Bilingual

Deep Agent Code Capabilities
23 days ago
- dcode is an open-source coding agent built on the Deep Agents SDK.
- It works with any large language model, allowing switching between providers mid-session.
- Features include persistent memory across conversations, customizable skills, and approval controls for code execution.
- Supports various built-in tools for file operations, shell execution, web search, and task planning.
- Offers interactive and non-interactive modes, with CLI commands for agent and session management.
- Includes capabilities like subagents, human-in-the-loop approval, and LangSmith tracing for observability.
- Configuration is stored in ~/.deepagents/ with per-agent directories for memory and skills.
- Provides slash commands in interactive sessions for model switching, skill invocation, and memory management.
- Supports remote sandboxes for code execution and MCP tools for external integrations.
- Non-interactive mode allows automated task execution with safeguards like turn and time limits.
Claude's Criminally Bad Electron Mac App Is an Inside Job
22 days ago
- Anthropic's Claude desktop app for MacOS is built on Electron, which is criticized for not providing a native Mac experience.
- The article suggests that despite Anthropic's advancements in AI coding tools like Claude Code, the team chose Electron due to engineers' familiarity, particularly Felix Rieseberg, a key Electron developer.
- The post argues that Electron limits app functionality and performance, citing examples like Notion migrating to SwiftUI for better performance, and questions why Claude hasn't followed suit.
$85,000 in tokens later: What I learned from scaling agentic coding at Lovable
21 days ago
- The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
- Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
- Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.
- AI workflows classify PRs by risk level (high, medium, low), enforcing human review for high-risk changes and using AI reviews for others to manage scaling and risk.
- Different AI review tools were tested: CI reviews serve as a safety net, while local reviews with subagents handle bulk work, though sharing setups between people remains unsolved.
- PRs remain useful as an abstraction for unverified changes, and stacked PRs help manage context and risk, as large PRs are not reviewable by AI or humans effectively.
- Managing human and agent attention involves using fast mode for interactive tasks, ensuring task autonomy to minimize permission requests, and batching validation work per PR stack.
- Agents require external task trackers like Beads, subagents for large tasks, small PRs to avoid context limits, and new context per task for basic hygiene.
- Task tracking for humans and agents should be separate to avoid clutter from agents' ephemeral text, keeping systems like Linear clean.
- Improving the development process involves agents writing skills, reflecting on past work, sharing context via git, and encouraging cultural sharing, though sharing skills is still challenging.
- The author merged 293 PRs in a week with no production defects, viewing the productivity gains as beyond unassisted development, despite token costs, with focus on outcomes over efficiency.
- The journey is described as fun, with anticipation for future innovations, and Lovable is hiring for those interested in this approach.
Mark Zuckerberg tells staff that AI agents haven't progressed enough
20 days ago
- Meta's AI agent development pace has not accelerated as expected, despite workforce reassignments to AI groups.
- CEO Mark Zuckerberg acknowledged that recent job cuts, involving layoffs and reassignments, were not executed cleanly.
- Zuckerberg expressed concerns about adapting quickly to tech industry changes, driving the restructuring.
- The anticipated benefits of Meta's AI-focused structure have not yet materialized, though improvements are expected within 3-6 months.
- Meta's AI unit has been described negatively by some engineers as a demanding environment.
- Meta plans significant investment in AI infrastructure, with spending potentially reaching $145 billion this year.
Artificial Adventures
24 days ago
- AI assistants are most valuable for reviewing code and finding bugs, with frontier models performing well.
- Refactoring tasks benefit from AI, reducing the cost of fixing design mistakes, but review can be challenging due to mixed-in incorrect changes.
- Writing code together with AI is problematic because AI makes poor decisions and often disregards instructions, though future harnesses may improve collaboration.
