Hasty Briefsbeta

All tags

ai development

20 stories total

Bilingual

Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App
2 months ago
- Boris Cherny of Anthropic discussed using Claude Code to rewrite the Electron-based Claude desktop app in Swift, running continuously over two weeks.
- Cherny emphasizes that the key skill is not prompt engineering but enabling Claude to handle hard tasks and verify its own work along the way.
- John Gruber criticizes the experiment, arguing it only addresses technical issues (Electron bloat) but not the underlying poor design of the Claude app.
- Gruber notes that the current Claude Mac app is still slow and unreliable, contradicting Cherny's claim that it is now 'pretty awesome'.
Code review (as we know it) needs to die
2 months ago
- AI-driven development is breaking traditional code review: faster code generation caps velocity at slow human reviews.
- Asynch diff-centered code review, ~20 years old, fails in an era where writing code is cheap but reviewing isn't.
- Diffs lack context, requiring reviewers to reconstruct intent and design, adding expensive overhead.
- All changes get equal scrutiny regardless of risk, wasting attention on low-risk changes and under-scrutinizing risky ones.
- Code review's core jobs—defect detection, maintainability, knowledge sharing, control, and norm enforcement—remain vital but must be unbundled.
- Defect detection shifts to automated checks, independent AI review, and human verification of product, not just code.
- Knowledge sharing moves up: review focuses on intent and design, while code comprehension becomes pull-based via AI.
- Norm enforcement moves into specs and automated tools, not comments on diffs.
- Control stays as merge gates: humans approve changes based on higher-level review, trusting agents and automated verification.
- New process: align design before code, surface changes with context, run adversarial AI review, show evidence, route by risk, and absorb mistakes through system safeguards.
- Examples like Linear Diffs, Meta RADAR, and Uber's Code Inbox show components of this new approach in practice.
- Costs include less vetted individual changes and shipping some bugs, but winners accept this and build resilient systems.
- Ultimately, redesigning review around AI accelerates velocity and lets engineers focus on creative, high-leverage work.
A non-technical CFO is shipping better code than the agencies he hired
2 months ago
- A non-technical CFO built a functional internal dashboard using Claude Code, succeeding where agencies and low-code tools failed.
- The combination of business domain knowledge and LLMs is powerful for building internal systems, especially for non-technical stakeholders.
- Less senior engineers with limited domain knowledge are at risk of automation; domain knowledge is becoming more important than ever.
- Senior developers remain essential for productionizing apps (SDLC, security, testing, scalability), but will likely focus on taking MVPs built by business users and making them production-ready.
- Organizations should prepare for a surge in internal apps (Jevons Paradox) and need a PaaS to manage authentication, security, and maintenance.
- Recommendation: Senior technical leadership should plan now for managing 100x more internal apps.
No-Look Coding and the Five Stages of Grief
2 months ago
- Uncle Bob Martin advocates not reading code written by AI agents, relying on constraints instead.
- The author, after initial skepticism, now fully embraces 'no-look' coding, having built five products without seeing code.
- The real debate should be about customer obsession, product judgment, and leadership, not about reading code.
- Code review was a proxy for truth; now AI can generate and review code, making specs the truth.
- The author's workflow: write spec, iterate with AI, generate code and tests, code review by AI, human testing, delete and rewrite if needed.
- Focus on spec, gates (automated tests), behavior (user testing), and rarely code.
- The shift is driven by fear of loss of identity; developers go through grief stages.
- The author finds joy in mastering new skills and solving customer pain.
The Vibe Coding Trap: Why AI-Assisted Projects Still Fail
2 months 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.
AI hasn't shifted the bottleneck from coding to code review
2 months 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.
OpenAI is everything it promised not to be: closed-Source and for-profit (2023)
2 months 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.
I couldn't pay for DeepSeek V4 with my card. So I built a proxy
2 months 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.
Code Was Our Medium for Thought
2 months 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.
Musk promises purge after Grok Build caught sending repos to the cloud
2 months 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.
Microsoft Agent Framework for Go
2 months 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.
Cdbx.ai – AI-powered browser IDE to describe, build, and publish apps
2 months 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.
Scarf has moved away from Haskell
2 months 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.
'I'm not a programmer' anymore: Linus Torvalds on the only two tools he uses now
3 months 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.
Kernels: Major Updates
3 months 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.
Artificial Adventures
3 months 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.
Mark Zuckerberg tells staff that AI agents haven't progressed enough
3 months 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.
$85,000 in tokens later: What I learned from scaling agentic coding at Lovable
3 months 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.
Claude's Criminally Bad Electron Mac App Is an Inside Job
3 months 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.
Deep Agent Code Capabilities
3 months 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.

firstprev0next

ai development

Boris Cherny on Trying to Get Claude Code to Rewrite the Claude App

Code review (as we know it) needs to die

A non-technical CFO is shipping better code than the agencies he hired

No-Look Coding and the Five Stages of Grief

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

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

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

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

Code Was Our Medium for Thought

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

Microsoft Agent Framework for Go

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

Scarf has moved away from Haskell

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

Kernels: Major Updates

Artificial Adventures

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

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

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

Deep Agent Code Capabilities