You.com Founder Socher's New Startup Recursive Valued at $5B for RSI
You.com Founder Socher's New Startup Recursive Valued at $5B for RSI
Richard Socher, an NLP veteran and CEO of You.com, has spun out a new startup called Recursive focused on recursive self-improvement (RSI), already valued at $5 billion. Socher is an early figure in natural language processing, and the new venture targets AI systems that can iteratively improve themselves. RSI is seen as a key path toward more capable AI, and the company’s valuation signals strong investor appetite for the direction.
The Cost of Compression: Rate-Distortion Limit on Factual Hallucination
A new arXiv paper argues that factual hallucination in closed-book QA is not just a coverage problem. Even when a fact has been observed during training, lossy compression of that fact introduces a second error source, yielding a rate-distortion lower bound on hallucination. The work separates hallucination into insufficient coverage and information lost through compression. The implication: simply adding more facts to training data cannot push hallucination below a certain limit. Developers evaluating QA models should distinguish “never seen” from “seen but corrupted” errors, since the two require different fixes.
Harness or Model? Study Isolates Framework Effect in Agentic Coding
A new paper introduces a contamination-controlled private suite to isolate the “harness effect” in agentic coding systems. These systems couple a language model with a harness — tools, prompts, and control flow — and vendors typically tune harnesses to their own models. The study aims to determine whether performance gains come from the model itself or from harness design.
Pre-Action Verification for LLM Agents Cuts Silent Failures
A new paper proposes pre-action verification for LLM agents. Agents act on the world by emitting actions such as shell commands and edits; a wrong action does not always fail loudly and can produce a plausible but incorrect effect with no error raised. The paper argues for inserting cheap deterministic checks before actions execute to catch these silent failures.
Occamy-1.0: Open 35B Pareto-Frontier Model for Co-Work Agents
An arXiv paper introduces Occamy-1.0, an open 35B model positioned on the Pareto frontier for co-work agents. Co-work agents execute complex workflows combining information gathering, tool use, coding, and file manipulation across many model invocations. Because cost and latency accumulate over the full episode, their practical value depends on more than peak performance. The model aims to balance performance and cost.
GAUGE: When Not to Trust LLM-as-a-Judge in User-Simulated Agent Evaluation
A new arXiv paper, GAUGE, challenges the low-cost offline evaluation pipeline widely used for task-oriented LLM agents. The common setup has persona-driven LLM user-simulators converse with each candidate agent, an LLM-as-a-judge scores the transcripts, and the highest scorer wins. The paper identifies systematic biases in this pipeline that can promote agents that are not actually the best performers. It offers a method for determining when LLM judges should not be trusted, giving teams a way to sanity-check their evaluation choices. For developers relying on offline evals for model selection, existing automated ranking results may need re-examination.
AI SDK Harness Layer Adds Native Subscription Auth for Coding Agents
Vercel’s AI SDK harness layer now supports authenticating harnesses through their native subscriptions where the underlying harness allows it. The layer runs different coding agents through the same HarnessAgent interface, letting developers switch agents without changing application code. This simplifies authentication for multi-agent setups and lowers integration overhead for teams that need flexibility across coding agents.
Claude Code v2.1.271 Adds Fast Mode for Remote Sessions and Mouse Support
Claude Code released v2.1.271. Remote sessions (cloud and self-hosted runners) gain fast mode: the host’s fast-mode setting or typing /fast in the session applies where the organization allows it. The /config panel in fullscreen mode now supports mouse input, letting the wheel scroll the settings list and clicks operate options. The changes make remote usage and configuration more convenient.
Fyxer Builds AI Executive Assistant Using OpenAI Models and Fine-Tuning
Fyxer uses OpenAI models combined with fine-tuning, memory, and real user feedback to build an AI executive assistant. It organizes inboxes and drafts emails in each user’s own voice, improving personalization by continuously learning user preferences. For knowledge workers drowning in email, the tool’s practical value lies in cutting the time spent manually sorting and writing messages.
