – Desktop app to query Postgres with AI
SnapQL is a local desktop app that allows users to explore their PostgreSQL database in seconds, generating schema-aware queries with AI and supporting any PostgreSQL database without exposing credentials. To use SnapQL, users can currently build it locally by cloning the repo, installing dependencies, and running a build command, with precompiled binaries to be shipped in the future.
Show HN: Tool to Automatically Create Organized Commits for PRs
Git Smart Squash is a tool that uses AI to transform messy commit histories into clean, logical commits by analyzing the entire diff and grouping related changes together. It can be installed and set up in minutes, and provides a safe and customizable way to reorganize commits, with options for local or cloud-based AI providers and easy recovery of original commits.
Show HN: ZenStack V3 – TypeScript ORM Combines Strength of Prisma + Drizzle
ZenStack V3 is a TypeScript database toolkit for developing full-stack or backend Node.js/Bun applications, offering a unified data modeling and access solution with features like a schema-first ORM, data access APIs, and built-in access control. It is currently in alpha phase and has been rewritten to no longer depend on Prisma at runtime, instead using Kysely as its ORM engine, while maintaining compatibility with Prisma's schema and API.