Bilingual

Some Data Should Be Code
17 hours ago
- Makefiles are often used as ad-hoc build systems for small projects, but their pattern matching and abstraction capabilities are limited.
- For complex tasks, users resort to generating Makefiles with scripts since Make doesn't scale well, highlighting the need for code-driven build graphs.
- Tools like doit, AWS CDK, and potential Python-based alternatives for CI systems demonstrate the benefits of using code to generate configuration files instead of hand-writing static data formats.
- Static data formats are often seen as simpler, but they lack the flexibility, abstraction, and type safety that code provides when used to emit data.
A sufficiently comprehensive spec is not (necessarily) code
17 hours ago
- A specification is an abstraction representing a set of possible implementations, not code itself.
- A spec is sufficient if any reasonable implementation within it satisfies the specifier, even if not fully comprehensive.
- Programmers are still needed to encode specs in formal languages like test suites, which remain distinct from code.
- LLM non-determinism and language vagueness mean specs may produce inconsistent results, reinforcing the distinction between spec and code.
- Even low-level programs (e.g., C or machine code) can be seen as specs due to varying implementations.
Wolfram LLM Benchmarking Project
9 days ago
- The Wolfram LLM Benchmarking Project focuses on evaluating large language models (LLMs) using Wolfram Language for a code generation task.
- The task converts English-language specifications into Wolfram Language code, based on exercises from 'An Elementary Introduction to the Wolfram Language'.
- The project utilizes tools to assess functional correctness, with test cases previously completed by millions of humans online.
- Results and previous versions are accessible in computable form via the Wolfram Data Repository.
- For LLM developers, Wolfram offers access to datasets, tools, and opportunities for LLM inclusion in benchmarking.
Code Was Our Medium for Thought
13 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.
AI Is a Bad Tool
15 days ago
- AI can be useful as a data distiller, condensing information retrieval and refinement steps that were previously manual.
- Using AI to generate code is inefficient; it often produces unverified, opaque outputs that lack proper abstraction and increase repetition.
- AI's opacity makes 'prompt engineering' questionable, as manipulating a black-box system without understanding its source code is unstable and unreliable.
- AI exposes a lack of proper abstraction in software development, revealing that many programming tasks could be automated by better tools or frameworks.
- The fear of AI making humans irrelevant stems from its ability to handle trivial, unabstracted work, highlighting that many software jobs were already inefficient.
- Dealing with AI's impact may require focusing on passion-driven work, full understanding, and proper abstraction to avoid reliance on AI for non-problems.

code generation

Some Data Should Be Code

A sufficiently comprehensive spec is not (necessarily) code

Wolfram LLM Benchmarking Project

Code Was Our Medium for Thought

AI Is a Bad Tool