Hacking AI Customer Service Agents: Intigriti Publishes Security Research
Security research platform Intigriti published an article on hacking AI customer service agents, exploring penetration techniques against these systems. As more companies delegate customer service to AI agents, their security becomes a real risk area. The article analyzes potential vulnerabilities from an attacker’s perspective, offering reference value for security teams deploying or evaluating AI customer service.
AI Doom Rhetoric Is Just Another Form of Hype
Erkan Saka argues in a blog post that AI “doom” and existential-risk rhetoric functions as a form of hype. The piece contends that safety language is used to attract attention and shape public perception rather than genuinely advance safety research. The post drew 122 points and 170 comments on Hacker News.
Laurie Voss: Coding Is Cheap Now, Defining What to Build Is the Hard Part
Laurie Voss argues that the cost of writing code has collapsed, and the cost of reviewing, fixing, and operating it is following. What remains in software development, he says, is finding out what people actually want, defining it precisely, and making it pleasant to use. That cost is per piece of software and does not scale away through automation.
Bryan Cantrill on AI Fear: 'The Contagion of Fear' Spreads Among Researchers
Bryan Cantrill responds to a tweet by former Anthropic employee Jacob Coxon, who confirmed that many Anthropic researchers believe AI “could kill us all by the end of the decade.” Cantrill shares a story of his own youthful mistakes that caused unjustified panic among less technical people, drawing a parallel to how AI fear spreads. His central point: excessive panic over AI risk can spread contagiously through groups, producing irrational judgments.
Simon Willison on the Blog Posts That Shaped His Thinking
Developer Simon Willison joined a Lobste.rs discussion on the blog posts that most influenced readers’ thinking. He cited Joel Spolsky’s early essay “The Law of Leaky Abstractions,” which he read early in his career and which encouraged him to always seek a better understanding of the layers beneath the one he works at. The thread collects recommendations from multiple engineers.
Adversarial Fashion Makes a Statement Against AI Surveillance
IEEE Spectrum reports that designers are using adversarial fashion to counter AI surveillance systems. These garments use patterns and textures designed to disrupt computer vision recognition, making the wearer harder to detect automatically. The story drew 92 points and 44 comments on Hacker News.
China Regulators Target "AI Boyfriend" Chatbots
IEEE Spectrum reports that Chinese regulators are taking aim at “AI boyfriend” emotional-companion chatbots. The story received 54 points and 49 comments on Hacker News. Specific details of the regulatory measures are not elaborated in the summary.
The AI Job Market in 2026: What's Growing
A blog post analyzes the landscape and trends of the AI job market in 2026. The piece drew 63 points and 68 comments on Hacker News, indicating strong interest. Specific job data and conclusions require reading the original post.
Google DevFest 2026 Returns with 800+ Global Events on Agentic AI
Google announced that DevFest 2026 is back, with more than 800 events planned worldwide. This year’s theme centers on building, securing, and scaling in the agentic AI era. Developers can use these events to connect with the community and learn about related technologies. Google has not yet released specific schedules or technical details.
Constant-Size Cache Boosts Block Diffusion at Scale
A new paper examines caching in block diffusion language models. Diffusion models decode tokens in parallel, but their bidirectional denoiser rules out the naive key-value cache used in fast autoregressive inference. Block diffusion restores caching by decoding block-by-block. The paper analyzes what a constant-size cache buys block diffusion at scale.
Astronaut Christina Koch and Google's James Manyika Talk Space and Tech
Astronaut Christina Koch sits down with James Manyika, Google’s Senior Vice President of Research, Labs, Technology & Society, to discuss space exploration, technology, and discovery. The conversation pairs first-hand spaceflight experience with a tech industry research perspective. For those interested in the intersection of space technology and AI, the cross-disciplinary dialogue offers views from both practitioners and researchers.