Don't credit the LLM

4 hours ago

Peers are disclosing LLM usage for tasks like answering queries or writing documents, which the author finds odd.
Possible reasons for disclosure include giving credit, showcasing impressive results, avoiding responsibility for errors, or complying with mandates.
The author argues that tools cannot be credited or held accountable; humans should take full credit and responsibility for their work.
Taking complete ownership—both for successes and failures—is key to creating valuable work, regardless of AI involvement.

The author confesses to shipping a bug fix without personally investigating the Sentry issue, reading the code, or verifying the fix, relying entirely on Claude AI.
The AI's solution was correct, but the author realizes the workflow would look identical regardless of correctness, highlighting a dangerous reliance on unchecked AI outputs.
The author emphasizes that the job of a software engineer involves critical thinking—understanding root causes, reading stack traces, and ensuring fixes are comprehensive—not just delegating to AI.

The author acknowledges valid criticisms of LLMs, including issues with copyrighted materials, environmental impact, ethical problems, and the unsustainable nature of current AI hype.
LLMs can flood open-source projects with low-quality contributions, eroding trust and potentially harming OSS, and may reduce incentives for seniors to teach juniors due to automation of mundane tasks.
Despite criticisms, LLMs are seen as inevitable tools that can enhance human thinking when used responsibly, such as for brainstorming, grammar checks, and improving writing quality, rather than replacing human thought.

<antirez>5 days ago

The author loves writing software by hand but acknowledges that AI is irreversibly transforming programming, making manual coding no longer sensible for most projects.
In 2020, the author left their job to write a novel about AI and universal basic income, and opened a YouTube channel on AI in 2024, recognizing early the shift in programming.
Recent LLMs can complete large subtasks and medium projects alone with minimal guidance, as demonstrated by the author's examples: modding linenoise, fixing Redis tests, creating a BERT library in C, and reproducing Redis Streams changes.

The author rejects becoming a 'reverse centaur'—a human reviewer of LLM-generated code, despite increasing AI-driven contributions.
Unsolicited pull requests (PRs) are now red flags, often being low-effort LLM output; the author closes such PRs immediately.
New contribution guidelines require discussion via issue before PR submission to ensure human engagement.

A good tool should be invisible, allowing users to focus on their work rather than the tool itself.
Tool users often reframe a tool's shortcomings as enjoyable puzzles, confusing the feeling of cleverness with actual productivity.
Tool choices can become tied to personal identity, leading to tribal signaling and blind defense of flaws.

The author confesses to shipping a bug fix without personally investigating the Sentry issue, reading the code, or verifying the fix, relying entirely on Claude AI.
The AI's solution was correct, but the author realizes the workflow would look identical regardless of correctness, highlighting a dangerous reliance on unchecked AI outputs.
The author emphasizes that the job of a software engineer involves critical thinking—understanding root causes, reading stack traces, and ensuring fixes are comprehensive—not just delegating to AI.

The author acknowledges valid criticisms of LLMs, including issues with copyrighted materials, environmental impact, ethical problems, and the unsustainable nature of current AI hype.
LLMs can flood open-source projects with low-quality contributions, eroding trust and potentially harming OSS, and may reduce incentives for seniors to teach juniors due to automation of mundane tasks.
Despite criticisms, LLMs are seen as inevitable tools that can enhance human thinking when used responsibly, such as for brainstorming, grammar checks, and improving writing quality, rather than replacing human thought.

<antirez>5 days ago

The author loves writing software by hand but acknowledges that AI is irreversibly transforming programming, making manual coding no longer sensible for most projects.
In 2020, the author left their job to write a novel about AI and universal basic income, and opened a YouTube channel on AI in 2024, recognizing early the shift in programming.
Recent LLMs can complete large subtasks and medium projects alone with minimal guidance, as demonstrated by the author's examples: modding linenoise, fixing Redis tests, creating a BERT library in C, and reproducing Redis Streams changes.

The author rejects becoming a 'reverse centaur'—a human reviewer of LLM-generated code, despite increasing AI-driven contributions.
Unsolicited pull requests (PRs) are now red flags, often being low-effort LLM output; the author closes such PRs immediately.
New contribution guidelines require discussion via issue before PR submission to ensure human engagement.

A good tool should be invisible, allowing users to focus on their work rather than the tool itself.
Tool users often reframe a tool's shortcomings as enjoyable puzzles, confusing the feeling of cleverness with actual productivity.
Tool choices can become tied to personal identity, leading to tribal signaling and blind defense of flaws.

Hasty Briefsbeta