Linux 7.3 Released
The Linux 7.3 kernel was released on 18 October 2026. The final tree contains 18,765 non‑merge commits contributed by 3,021 developers, including 807 first‑time kernel contributors. Key technical additions in 7.3 are:
- Extensions to binfmt_misc for additional binary formats.
- Group‑scheduling improvements targeting performance on large systems.
- Introduction of the failfs pseudo‑filesystem, which deliberately fails all operations.
- New fchroot() system call for changing the root directory of a process.
- Global per‑CPU data structures for BPF programs.
- Completion of the sched_ext subscheduler feature.
- Working‑set tracking enhancements for virtual machines.
- Alternative data‑stream support added to the ntfs3 driver.
- Numerous SMB server improvements.
The merge window for the upcoming 7.4 release opens the following day. (Author: Jonathan Corbet).
Rabbits were a medieval royal luxury in England until the Black Death
Comments focus on rabbit meat’s historical rarity, regional availability, and culinary potential, mixing curiosity about naming and recipes with factual references to medieval statutes, price comparisons, and disease impacts. Several contributors express interest in trying rabbit dishes, while others note personal aversion or view the idea of rabbits as luxury food as contradictory to their prolific breeding. Skepticism appears regarding claims of rarity, and humor surfaces in references to “Bunny Burgers” and British medieval illustrations. Overall, the discussion balances historical inquiry, culinary curiosity, and critical questioning of the presented narrative.
Hybrid Reasoning Agents: LLM agents that think in latent space, act through text
The document “Hybrid Reasoning Agents” consists primarily of visual elements identified by their alternative‑text labels. The images include: an “archive” icon, a “license” icon, references to “BibSonomy” and “Reddit,” and logos for the Simons Foundation, Simons Foundation International, and Schmidt Sciences. These graphics suggest the paper’s association with open‑access archiving, licensing information, bibliographic/social‑network resources, and institutional sponsorship or affiliation from the Simons Foundation entities and Schmidt Sciences. No additional textual content, abstract, methodology, results, or conclusions are provided within the excerpt.
The discussion centers on hybrid‑agent architectures that offload reasoning to latent‑space passes, with participants noting potential speed and quality gains while questioning safety, interpretability, and alignment implications. Technical feedback highlights implementation problems such as looping failures and quantization issues on specific models, prompting calls for clearer benchmarks and evaluation of accuracy versus token efficiency. Opinions diverge between enthusiasm for the approach’s novelty and caution about “thinking” terminology, auditability, and the broader impact on model control and transparency.
William Gibson and the trademarking of cyberspace
Autodesk’s research lab applied for a “cyberspace” trademark in the late 1980s while developing an early virtual‑reality system (386 PC, dual Matrox cards, VPL EyePhone, Polhemus tracker, data glove, 6‑10 fps, a few hundred polygons). Timothy Leary, a friend of the author and fan of William Gibson’s Neuromancer, nicknamed the lab “Cyberia” and later circulated a rumor that Gibson was retaliating by trying to trademark “Eric Gullichsen”. Gibson’s lawyers did send Autodesk a letter objecting to the “cyberspace” filing, but Autodesk ignored it. The “Eric Gullichsen” claim was a prank by Leary, later amplified by others. At Art Futura 1990 in Barcelona, Gibson met the author and remarked that “Autodesk is not the street,” confirming his objection but indicating no further trademark action. Thus, Autodesk did seek the “cyberspace” trademark; Gibson objected; the alleged Gibson retaliation trademark is a myth.
The comments convey a generally critical view of corporate actions in tech and culture, noting that attempts to trademark genre terms like “cyberpunk” are seen as inappropriate and that companies such as Autodesk and Meta receive blame for lacking self‑awareness and failing to acknowledge the creators they borrow from. There is skepticism about the practicality of VR for precise CAD work, and a nostalgic observation that the term “cyberspace” emerged alongside countercultural movements, suggesting a disconnect between original ideas and later commercial exploitation.
How to Train Your Robot
The recent surge in AI performance is attributed to the “scaling hypothesis”: larger models, larger datasets, and more compute yield higher intelligence. While language models have access to hundreds of trillions of text tokens, robotics lacks comparable public data. Training robots requires massive examples of correct actions, obtained through several methods:
- Teleoperation (behavioral cloning) – humans control robots, recording state and visual data; high fidelity but expensive and limited in dexterity.
