Hasty Briefsbeta

All tags

#software development

297 stories total

Bilingual

AI Can't Even Fix a Simple Bug – But Sure, Let's Fire All Our Engineers
a year ago
- GitHub Copilot AI submitted broken code to Microsoft’s .NET runtime, repeatedly failing to fix issues despite human corrections.
- CEOs are using AI advancements as a cover to justify layoffs, masking poor planning during the pandemic hiring boom.
- AI tools enhance productivity but cannot replace human engineers, especially in complex problem-solving and system design.
- Developers should master AI tools, document their limitations, and publicly share real-world AI successes and failures.
- The most effective approach is skilled developers using AI as a force multiplier, not as a replacement for human expertise.
Developers spend most of their time figuring the system out
a year ago
- Developers spend most of their time figuring out the system, with maintenance and comprehension being significant time consumers.
- Research from 1979 and 2018 shows that comprehension and navigation account for a large portion of development effort (~58% and 24% respectively).
- The 'figuring out' time is the largest expense in development, yet it's often not explicitly discussed or optimized.
- Reading code is the primary method for gathering information, but it's manual and inefficient, leading to incomplete information and uncertainty.
- Moldable development is proposed as a solution, where tools are customized to the context of the problem, reducing the need for manual code reading.
- Glamorous Toolkit is introduced as a moldable development environment to facilitate the creation of custom tools for software systems.
The Copilot Delusion
a year ago
- The author describes a frustrating experience with a coworker who resembles GitHub Copilot—fast but reckless, pasting code without understanding.
- Copilot is useful for syntax help and quick scaffolding but lacks nuance and can be dangerous without supervision.
- The author criticizes the over-reliance on AI tools, arguing it degrades programming skills and leads to bloated, inefficient code.
- Real programming requires deep understanding of systems, performance tuning, and hands-on problem-solving, which AI cannot replicate.
- The article warns against the normalization of mediocrity in software development, where AI-generated code replaces true craftsmanship.
The Myth of the 'Waterfall' SDLC
a year ago
- The article explores the misconceptions and historical context of the Waterfall model in software development.
- It challenges the common 'Frozen Waterfall' interpretation, arguing that no original sources advocate for such a rigid process.
- The author traces the origins of Waterfall to papers by Herbert D. Bennington (1956) and Winston W. Royce (1970), neither of whom used the term 'Waterfall'.
- Royce's model is described as iterative and flexible, contrary to the rigid, linear process often attributed to Waterfall.
- The U.S. Department of Defense's DOD Standard 2167 is often blamed for popularizing the 'Frozen Waterfall,' but the standard itself supports iterative development.
- The article suggests that the 'Frozen Waterfall' myth may have been propagated by salespeople and Agile proponents to promote their methodologies.
- Key characteristics of Royce's Waterfall include heavy documentation, iterative phases, and customer involvement throughout the process.
- The author argues that no single development methodology is universally applicable and advocates for tailoring processes to specific project needs.
- Alternative terms like 'Planned Development' are proposed to avoid the negative connotations associated with 'Waterfall.'
- The article concludes with alternative visual representations of Waterfall to better reflect its iterative and flexible nature.
On "Vibe Coding"
a year ago
- AI hype cycles shift from 'reasoning models' to 'agentic AI' and 'Vibe Coding'.
- 'Vibe Coding' involves using AI to generate code without deep understanding of software architecture.
- Narratives around 'Vibe Coding' promote democratization and empowerment but overlook complexities of software development.
- Software creation involves understanding problems, domains, and social contexts, not just writing code.
- 'Vibe Coding' risks creating untested, bug-ridden software that becomes permanent due to management pressures.
- Democratization via AI commodifies skills, centralizes power, and undermines personal growth and expertise.
- AI tools devalue expertise and discourage learning, infantilizing users by providing instant, low-quality results.
- Marginalized groups are further disadvantaged by opaque, unreliable AI-generated solutions.
- Untested AI-generated software poses risks to users and the broader digital ecosystem.
- The 'Vibe Coding' trend reflects a disrespect for craftsmanship and long-term problem-solving.
I Let Google's Jules AI agent into my code, it did hours of work in an instant
a year ago
- Google's Jules AI agent can add new features to software by following text instructions, significantly reducing development time.
