Researchers propose a metric called NLS depth to quantify how much unobservable reasoning AI models can perform internally, separate from chain-of-thought outputs. This matters because architectural changes could let models do complex reasoning without showing their work, making them harder to monitor.

• Current chain-of-thought monitoring works because models are trained on human text, but new architectures could bypass this by performing reasoning in opaque hidden states instead of readable tokens.
• NLS depth measures the longest computational path that doesn't pass through natural-language outputs (tokens initialized from pre-training), making it possible to assess monitorability risk before a model is even trained.
• The metric deliberately ignores some post-training choices (like RLHF or distillation) to focus on architectural differences, distinguishing between standard transformers and genuinely concerning designs like full-bandwidth transformers that can hide reasoning across entire context windows.

The author argues that you don't need massive AI models for most real-world applications—smaller models handle instruction-following well and fail predictably when they don't know something, rather than confidently making things up like larger models do. He's building a system that lets you swap between different AI providers through a single interface.

• Small models hallucinate less reliably than large ones; when told clearly that information is missing, they admit it instead of inventing plausible-sounding answers
• Larger models produce more polished fabrications, making their hallucinations harder to catch in production
• The author's architecture (Pepper) uses a provider-neutral layer that treats adding new AI backends as simple routing and translation work, not a full rewrite

This article breaks down how to design a web system for scale, starting with scope, use cases, and constraints. It then walks through high-level and component design, scaling strategies (caching, load balancing, replication), performance tuning, reliability, and front-end considerations.

• Start every system design by pinning down use cases and traffic patterns (RPS, read/write ratio) before drawing architecture
• When the database becomes the bottleneck, add Memcached/Redis caching before jumping to vertical or horizontal scaling
• Choose replication vs. partitioning based on whether you need more read capacity or more write throughput
• Isolate shared concerns (auth, caching, DB access) in a platform layer so web, mobile, and API servers can scale independently

Three major AI agent platforms—Manus, OpenClaw, and Claude Code—store their memory in plain Markdown files instead of vector databases. The article breaks down how file-based context boosts token-cache economics, enables attention control, and layers optional semantic retrieval, plus when this approach starts to break down.

• Manus hit $100M ARR in eight months using append-only Markdown for memory instead of a vector database, since its 100:1 input-to-output token ratio makes cache-friendly file appends roughly 10x cheaper than invalidating a cache with database writes.
• OpenClaw layers SQLite-based vector search directly on top of its Markdown notes (vectorWeight = 0.7 plus temporal decay) rather than using an external vector database, blending keyword and embedding scores.
• Claude Code's CLAUDE.md files use directory/scope-based progressive disclosure to load only relevant context, replacing the need for a retrieval database entirely.
• All three systems treat the filesystem hierarchy itself as the retrieval mechanism, using plain text files for both long-term memory and active attention control (e.g. Manus's constantly rewritten todo.md).

A Lobste.rs thread where developers share untested ideas—from building GUI toolkits with an accessibility-first architecture to enable better testing and scripting, to using 32-bit pointers on 64-bit systems for memory savings, and applying property-based approaches to infrastructure. Contributors discuss tools like AccessKit, the x32 ABI, and the benefits of driving design by fundamental properties rather than legacy constraints.

• Building GUI toolkits with accessibility as the core data model (not bolted on) enables pixel-free testing, automation via high-level actions, and multiple "themes" (graphical, voice, terminal, remote) on one semantic backbone—AccessKit already does much of this for Rust/Egui.
• Reviving x32 ABI (32-bit pointers on 64-bit hardware) cut Mastodon's RAM use from 650MB to 350MB by shrinking pointers and improving cache density, though Debian disables it by default and most distros don't ship it.
• V8 and other JIT engines already use pointer compression for sandboxing, suggesting broader adoption of narrow pointers could yield wider resource savings.

Developers unveiled plans for a glass skyscraper in Miami that will house Donald Trump's presidential library. The design includes space for a hotel, a 747, and replicas of White House features, with funding currently being sought for the project. AI-generated imagery was used to create a promotional video for the library.

• Trump shared a video of a proposed glass skyscraper in Miami for his presidential library, featuring a gold entryway, a lit-up American flag, a "TRUMP" sign, a hotel, and a replica Air Force One in the lobby
• Reuters confirmed the promotional video was made with Google's AI generative tools, identified via watermarks from an AI detection tool
• Construction won't start until after Trump leaves office, and fundraising is still underway
• The video reflects a growing trend of Republicans using AI-generated imagery in political messaging