AI

AI Engineers aren't safe from being replaced by AI (dmanco.dev) AI

The post argues that “AI engineers” may be among the jobs most likely to be replaced sooner than other developer roles because broad, general foundation models are increasingly being applied across many AI domains, reducing the need for specialized human-built solutions. The author also notes that the job market for “AI engineer” is confused by the umbrella use of “AI,” and concludes that while AI may not fully replace developers because humans still need to define and integrate applications, many AI engineering tasks could be automated.

Introducing RadixAttention to Trellis (trellis.unfoldml.com) AI

Trellis introduces RadixAttention, a KV-caching approach that uses a radix tree to reuse cached key/value activations for shared prompt prefixes (common in chat systems), reducing redundant prefill compute and improving throughput, latency, and memory usage. The post details the block-paged cache design for concurrent requests and reports benchmark results showing faster (30–40%) and more memory-efficient inference as the shared-prefix fraction increases.

Leiden Declaration on Artificial Intelligence and Mathematics (leidendeclaration.ai) AI

The Leiden Declaration on Artificial Intelligence and Mathematics calls on the mathematical community to address risks from AI tools—such as unreliable or opaque proofs, weakened attribution, and shifting incentives—while urging transparent tool disclosure, support for peer review, adherence to open-science principles, and continued human responsibility for correctness and authorship.

Type-Error Ablation and AI Coding Agents (arxiv.org) AI

The paper titled “Type-Error Ablation and AI Coding Agents” (arXiv:2606.01522) examines how removing or varying type errors affects the performance of AI coding agents, based on the information available from the title and URL alone.

Ad Infini­Tum (matthiasott.com) AI

A blog post argues that Google’s new “generative UI” and Gemini-based “AI Mode” search replace traditional blue links and ad slots with dynamically generated, personalized answers where paid placements are integrated as “Highlighted Answers,” raising concerns about monetization, privacy opt-in, and whether users will want this model.

Skillhound: Give your AI access to every public SKILL.md (skillhound.ai) AI

Skillhound describes a search engine that aims to give AI access to public SKILL.md files across GitHub.

Dense Retrievers Know More Than They Can Express (mixedbread.com) AI

Mixedbread argues that dense retrieval models contain richer, “latent” information than their scoring operators can express, and that this hidden vocabulary can be extracted using sparse autoencoders to obtain features that resemble lexical signals useful for retrieval.

1 in 5 U.S. Adolescents and Young Adults Use AI Chatbots for Mental Health (rand.org) AI

A RAND report says nearly 1 in 5 U.S. adolescents and young adults use AI chatbots for mental health, based on the study referenced by the article's title and URL.

Google Is Quietly Buying Code from Play Store Developers to Train AI (404media.co) AI

404 Media reports that Google is inviting some Google Play app developers to join a “confidential content offer pilot” that would pay them for access to their app codebases and archived projects to help improve Google’s developer tools and AI products, while keeping developers’ IP and retaining non-exclusive licensing.

I Don't Want My Search Engine to Think for Me (searchzee.com) AI

The SearchZee blog argues against AI “overview” or summary boxes in search results, saying they hide uncertainty and disagreements, increase the risk of users acting on confident-but-wrong synthesis, and reduce traffic to source websites—ultimately weakening the web’s incentive to keep content updated. The post says results-only search forces users to scan and click through to build an evidence-based picture, which it claims is slower but more accurate, while acknowledging AI summaries may be helpful for simple, low-ambiguity factual lookups.

More than 6 out of 10 people turn to AI for psychological support (axa.com) AI

An AXA and Ipsos Mind Health1 report finds mental health is worsening, with 46% of surveyed people saying they are struggling or languishing and two-thirds reporting negative effects from screen time. The study also reports that 61% already use AI for mental health questions and 42% of those users “almost always” follow its advice, though many express mixed views about quality and some report harmful or uncomfortable experiences. AXA says AI can help with early, supervised support but should not replace professional therapy and calls for better access to first consultations.

AI enthusiasts are in race against time, AI skeptics are in race against entropy (charitydotwtf.substack.com) AI

The piece argues that AI “enthusiasts” and “skeptics” face different existential risks—enthusiasts fear lost competitive time, while skeptics warn that shipping faster than humans can understand erodes trust and reliability—and says teams need to rebuild shared context by telling the full story of AI wins and costs, then treating the policy and safety questions as an engineering problem rather than a debate.

DeepSeek-V4-Flash (official FP8) running across 2x DGX Spark (forums.developer.nvidia.com) AI

A NVIDIA developer forum post shares a working end-to-end setup for running DeepSeek-V4-Flash in official FP8 on a dual-node 2x DGX Spark cluster (tensor parallel TP=2) with MTP and a 200K context window, including specific build/run steps, key recipe flags, and reported throughput/TTFT numbers (e.g., ~44 tok/s decode warm, ~2s TTFT on short prompts, ~6-minute cold start). The thread also documents issues encountered—especially cross-node NCCL/version “pinning,” long-context prefill being slow, a SparkRun benchmark hang caused by tokenizer resolution in the benchmark harness—and follow-up results suggesting further performance tuning, including a later test with 256K context and updated benchmark figures.