- Jules is currently free and allows five requests per day, demonstrating high efficiency in coding tasks.
- The author successfully used Jules to add a 'Hide Admin Bar' feature to a WordPress plugin in under 10 minutes.
- Jules autonomously plans and implements code changes, requiring minimal human intervention but raising concerns about oversight.
- The AI's ability to understand and modify entire codebases quickly presents both opportunities and risks for software development.
- Junior developer roles may be at risk as AI agents like Jules can perform their tasks more cost-effectively.
- The author highlights the need for experienced coders to review AI-generated code to prevent errors and ensure quality.
- Jules' beta status includes occasional downtime and interface issues, expected to be resolved in future updates.
Domain Modelers Will Win the AI Era
a year ago
- AI tools are closing the 'implementation gap' by allowing domain experts to directly translate their visions into working code without needing deep coding knowledge.
- The bottleneck has shifted from coding ability to understanding what should be built, emphasizing the importance of domain knowledge over syntax.
- Low-level coding is becoming commoditized, while the ability to define entities, relationships, and constraints remains scarce and valuable.
- AI can generate code, but without a clear mental model of the domain, the output may be flawed or nonsensical.
- Domain modelers—those with deep understanding of specific fields—will thrive in the AI era, as they can leverage AI to build and iterate quickly.
- Examples include doctors modeling workflows, teachers creating custom tools, and logistics experts automating scheduling.
- AI is reversing the trend of specialization, bringing us back to a time when problem-solvers could also build solutions.
- Prompting AI is easy, but effective modeling is what differentiates a functional product from a mere demo.
- The era of needing a developer to realize an idea is fading, as AI empowers domain experts to build directly.
The Myth of Developer Obsolescence
a year ago
- New technologies often promise to make developers obsolete but end up transforming roles instead.
- NoCode created specialists and backend integrators, not fewer developers.
- Cloud computing turned sysadmins into DevOps engineers with higher salaries.
- Offshore development led to distributed teams with higher costs than expected.
- AI-assisted development highlights that architecting systems is more valuable than writing code.
- AI can generate code but struggles with system architecture and global design.
- The ability to manage and minimize technical liability becomes more critical with AI.
- Architecting systems is the skill AI cannot replace.
As a developer, my most important tools are a pen and a notebook
a year ago
- The author considers a notebook as their most important tool as a software developer.
- Writing and sketching help in transforming abstract ideas into tangible forms, aiding in problem-solving and design.
- Stepping away from the computer and using a notebook fosters creativity and better understanding of problems.
- Writing about code as if explaining it to someone else helps in identifying inconsistencies and mistakes.
- Notebooks serve as a record of thoughts and processes, useful for future reference.
- The author invites discussions on the topic via email, aiming for deeper conversations in 2025.
Vibe coding for teams, thoughts to date
a year ago
- LLMs are the most disruptive change to coding since the World Wide Web.
- Writing code is easier than understanding code, a dynamic unchanged by LLMs.
- Every line of code represents tech debt, encoding current, possibly flawed understanding.
- Updating mental models is harder with LLMs than with engineers.
- Convincing engineers to abandon outdated tests is likened to reasoning with toddlers.
- LLMs struggle to prioritize existing codebase elements over new creations.
- LLMs frequently reinvent existing functionality, leading to 'bushy' codebases.
- Productive teams rely on deep toolchain expertise, a practice challenged by LLMs.
- LLMs lower coding costs but increase ecological diversity in software environments.
- Current trends suggest a future of large, complex, bespoke codebases.
- Potential exists for LLMs to aid in code reading and reasoning, reducing complexity.
Why GUIs are built at least 2.5 times
a year ago
- The article critiques Lean Software Development, arguing it misapplies manufacturing metaphors to software development.
- Software development is likened to designing a factory rather than working in one, with the program itself being 'The Factory'.
- The process of software development is described as iterative, involving deep understanding and feedback loops to discover what users truly want.
- GUIs are highlighted as particularly challenging to design due to subjective user experiences, often requiring multiple iterations to reach an acceptable solution.