- AI is effective for one-off scripts and small tasks where output can be easily verified, but fails for complex projects like implementing board game rules.
- Search and verification tasks work well when precision is needed, but caution is required for unverifiable answers.
- Brainstorming with AI yields banal suggestions, and existing harnesses are limited, highlighting the need for better control and interaction methods.
- AI subscriptions offer good value, but token-based pricing may change cost-effectiveness, with cheaper models often misaligned or less useful.
- AI use requires adapting practices to manage mental models and verification, with potential for future improvements in model capabilities and tools.
Kernels: Major Updates
19 days ago
- Introduction of a new 'kernel' repository type on the Hugging Face Hub to cater to compute-specific needs, allowing users to browse accelerators, operating systems, and backend versions.
- Enhanced security measures including trusted kernel publishers, code signing with Sigstore's cosign for ephemeral keys, and reproducibility via Nix builds to prevent malicious kernel execution.
- Revamped CLI tools with better separation between kernels and kernel-builder, focusing on library functions versus building processes, and improved support for agentic kernel development workflows.
- Extended framework and backend coverage, such as Torch Stable ABI support and Apache TVM FFI integration, enabling cross-framework kernel interoperability.
- Foundation for agentic kernel development with tools for scaffolding, building, benchmarking, and optimizing kernels via HF Jobs integration for performance testing across hardware configurations.
- Miscellaneous updates including an easier environment setup script, system cards for kernels, compatibility checks via has_kernel() and get_kernel_variants(), and improved manylinux_2_28 support to resolve libstdc++ linking issues.
'I'm not a programmer' anymore: Linus Torvalds on the only two tools he uses now
16 days ago
- Linus Torvalds now sees himself as a development lead rather than a programmer, focusing on understanding the bigger picture through pull requests and handling merge conflicts.
- He uses only two main tools: Git for version control and email for communication, relying on maintainers and avoiding last-minute fixes to maintain a steady release schedule.
- Torvalds emphasizes incremental improvements over splashy releases, though AI-generated bug reports and patches have added stress, often producing junk or superficial fixes.
- He supports removing obsolete hardware and software, like discontinuing support for 486 SX CPUs and old networking standards, to reduce maintenance burdens.
- While acknowledging Rust's value, he believes C remains simpler and more powerful, noting that Rust doesn't prevent logic errors and its guarantees break in mixed C/Rust codebases.
- AI and LLMs are used cautiously; they can prototype ideas but often generate invalid bug reports or band-aid patches, requiring human verification and suggested fixes.
- Torvalds values finding bugs, even embarrassing ones, over ignoring them, and uses AI for personal projects like adding Godzilla to family photos from his travels.
Scarf has moved away from Haskell
15 days ago
- Scarf moved from Haskell to Python due to AI changing development economics, making compile times and ecosystem friction significant bottlenecks.
- AI allows faster code generation, but Haskell's long compilation cycles hinder rapid feedback, especially in parallel AI-assisted workflows.
- The shift enabled faster development cycles, improved test coverage, and quicker hotfixes, with no noticeable loss from reduced type safety.
- Haskell faces stagnation risk if it doesn't adapt to AI by optimizing for fast builds, documentation, and agent-friendly tooling.
Cdbx.ai – AI-powered browser IDE to describe, build, and publish apps
12 days ago
- cdbx is a browser-based IDE that requires no setup, featuring a full Monaco editor, real terminal, and AI pair programmer.
- Users can describe ideas in plain English to generate apps in real-time, with one-click publishing and no learning curve.
- It integrates external services like Notion, GitHub, and Slack, allowing AI to query them on demand without copy-pasting.
- Agents are embedded throughout the workflow, including planning, coding, error-fixing, and UI refinement with live preview.
- The platform supports multiple programming languages and frameworks for both frontend and backend, with a Runner API for executing code securely.
- Pricing includes a free tier with basic features and Pro/Team plans offering more credits, AI models, and collaboration tools.