- Reinforcement learning on real robots – autonomous robots collect success/failure scores; promising “data flywheel” but currently sparse and repetitive.
- Simulation – cheap, fast generation of billions of steps; effective for locomotion, limited for manipulation due to the sim‑to‑real gap. Techniques like domain randomization and Automatic Domain Randomization expand variability.
- Synthetic data – converting teleoperation recordings into simulated variants.
- Human‑collected video (egocentric, UMI, tactile gloves) – cheap, diverse; requires mapping to robot actions and suffers from imperfect robot‑human correspondence.
Existing robotic datasets range from 1 M to >10 M task trajectories (≈4 k–10 k h) to tens of thousands of hours of UMI video. Compared to LLMs (GPT‑4 ≈ 685 M h equivalent), robotics faces a “100 k‑year data gap.” Rough cost estimates for closing this gap span $10–100 B. Researchers anticipate that scaling, improved sample efficiency, or leveraging generalist AI may eventually bridge the gap, but the exact data requirements remain uncertain.
The comments express a positive reaction to the NVIDIA presentation, highlighting its entertainment value and usefulness. Appreciation is shown for the linked video and the accompanying article, with gratitude conveyed for the shared resources. Additionally, there is a lament that topics related to robotics receive limited attention on the platform, indicating a desire for greater visibility of such subjects. Overall, the tone is appreciative yet slightly disappointed about the community’s focus.
There's never been a better time to go into math (by: 3blue1brown) [video]
The footer of the YouTube site presents a compact navigation menu that lists corporate and user‑oriented sections: About, Press, Copyright, Contact us, Creators, Advertise, Developers, Terms, Privacy, Policy & Safety, How YouTube works, Test new features, and NFL Sunday Ticket. These links direct users to informational pages about the platform, media relations, intellectual‑property notices, support contacts, resources for video creators, advertising programs, developer tools, legal agreements, privacy guidelines, safety policies, operational explanations, beta testing of features, and the NFL streaming service. The footer concludes with a copyright notice indicating ownership by Google LLC as of 2026.
Comments express a mixed view of AI’s impact on mathematics. Many note that automated theorem‑proving may replace routine problem‑solving, shifting mathematicians toward interpreting, verifying, and managing AI‑generated results, similar to past mechanization in other fields. Optimists highlight new opportunities in oversight, education, and interdisciplinary work, while skeptics stress loss of personal fulfillment, diminished career prospects, and a dehumanizing effect on the discipline’s culture. Overall, the consensus acknowledges both potential benefits and significant concerns about how the profession’s identity and value will evolve.
If you don't run your own LLM proxy, how do you know whats going on?
Comments reveal a split view on using proxies and API keys versus subscription plans. Many highlight the benefits of proxies for credential safety, observability, and centralized control, citing tools such as LiteLLM, Bifrost, OpenRouter, and custom extensions. Others argue that proxies add complexity and potential failure points, questioning their cost‑effectiveness and suggesting that subscription tiers often feel unlimited for typical home use. Cost concerns surface, especially regarding high‑spend API usage, while a minority dismiss the practice entirely as unnecessary or hobby‑level. Overall, the discussion balances perceived technical advantages against practicality and expense.
Tmux and Claude Code: Sessions, Send-Keys, and the Ceiling
- Claude Code can automatically create a Git worktree and a tmux session with claude --worktree --tmux (or --tmux=classic for a real tmux server).
- Recommended tmux organization: one session per repository, named after the repo; each agent runs in its own window, with optional tiled or main‑vertical pane layouts for simultaneous viewing.
- Key tmux configuration for agents includes large history-limit, 24‑bit colour (default-terminal "tmux-256color"), mouse support, disabled automatic renaming, 1‑based indexing, activity monitoring, short escape-time, and custom pane‑switch bindings.
- Automation: tmux send-keys -t <session>:<window> -l "<text>" (followed by a short sleep and an Enter) to feed prompts; tmux capture-pane -p -J -t <session>:<window> to retrieve pane output, with -J preserving wrapped lines.
- Continuous logging via tmux pipe-pane -o -t <session>:<window> 'cat >> ~/agent-logs/<name>.log'.
