Published on April 27, 2026 at 18:00 CEST (UTC+2)
Microsoft and OpenAI end their exclusive and revenue-sharing deal (166 points by helsinkiandrew)
Microsoft and OpenAI end their exclusive and revenue-sharing deal
Microsoft and OpenAI have reportedly terminated their exclusive partnership and revenue-sharing arrangement. This marks a significant shift in the AI landscape, as the two companies had been closely aligned since OpenAI’s early days. The breakup could reshape competitive dynamics, with Microsoft likely to invest more in its own AI efforts and OpenAI seeking new partners.
"Why not just use Lean?" (87 points by ibobev)
"Why not just use Lean?"
The author reflects on the pressure in formal mathematics circles to justify any choice other than Lean, critiquing what they see as a cult-like mindset around the tool. They trace the history of formal mathematics back to AUTOMATH in 1968 and Jutting’s 1977 formalization of Landau’s Foundations of Analysis, arguing that Lean’s achievements, while impressive, are part of a long tradition. The piece warns against forgetting earlier contributions and the value of diverse approaches.
Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview (158 points by GodelNumbering)
Show HN: OSS Agent I built topped the TerminalBench on Gemini-3-flash-preview
Dirac is an open-source coding agent that claims to reduce API costs by 50–80% while improving code quality through techniques like hash-anchored edits, parallel operations, and AST manipulation. It achieved top performance on the TerminalBench benchmark using a Gemini 3 flash preview model. The project emphasizes efficiency and context curation, offering a practical alternative to more expensive agent pipelines.
Pgbackrest is no longer being maintained (301 points by c0l0)
Pgbackrest is no longer being maintained
The maintainer of pgBackRest, a popular PostgreSQL backup and restore tool, announced the project is archived after 13 years of development. The decision was not made lightly and follows years of corporate sponsorship. The repository is now read-only, and the maintainer requests that any forks use a new name. This leaves PostgreSQL users seeking alternative backup solutions.
4TB of voice samples just stolen from 40k AI contractors at Mercor (192 points by Oravys)
4TB of voice samples just stolen from 40k AI contractors at Mercor
An extortion group, Lapsus$, leaked 4 terabytes of data from Mercor, containing voice biometrics paired with government IDs of over 40,000 AI contractors. The breach is especially dangerous because voice samples can be used for deepfake attacks. Five lawsuits were filed within ten days, alleging the company collected excessive personal data without adequate protection.
The Woes of Sanitizing SVGs (23 points by varun_ch)
The Woes of Sanitizing SVGs
The article details Scratch’s long history of SVG-related XSS vulnerabilities, stemming from the platform’s unsafe practice of parsing user-controlled SVGs into the main document. Despite building increasingly complex sanitization infrastructure, the approach is deemed fundamentally flawed. New attack vectors, like the use element, continue to bypass filters, highlighting the difficulty of safely handling user-generated SVG content.
Running Local LLMs Offline on a Ten-Hour Flight (50 points by darccio)
Running Local LLMs Offline on a Ten-Hour Flight
An engineer tested Gemma 4 31B and Qwen 4.6 36B on a MacBook Pro M5 Max during a ten-hour flight. They built a billing analytics tool and processed 4M tokens, finding local LLMs comparable to frontier models for tight-scope tasks. However, power consumption (1% battery per minute), heat (70–80W sustained), and limited RAM bandwidth were major constraints, though the offline experience was otherwise viable.
Men who stare at walls (172 points by aselimov3)
Men who stare at walls
This article explores a productivity technique where one deliberately stares at a wall to recover focus after mental drain. The author connects this to the problem of information overload, citing a study that daily information exposure has grown from 34 GB in 2008 to an estimated 87 GB today. The practice helps reduce “brain fog” by giving the brain intentional downtime away from screens and media.
Fully Featured Audio DSP Firmware for the Raspberry Pi Pico (168 points by BoingBoomTschak)
Fully Featured Audio DSP Firmware for the Raspberry Pi Pico
DSPi is an open-source firmware that turns a Raspberry Pi Pico (RP2040 or RP2350) into a low-cost digital audio processor. It functions as a USB sound card with onboard DSP for room correction, crossovers, parametric EQ, and more. The project aims to make the Pico a “swiss army knife of audio” for hobbyists and professionals alike.
Tendril – a self-extending agent that builds and registers its own tools (27 points by walmsles)
Tendril – a self-extending agent that builds and registers its own tools
Tendril is an agentic sandbox built with AWS Strands Agents SDK and Tauri. It demonstrates an “Agent Capability” pattern: when asked to perform a task, it checks its registry, and if no tool exists, it writes and registers one autonomously. The agent learns across sessions, reusing tools later. This approach reduces repetitive coding and enables dynamic capability expansion.
Fragmentation of AI partnerships and model supply chains
The Microsoft-OpenAI breakup signals that major tech companies are moving from exclusive arrangements to multi-provider or in-house strategies. For AI/ML developers, this means greater choice in foundation models but also more complexity in managing integrations, licensing, and vendor lock-in. Expect a wave of new partnerships and proprietary models as companies diversify their AI stacks.
Cost efficiency and local inference become practical priorities
Dirac’s 50–80% cost reduction and the successful offline use of local LLMs on a laptop demonstrate that efficiency is now a first-class concern. Running 31B–36B models on consumer hardware is feasible for many tasks, especially when power and heat are managed. This trend reduces dependency on cloud APIs, enables privacy-sensitive applications, and lowers the barrier for individual developers to experiment with state-of-the-art models.
Self-extending agents signal a shift toward autonomous tool-building
Tendril’s ability to write and register its own tools without human intervention represents a new paradigm: agents that dynamically expand their capabilities. This could dramatically reduce the need for pre-defined toolkits and manual integration work. For AI/ML development, such agents could automate infrastructure management, data pipeline creation, and even model fine-tuning, accelerating the pace of software engineering.
Formal mathematics and AI are converging, but cultural friction persists
The Lean debate highlights tensions between established formal methods communities and newer, hype-driven AI-assisted proof systems. As LLMs become capable of generating formal proofs, the need for rigorous benchmarks and diverse tooling (beyond Lean) grows. AI/ML practitioners should watch for models that integrate with multiple proof assistants, as this could unlock automation of verification in critical systems.
Voice biometrics and deepfake risks escalate with data breaches
The Mercor breach of 4TB of voice samples paired with government IDs underscores a dangerous trend: AI training data is becoming a high-value target for extortion. Voice cloning technology is now cheap and effective, making stolen voice prints a direct threat to identity security. For AI/ML, this reinforces the need for stricter data governance, federated learning, and synthetic data generation to minimize exposure of sensitive biometrics.
Edge AI hardware limitations are real but manageable
The local LLM experiment showed clear constraints (battery drain, heat, memory bandwidth) even on top-tier hardware. Yet the work was still productive for focused tasks. This suggests that edge AI is viable for specific use cases but not a complete replacement for cloud computing. Developers should optimize for efficient token generation and leverage specialized hardware (e.g., NPUs) as they become available in consumer devices.
Open-source agent frameworks are maturing rapidly
Both Dirac and Tendril are open-source projects that directly compete with or complement commercial offerings (e.g., from OpenAI, Anthropic). Their emergence indicates a healthy ecosystem where community-driven innovation can challenge proprietary solutions. For AI/ML engineers, investing in open-source agent frameworks reduces risk of vendor lock-in and allows customization for niche domains like backup management, audio processing, or SVG sanitization (as seen in other articles).
Analysis generated by deepseek-reasoner