🐧 Codex in ChatGPT desktop app for Linux is now in preview
The workaround suggests that the app stack already supports native Wayland IME correctly, but the necessary configuration may not be enabled by default—a potentially straightforward packaging or launcher fix.
A Fedora 44 KDE Plasma user reports that Japanese input with Fcitx 5 does not work in the preview ChatGPT/Codex Linux desktop app when launched normally under Wayland. The issue can be fixed without XWayland by launching the app with Chromium flags that explicitly enable the Wayland platform and Wayland IME support, then adding those flags to the desktop launcher.
🔗 Read more 🔗
🧩 DeepSeek Harness developer preview
This framing shifts attention from raw model capability to the surrounding execution layer, highlighting the growing importance of tool interfaces, orchestration, state, and environment integration in practical agent systems.
DeepSeek describes an agent as the combination of a model and a harness. While the model provides the core intelligence, the harness enables the agent to understand its environment, use tools, and keep operating in real-world settings.
🔗 Read more 🔗
📈 Gloomberb
The product appears positioned as an all-in-one terminal-style alternative, combining traditional market data, portfolio workflows, and AI features rather than treating AI as a standalone research tool.
Gloomberb presents itself as a financial research and market-monitoring workspace. It combines company fundamentals, filings, ownership data, options and analyst information with breaking news, macro and market feeds, portfolios, watchlists, broker connections, alerts, AI screening, prediction markets, and chat.
🔗 Read more 🔗
⚡ Gemini 3.7 Flash
The key theme is broader workflow competence, not just stronger coding scores, suggesting Google is positioning Flash as a low-latency production workhorse spanning code, documents, interfaces, and business processes.
Google says Gemini 3.7 Flash substantially improves on 3.6 Flash across coding, web development, UI generation, document reasoning, and business automation. Reported benchmark gains include FrontierCode 1.1 Main at 43.6% versus 34.4%, DeepSWE v1.1 at 65.3% versus 49.0%, GDP.pdf at 34.0% versus 22.0%, and AutomationBench at 30.4% versus 17.0%.
🔗 Read more 🔗
🧠 Spaghettifying DRAM
The work targets assumptions beneath conventional software privilege boundaries: if physical memory translation itself can be altered, protections enforced by kernels, firmware, or isolated regions may no longer provide the expected separation.
The skitter-creek-bath-salts project explores manipulating DRAM controller address translations so physical addresses can be redirected to unexpected memory locations. By scrambling the platform’s memory mapping, the technique is described as exposing protected DRAM regions such as firmware and system-management carveouts that are normally hidden even from the operating-system kernel.
🔗 Read more 🔗
🐘 Does anyone run Postgres without PgBouncer?
The provider survey turns an old best practice into an architectural argument: when nearly an entire ecosystem independently adds the same missing layer, that may be evidence the capability belongs closer to the database itself.
Brandur Leach revisits the longstanding need for connection pooling in Postgres and finds that nearly every prominent managed Postgres provider offers PgBouncer or an equivalent managed pooling layer. He argues that this widespread workaround exposes a core operational weakness in Postgres and suggests connection management should ideally be transparent rather than requiring separate proxies, configuration, and caveats.
🔗 Read more 🔗
📝 Let Them Write RFCs
Using RFCs as an intermediate artifact can force both human and model to expose assumptions and compare competing designs before implementation, turning documentation into part of the reasoning process rather than merely a record of it.
The author describes using hundreds of millions of LLM tokens to produce multiple versions of an RFC while working on a project beyond their existing expertise. They highlight a recurring problem in AI-assisted development: not fully understanding the problem, failing to recognize knowledge gaps, and being vulnerable to plausible-sounding but weak ideas.
🔗 Read more 🔗
🚩 C3 0.8.3 Feature flags
A dedicated attribute for feature-gated declarations should make conditional compilation more explicit and easier to recognize, rather than overloading a general conditional construct for two different purposes.
C3 version 0.8.3 introduces the @feat attribute as the new mechanism for top-level conditional compilation. It replaces the previous use of top-level @if with feature checks on non-generic declarations, while top-level @if is being phased out except in generic contexts.
🔗 Read more 🔗
🐍 python’s pre-declared constants are kinda weird
These inconsistencies offer a useful window into Python’s historical evolution: concepts that appear uniform to programmers can be represented very differently across the grammar, compiler, and runtime because they were introduced at different times for different reasons.
The article examines Python’s six pre-declared constants and highlights their surprisingly different language semantics. True, False, and None are keywords, while debug is an identifier with special restrictions on assignment and deletion, resulting in behavior and error handling unlike ordinary names.
🔗 Read more 🔗