- Status monitoring can be built with a status-right script that scrapes recent pane content, or with native monitor-silence/visual-silence hooks.
- tmux persists across disconnects and SSH drops; sessions survive detach/attach but not reboots, and scrollback is lost when the server stops.
- crystl (macOS terminal) extends tmux by surfacing approval prompts, providing structured agent state, searchable transcript storage, and persistent project‑based sessions, addressing tmux’s limitations for multi‑agent workflows.
The comments express mixed but generally positive sentiment toward integrating LLM agents with terminal multiplexers. Users note that tmux and Zellij can be effective for handing off sessions to Claude, especially when combined with security measures like fingerprint‑protected SSH keys. Several participants highlight that the newer Herdr tool offers a more streamlined, out‑of‑the‑box experience for agentic coding, reducing the need for multiple tmux sessions, while still acknowledging tmux’s flexibility and sophisticated configurations. A recommendation for the agent‑deck project is also mentioned.
Someone behind the software
The author argues that software products reflect the perspectives and intuition of their creators rather than a purely process‑driven, data‑centric pipeline. They contrast two teams with similar resources, noting that differences arise from founders’ backgrounds, company culture, and how teams interpret customer feedback. AtLinear emphasizes intuition built from experience, mistakes, and direct conversations with users, using this judgment to decide which features to pursue or discard. The piece warns that contemporary products often become “messy” because adding capabilities is easier than reevaluating core design, leading to fragmented, additive feature sets that lack a unified vision. The author attributes this trend partly to organizational structures where individuals work with autonomous agents, producing isolated changes that cumulatively form incoherent systems. They advocate for deeper problem understanding and sustained, cohesive design to preserve a product’s “soul” and ensure a discernible creator behind the software.
The comments converge on the view that effective software stems from a clear, human‑driven vision rather than a factory‑like, feature‑additive process. Many note that developers who are also the product’s users tend to make better judgment calls, while organizational fragmentation and AI‑generated code often lead to incoherent, “Frankenstein”‑style products. There is broad agreement that intuition, taste, and craftsmanship remain essential, especially for complex or high‑value experiences, and that reliance on data‑only or purely automated approaches risks losing product coherence and lasting value.
Superlite: A Faster Alternatie to Superpowers
Superlite is a Claude Code plugin that automates end‑to‑end software development from a single specification. It replaces the highly interactive “superpowers” workflow with a single approval, then generates a parallel task graph, assigns each task to the cheapest capable sub‑agent, runs tasks concurrently, and performs automated seam checks and reviews. Core components include: Opus (orchestrator, never edits code), Sonnet (primary builder), Haiku (small mechanical tasks), dedicated review agents, and tiered fix‑only agents. Hooks enforce rules (no coordinator edits, background test execution, lint stability) and record token usage, cost, and time for budgeting. Benchmarks on a static‑site‑generator + spreadsheet engine project (177 hidden tests) show Superlite completing builds in 16.1 minutes—about half the time of the prior Superpowers setup—by reducing human gates, eliminating strict sequential dispatch, and moving test runs to background agents. Installation requires Claude Code, git, Python 3.10+, and Bash; usage involves providing a spec, optionally choosing a human or AI orchestrator, and monitoring progress via generated logs (progress.md, test-summary.json, metrics.jsonl). The plugin is open‑source on GitHub (iha2/superlite).
Comments convey growing skepticism toward multi‑agent fleets and “superpowers” skills, noting little speed benefit and added refactoring overhead, especially for legacy code. Users report that newer models (e.g., Gemini 3.8 Flash, Claude, Opus 5.5, Sonnet 5.5) already embed robust processes, making external skills redundant and token‑inefficient. Early enthusiasm for brainstorming and planning tools has faded as prompts become smoother, while concerns persist about approval prompts, token costs, and the need to continually supervise agents. Overall sentiment favors simpler, single‑agent workflows and questions the long‑term value of elaborate skill frameworks.
The comments convey overall enthusiasm for recent Linux kernel developments, highlighting rapid Apple‑silicon support, the record‑breaking number of commits and new contributors, and the addition of an in‑kernel SMB server with optional Time Machine compatibility. Observations note the prevalence of AI‑assisted patches without criticism, while curiosity is expressed about niche features such as failfs. The tone is generally positive about the pace of progress, with neutral inquiries into new functionalities and their practical use cases.