Microsoft Agent Framework for Go
12 days ago
- Microsoft Agent Framework for Go public preview announced for AI agent and multi-agent workflow development.
- Framework designed to transition from single prompt calls to production agent systems with tools, context, coordination, and observability.
- Go SDK integrates into services, CLIs, workers, and cloud-native applications where Go is suitable.
- Installation via 'go get github.com/microsoft/agent-framework-go' includes a sample agent using Microsoft Foundry.
- Example requires environment variables (FOUNDRY_PROJECT_ENDPOINT, FOUNDRY_MODEL) and Azure authentication (e.g., az login).
- Preview features: providers for Microsoft Foundry, Azure OpenAI, Anthropic, Gemini, A2A; tools, MCP, middleware; multi-agent workflows; OpenTelemetry tracing.
- Go SDK joins existing .NET and Python SDKs, with potential API changes during preview based on feedback.
Musk promises purge after Grok Build caught sending repos to the cloud
11 days ago
- OpenAI encrypts Codex agent instructions, causing developer concerns over debugging and auditing.
- Anthropic's Claude exhibits varied values across languages, such as being nicer in Hindi or Arabic.
- New York halts datacenter buildouts over 50 MW pending environmental and ratepayer protection rules.
- IBM's mainframe sales decline due to AI hardware spending, with stock losing over a quarter of its value.
- Anthropic's tokenizer complicates AI pricing as token consumption doesn't fully reflect costs.
- Russian actors pose as Signal support for phishing, while US disrupts Iranian propaganda sites.
- Microsoft's patches fail to fix on-prem SharePoint, leaving it vulnerable to zero-day attacks.
- DEF CON's Franklin project expands to include all hackers in hardening critical infrastructure.
- EQT acquires a majority stake in Acronis at a $3.5B+ valuation for the cybersecurity firm.
- Ransomware remains a persistent threat a decade after the first corporate attack, but infosec offers stable career prospects.
- Debian releases final version for x86-32 with updates 13.6 and 12.15.
- Joomla extensions iCagenda and Balbooa Forms exploited with critical vulnerabilities impacting a million sites.
- Frame, a new X11 server written in assembly, joins other X11 alternatives.
- Cinnamon 6.8 adds Wayland support optionally, while Linux Mint's desktop offers both display servers.
- KDE Plasma 6.6.6 is released, with future versions like 6.8 mandating Wayland support.
- Collabora's CODE 26.04 enhances FOSS office suites with Markdown, formula handling, and optional AI integration.
Code Was Our Medium for Thought
11 days ago
- The author created a beeswarm chart showing AI model releases over time, with each mark representing a model and its horizontal position indicating the publication date.
- They discuss the initial excitement of AI-generated code feeling like magic, but overuse leads to producing low-quality work and overwhelming PR reviews, forcing a choice between being a bottleneck or relying on agents with half-hearted efforts.
- Reflecting on the past, the author notes that coding involved problem exploration, decision-making, and team collaboration, with code evolving from fuzzy ideas to clear implementations as a medium for thought.
- They acknowledge that while reverting to old ways isn't desirable, achieving a flow state requires new workflows and tooling, similar to how IDEs evolved over decades to support coding practices.
- The history of programming is traced from punch cards to assembly, compilers, scripting languages, and libraries, each step abstracting details to allow faster and more ambitious creation.
- Critiquing current AI use, the author argues that relying on prompts and raw code reviews leads to unreadable code and agent overuse, whereas modern AI models, rooted in language and translation research, can robustly translate code across mediums.
- They propose that AI agents could create custom whiteboards or playgrounds for exploration, starting with images or prototypes instead of code, to maintain intuition, mental mapping, and team alignment.
- The conclusion emphasizes that if AI is only used to edit code faster, it risks losing the thinking process; instead, tools should enable human-centric thinking to make coding feel grounded and engaging.
I couldn't pay for DeepSeek V4 with my card. So I built a proxy
11 days ago
- DeepSeek V4-Flash API offers a cost-effective pricing of about $0.14 per million input tokens via the official platform, significantly cheaper than OpenAI or Anthropic.
- The official DeepSeek platform limits payment methods to Alipay and WeChat Pay, excluding international developers using Visa or Mastercard.
