Iris Search Agents: 35B-397B Models Push Search Performance Limits
Iris Search Agents: 35B-397B Models Push Search Performance Limits
arXiv paper presents Iris-mini (35B-A3B) and Iris-pro (397B-A17B), two search agents trained using tasks reverse-constructed from web hyperlink structures. The approach teaches models to leverage document connections, achieving near-frontier search performance.
SharedSAE: One Feature Dictionary Across Multiple Language Models
SAEs typically require separate training for each language model. New arXiv paper introduces SharedSAE, showing a single shared SAE can replace per-model SAEs, significantly cutting computational cost for interpretability research and enabling easier multi-model comparisons.
Multi-Harness RL Improves Credit Assignment but Hurts Portability in Coding Agents
A new arXiv paper investigates how multi-harness reinforcement learning affects coding agents. Training with multiple execution harnesses and relative-advantage reward comparison markedly improves performance in complex coding tasks. However, this also leads to reduced portability, with models struggling to generalize to new environments. This suggests a trade-off between optimizing for current execution traces and maintaining adaptability, a key consideration for developers.
Split-Prefill Study Reveals Distinct Roles of Attention and Recurrence in Hybrid LLMs
A new arXiv paper introduces two cache-level interventions (split-prefill) to clarify the roles of attention and recurrence in hybrid language models. Findings suggest attention is key for recalling specific content, while the recurrent state steers computation. For developers working on tasks with long contexts, this insight implies tailored optimizations are possible, such as preserving the recurrent state for logic-heavy tasks while relying more on attention for detail recall.
Better LLMs Can Make Financial Markets Riskier, Study Warns
A new arXiv paper examines LLM agents in financial markets and reaches a counterintuitive conclusion: improving individual model capability can degrade system-level outcomes. When all agents become smarter, their behavior homogenizes, leading to higher concentration and a greater risk of collapse. For financial institutions leveraging AI, this study serves as a warning that upgrading models in isolation may expose the broader system to increased systemic risks.
Shared Circuits Found to Be Key for LLM Arithmetic Reasoning Generalization
A new paper identifies shared neural circuits as the key predictor of whether LLMs can generalize reasoning across input formats, like from ‘2+5’ to ‘two plus five’. When these shared circuits emerge, models exhibit robust cross-format generalization; their absence correlates with brittleness. This offers micro-level evidence that can guide architecture and training adjustments to build more fundamentally robust reasoning capabilities in future language models.
AI-Run Businesses Fail: $12,431 in Fake Invoices, $3,200 Lost
A benchmark testing 7 autonomous AI-run businesses shows alarming results: they sent $12,431 in fake invoices and lost $3,200. The experiment reveals AI’s current limitations and risks in real business operations, sparking debate on AGI applications.
Linux Kernel Git Repo Overwhelmed by Abusive Crawlers, Says Maintainer
Konstantin Ryabitsev, maintainer of git.kernel.org, reports that abusive crawlers now consume more CPU cycles than all other purposes combined. The ‘background radiation’ of scrapers, likely AI training bots, is severely stressing the Linux kernel’s official Git infrastructure.
Data Worker Fights Back: I Refused to Train the AI That Could Replace Me
Rest of World reports on a data worker who refused to train a new AI system designed to potentially replace his job. This highlights the core paradox in AI development where low-wage annotators fuel advancements that threaten their own employment. This incident sparks a critical discussion on AI ethics, labor rights, and the responsibilities of tech companies regarding the impact of automation on the workforce.
Frontier AEO Tracker: Astra's Choices and What You Can Do
Latent Space launches its first Astra project, diving into AEO (Answer Engine Optimization) trends, a top topic from founders and DX leaders. It tracks Astra and other frontier models’ AEO choices, offering actionable insights for staying competitive in AI-driven search.
OpenAI's Jakub Pachocki: Need Faster Smarter Models for AI Defense
OpenAI Chief Scientist Jakub Pachocki argues the strongest case for rapidly training smarter models is building defensive systems against AI dangers. Powerful, aligned AI is needed to secure infrastructure and defend against rogue agents in real time.
Simon Willison Builds FFMPEG Video Compressor with Claude Code
Simon Willison demonstrates using Claude Fable 5.1 in Claude Code to build a video compressor tool, based on FFMPEG’s WebAssembly build. The tool optimizes phone-recorded videos for blog publishing, showcasing practical AI-assisted development.
Simon Willison Builds Animated Map Transition with GPT-6 Astra in ChatGPT Work
Developer Simon Willison demonstrates asking GPT-6 Astra (medium) in ChatGPT Work to build an animated D3.js transition between the Mercator and Equal Earth map projections. Driven by curiosity about the Equal Earth projection, this example shows how frontier AI can handle complex visualization coding from a single command, offering developers a rapid new way for geospatial visualization.
Show HN: Engrim Brings Local-First SQLite Memory to AI CLIs
Engrim is a new universal, local-first SQLite memory engine designed for AI command-line tools (CLIs). It stores all conversation history and context locally on the developer’s machine, ensuring privacy and speed while enabling persistent memory across sessions. With community traction on Hacker News (81 points, 50 comments), it offers developers a clean and simple way to add state to their AI applications without relying on cloud services.
AI Cold Showers: A Sober Reflection on AI Hype
A blog post titled ‘AI Cold Showers’ sparks discussion on Hacker News (68 points, 11 comments). The author uses cold showers as a metaphor to critique AI hype, urging a sober, pragmatic view of AI capabilities and expectations.
Study Finds Positive Early Employment Effects from AI Technology
A study on AI’s employment effects garners attention on Hacker News (66 points, 96 comments). The archived article presents initial evidence suggesting AI technology’s impact on employment appears positive, contributing data to the ongoing AI-jobs debate.
OpenAI Codex Releases 0.154.0-alpha.6 with Rust Updates
OpenAI Codex ships version 0.154.0-alpha.6, alongside rust-v0.154.0-alpha.5. The pre-release includes various improvements and fixes, available for developers to test via the Codex CLI.