- The article emphasizes that software development's economic model differs from manufacturing, with negligible costs for copies but high initial design costs.
- The concept of 'Waste' in Lean Software Development is criticized for being subjective and potentially undermining essential creative processes.
- The author suggests that understanding software development requires recognizing its unique nature, distinct from manufacturing or other analogies.
What do software developers need to know to succeed in an age of AI?
a year ago
- Generative AI shows early signs of boosting productivity for software developers.
- Concerns exist about workforce disruption and deskilling due to AI.
- Research involved 21 developers at the forefront of AI usage.
- Identified 12 work goals, 75 associated tasks, and required skills/knowledge.
- Findings distilled into 5 key insights.
- Successful AI-enhanced developers need skills in four domains: using Generative AI, core software engineering, adjacent engineering, and non-engineering.
- Skills are applied across a 6-step task workflow.
- Future-proofing developers requires on-the-job learning and updated CS degree programs.
- Both 'soft' and technical skills are essential to reskill, upskill, and prevent deskilling.
Why I Wrote the Beam Book
a year ago
- The author wrote 'The BEAM Book' after a decade of maintaining Klarna’s core system, where a 15-millisecond pause in the BEAM could cause major disruptions.
- The project started in 2012 with optimism but faced challenges, including switching tools (DocBook to AsciiDoc) and dealing with publishing issues with O’Reilly and Pragmatic Bookshelf.
- Community feedback and open-sourcing the project on GitHub in 2017 provided motivation and improvements from contributors.
- The book covers essential BEAM topics like schedulers, garbage collection, memory management, and performance tuning for Erlang/Elixir systems.
- Lessons learned include persistence over perfection, the importance of boundaries, and the value of community involvement.
- The book is now available in paperback on Amazon, and the author encourages contributions via GitHub and offers workshops on BEAM internals.
10 Years of Betting on Rust
a year ago
- The author reflects on 10 years of using Rust, starting from its early days with version compatibility issues and challenges with the borrow checker.
- Rust's ecosystem is praised for its programming 'taste', simplicity, and robustness, making it a favorite among developers.
- Early adoption of Rust required significant effort, including creating workarounds for missing std features, but the language has since matured.
- Rust's reliability has improved, with fewer compiler bugs, better dependency management, and a more predictable crate ecosystem.
- The language is commended for its programmer empathy, offering great error messages, linting, docs, and IDE integration.
- Looking ahead, the author anticipates improvements in build simplicity, portability, const evaluation, concurrency, and broader domain adoption.
- Rust's future is seen as bright, with a positive feedback loop of adoption and development driving further innovation.
I Think I'm Done Thinking About GenAI for Now
a year ago
- The author compares their problem-solving approach to that of the fictional detective Monk, who leverages neurotic tendencies to solve problems, rather than the systematic Holmesian method.
- They express skepticism towards generative AI (genAI), citing negative aesthetic experiences, ethical concerns, and practical inefficiencies in software development.
- The author discusses the economic and environmental impacts of genAI, including high energy usage and potential industry-wide crashes.
- Concerns about genAI's educational impact highlight rampant cheating and diminished learning outcomes.
- Privacy issues arise from sending sensitive data to cloud-based AI models, with potential misuse by companies.
- The author criticizes the use of stolen training data and the legal ambiguities surrounding AI-generated content.
- They describe the fatigue from engaging with genAI discourse, noting the difficulty in proving the technology's flaws due to its complexity and rapid evolution.
- The piece concludes with the author's decision to disengage from genAI discussions, acknowledging the overwhelming challenges in opposing the technology's widespread adoption.
How do you prototype a nice language?
a year ago
- Author has been prototyping a codeCAD language for the past month, making progress in grammar, parser, and evaluation.
- Current demos are underwhelming, akin to basic examples in game programming or hardware engineering.
- The goal is software hygge: instant loading, no crashes, and a comfortable user experience.
- Bidirectional editing is an idea being explored, allowing manipulation of geometry both graphically and through source code.
- Prototyping challenges include balancing between minimal viable product and ignoring flaws.
- First-time building a language from grammar, with challenges in binding, function resolution, and evaluation.