- A solution called api-hub.cc was created as a proxy that accepts credit cards and generates API keys with spending limits via LiteLLM's /key/generate endpoint.
- api-hub.cc provides a flat rate of $0.20 per million input tokens for DeepSeek V4-Flash, which is cheaper than DeepSeek's peak-hour pricing of ~$0.28 per million tokens introduced in June 2026.
- The proxy offers three pricing tiers: $5 for casual use, $20 for daily use, and $100 for team or power users, addressing the payment barrier for international developers.
OpenAI is everything it promised not to be: closed-Source and for-profit (2023)
10 days ago
- OpenAI evolved from a nonprofit research organization in 2015 with ideals of transparency and benefiting humanity to a for-profit entity in 2019, driven by competition and profit motives.
- The company's shift included forming a 'capped profit' structure, securing investments like $1 billion from Microsoft, and moving away from openly sharing research and code, as seen with GPT-2 and GPT-3.
- OpenAI's recent actions, including partnerships with companies like Microsoft and Bain, and the release of ChatGPT, highlight a focus on commercialization and rapid deployment, raising concerns about ethics and transparency.
- Critics, including co-founder Elon Musk, argue that OpenAI's current direction prioritizes speed and profit over its founding principles, leading to issues like unreliable AI behavior and lack of openness.
- The hype around AGI and chatbots like ChatGPT is contrasted with practical flaws, such as misinformation spread, inability to detect AI-generated text, and questions about the feasibility and responsible development of AGI.
AI hasn't shifted the bottleneck from coding to code review
8 days ago
- AI hasn't shifted the bottleneck from coding to code review, and coding was never the bottleneck.
- Many teams have multiple approved changes waiting in batches after code review, indicating the bottleneck is downstream (e.g., testing, deployment).
- Industry studies show AI increases coding speed and merge rates, but review times and PR sizes also rise.
- Accumulated unreleased changes in batches increase risk; speeding up code review just moves pressure to the real bottleneck.
- The real constraint may be manual steps, cumbersome approval processes, or deployment issues, not coding or reviewing.
- Research often stops at code merge, missing post-review delays, which hides the true bottleneck.
- AI initiatives succeed or fail based on addressing batch problems; fixing code review alone won't improve value delivery.
The Vibe Coding Trap: Why AI-Assisted Projects Still Fail
3 days ago
- AI-assisted development often fails not at the generation stage but during integration and deployment, as the produced code is fragile and doesn't fit production systems.
- A key debate is whether the main bottleneck is procedural (lack of rigorous code review and testing) or epistemic (tacit knowledge that can't be documented or caught by process).
- AI-generated code suppresses review instincts because it is fluent and internally coherent, unlike human mistakes that are more obvious.
- Unlike a junior engineer who learns from corrections, an AI model resets after each session, leading to repeated, costly mistakes.
- Some production constraints are tacit—known only through experience—and cannot be fully captured in specifications, limiting the effectiveness of procedural fixes.
- Economic incentives favor patching AI mistakes over documenting tacit knowledge, creating a cycle of recurring failures.
- The two diagnoses—procedural and epistemic—are intertwined in practice, making it hard to determine which failure mode is occurring until after attempts to fix it procedurally fail.

Hasty Briefsbeta

ai development

Deep Agent Code Capabilities

Claude's Criminally Bad Electron Mac App Is an Inside Job

$85,000 in tokens later: What I learned from scaling agentic coding at Lovable

Mark Zuckerberg tells staff that AI agents haven't progressed enough

Artificial Adventures

Kernels: Major Updates

'I'm not a programmer' anymore: Linus Torvalds on the only two tools he uses now

Scarf has moved away from Haskell

Cdbx.ai – AI-powered browser IDE to describe, build, and publish apps

Microsoft Agent Framework for Go

Musk promises purge after Grok Build caught sending repos to the cloud

Code Was Our Medium for Thought

I couldn't pay for DeepSeek V4 with my card. So I built a proxy

OpenAI is everything it promised not to be: closed-Source and for-profit (2023)

AI hasn't shifted the bottleneck from coding to code review

The Vibe Coding Trap: Why AI-Assisted Projects Still Fail