LLMs are not the black box you were promised (jay.ai) AI

The article argues that large language models are increasingly understandable rather than true “black boxes,” highlighting mechanistic interpretability work (notably Anthropic’s “circuit tracing”) that decomposes model computation into human-interpretable, causally linked features to observe multi-step reasoning.

AI Outperforms Law Professors in Stanford Law Study (law.stanford.edu) AI

A Stanford Law School-led blind study found that law professors overwhelmingly preferred AI-generated answers to student contract-law questions over answers written by other law professors, with AI winning 75% of head-to-head comparisons and being flagged as potentially pedagogically harmful only 3.5% of the time versus 12% for peer-written responses. The researchers tested nearly 3,000 anonymized comparisons using multiple AI systems and reported that professors rated the AI’s performance as comparable to the best human instructor in the study.

California’s university system went all in on AI, now it's tearing itself apart (nytimes.com) AI

The article argues that California’s higher-education system’s rapid push into AI has led to internal breakdowns, suggesting the strategy is now causing significant strain.

Paseo – Beautiful open-source coding agent interface (desktop, mobile, CLI) (github.com) AI

Paseo is an open-source interface for running and orchestrating coding agents (including Claude Code, Codex, Copilot, OpenCode, and Pi) across desktop, mobile, web, and CLI using a self-hosted local “daemon” with no telemetry or forced log-ins. The project supports voice control and cross-device workflows, and includes a “skills” system to let agents hand off, loop against acceptance criteria, or act as committees/advisors.

Self-extensible software is here and its amazing (slopfluencer.com) AI

The article argues that prepackaged software is not “dead” and that AI-enabled coding agents make a new model—self-extensible software—possible by embedding a coding agent, an extension runtime, an extension API, and agent-readable documentation so software can generate and integrate extensions for users.

The AI Trap We're Walking Into (unvoid.substack.com) AI

The article argues that while LLM capability is becoming cheaper and more “commoditized,” the cost of running effective agentic systems is rising sharply due to repeated context and token-heavy reasoning, pushing advantage toward capital-rich enterprises. It also claims that improvements increasingly depend on extracted human labor—such as data labeling, corrections, and scraped open-source contributions—which the piece says concentrates value with AI providers while leaving non-buyers to do lower-paid handwork. The author concludes that policy and technical shifts like open weights, local inference, and paying people for “data dignity” are needed to avoid repeating past cycles of power concentration.

RSS is back. AI agents are reading it (julienreszka.com) AI

The article argues that RSS “never really died” and is well-suited for AI agents because it offers deterministic, structured, and publicly accessible updates that are easier to parse and less brittle than scraping or relying on social-platform APIs.

Show HN: MetaBrain – A local document memory for AI agents (metabrain.eu) AI

MetaBrain is an open-source “local document memory” for AI agents, offering a durable, searchable store for notes, task context, metadata, tags, links, and version history via an `mb` CLI, optional `mbd` daemon, and an embeddable `MetaBrainCore` library. It supports searching by content and filters, applying unified diffs to documents while retaining prior versions, and uses a default local LevelDB-backed store at `.metabrain/store.leveldb`.

Microsoft CEO: We’re moving from OS and apps to agents instead (9to5mac.com) AI

At Microsoft Build, CEO Satya Nadella said the company is shifting from building operating systems and app platforms to an “agent-first” computing future, framing this as a platform change. The announcement was tied to Project Solara, a chip-to-cloud initiative with Qualcomm meant to power AI agents that are more personal and “always with you.”

Uber caps employee AI spending after blowing through budget in four months (techcrunch.com) AI

Uber has introduced internal caps on employee AI spending, limiting usage to $1,500 per employee per month and per agentic coding tool (such as Claude Code or Cursor), with some exceptions possible by permission, after the company said it exhausted its annual AI budget within four months.

Microsoft announces Scout, an autonomous AI agent built on OpenClaw (computerworld.com) AI

Microsoft has unveiled Scout, an always-on “autopilot” AI agent for Microsoft 365 built on the OpenClaw framework, which can independently handle tasks across apps like Teams, Outlook, OneDrive and SharePoint using a governed Entra identity. The experimental release is available through Microsoft’s Frontier program and can help with scheduling, coordinating, and flagging stalled decisions, while Microsoft says it includes enterprise-grade security and will contribute upstream to OpenClaw.

What happens when companies replace managers with AI? (analysis.infocentral.net) AI

The article argues that while LLMs may have some “management” knowledge from general web and book sources, there is little public evidence that frontier models were deliberately trained and benchmarked for the relational, contextual, and accountable functions managers perform. It points to “flat org” experiments at companies like Google, Zappos, Valve, and GitHub as recurring cases where removing formal management roles led to coordination problems, invisible hierarchies, and eventual reinstatement or redesign of managerial functions. The author concludes that using AI to replace middle management could similarly concentrate decision power at the top, with managers’ protective and translation roles being especially hard for today’s systems to replicate.