- Found 'Crafting Interpreters' by Nystrom useful, alongside Clojure tools.analyzer for data processing.
- Considering but decided against Tree-sitter for parsing, opting for familiar tools like Instaparse.
- Exploring unified approaches for language implementation with integrated editor tooling.
- Looking for prior art in programming languages with first-party LSP support and bidirectional editing.
- Mentions various resources and tools like Gleam, Lady Deirdre, Roslyn, and ClojureStorm.
- Shares personal interests and readings, including language design, 3D printing, and government inefficiency.
May 2025 (Version 1.101)
a year ago
- May 2025 release of Visual Studio Code (version 1.101) with numerous updates.
- Introduction of Chat tool sets for grouping and managing related tools.
- MCP support expanded to include prompts, resources, sampling (experimental), and authentication.
- Chat UX improvements including distinct user messages and undo functionality.
- Efficiency enhancements for applying edits with aligned keybindings.
- Implicit context streamlined for adding current file as context in chat.
- Custom chat modes (Preview) allow tailored chat behaviors and tool availability.
- Task configuration errors can now be fixed using GitHub Copilot.
- Accessibility improvements including new sounds for user actions and code actions.
- Editor experience updates like find-as-you-type and custom menus.
- Process explorer now supports web when connected to a remote.
- Python extension updates including chat tools and project creation from templates.
- Extension authoring updates include MCP extension APIs and secret scanning.
- Electron 35 update with Chromium 134.0.6998.205 and Node.js 22.15.1.
- Thank you to contributors for their pull requests and issue tracking.
Andrew Ng says vibe coding is a bad name for a real and exhausting job
a year ago
- Andrew Ng criticizes the term 'vibe coding' as misleading, emphasizing that AI-assisted coding is a deeply intellectual and exhausting task.
- Despite the name, Ng supports AI-assisted coding, noting its efficiency in speeding up software development.
- Ng advocates for everyone to learn coding, arguing it's a crucial skill for interacting with computers, regardless of AI advancements.
- At his startup AI Fund, coding knowledge across all roles has led to significant productivity improvements.
- The rise of AI in coding has sparked debates on the future of software development and the necessity of technical skills.
Using Codex Well (Experiences so far)
a year ago
- OpenAI's Codex represents a significant advancement in AI coding tools, shifting from assisting in writing code to managing teams of AI agents for coding tasks.
- Effective use of Codex requires implementing numerous sanity checks to catch its occasional mistakes, leveraging its capability to set up these checks.
- Ensuring code cleanliness before employing Codex enhances its performance, with the tool itself being adept at necessary refactoring.
- Documentation of coding standards is crucial for guiding Codex to produce desirable outputs, a task Codex can assist in creating.
- Breaking down large tasks into smaller, well-defined sub-tasks is recommended to prevent Codex from getting stuck or producing unexpected results.
- Codex's ability to assist in writing tests, cleaning code, and documenting processes exemplifies the principle of using AI to improve AI's own workflow.
- The future of software engineering may involve managing AI agents more than writing code, with implications for early-career developers and senior engineers alike.
Signs Your Agile Process Sucks
a year ago
- Daily standups have turned into long, unproductive meetings.
- The backlog is full of neglected tickets and lacks proper management.
- Multiple PMs have conflicting priorities, causing confusion and delays.
- Tech debt is ignored, leading to a deteriorating codebase.
- Sprint goals are vague and often result in last-minute fixes.
- Retrospectives are ineffective and feel like a waste of time.
- Despite busy work, progress is minimal, and burnout is common.

first prev2next

About|Login

#software development

AI Can't Even Fix a Simple Bug – But Sure, Let's Fire All Our Engineers

Developers spend most of their time figuring the system out

The Copilot Delusion

The Myth of the 'Waterfall' SDLC

On "Vibe Coding"

I Let Google's Jules AI agent into my code, it did hours of work in an instant

Domain Modelers Will Win the AI Era

The Myth of Developer Obsolescence

As a developer, my most important tools are a pen and a notebook

Vibe coding for teams, thoughts to date

Why GUIs are built at least 2.5 times

What do software developers need to know to succeed in an age of AI?

Why I Wrote the Beam Book

10 Years of Betting on Rust

I Think I'm Done Thinking About GenAI for Now

How do you prototype a nice language?

May 2025 (Version 1.101)

Andrew Ng says vibe coding is a bad name for a real and exhausting job

Using Codex Well (Experiences so far)

Signs Your Agile Process Sucks