Bilingual

LLM's as a Different Kind of Intelligence
16 days ago
- LLMs represent a different kind of intelligence, distinct from human intelligence.
- Human analogies like 'junior engineer' or 'abstraction layer' are flawed because LLMs lack intent and deterministic reasoning.
- LLMs combine vast knowledge with an inability to understand intent, leading to subtle drift from user goals over time.
- Probabilistic behavior from LLMs challenges the deterministic foundations of software engineering.
- Effective use requires learning LLMs' unique failure modes rather than forcing familiar comparisons.
Canonization and the Overhang
15 days ago
- David Bessis argues that mathematics overvalues theorem proving while undervaluing creativity, clarity, and teaching, which are critical for progress.
- LLM-powered theorem provers can prove theorems without the beneficial side effects of human collaboration and intuition.
- In software, there's a parallel between code that passes tests and code that is socially constructed for team usability and iteration.
- Canonization is the process of turning one-off solutions into reusable, coherent, and compatible systems or libraries, which is time-consuming but essential for long-term value.
- Software benefits from canonized code through stable abstractions and reliable systems, contrasting with disposable, single-use software that may seem 'bad' by traditional standards.
- The Overhang refers to unrealized potential from past creativity, where connecting existing ideas can yield new solutions, a concept applicable to both mathematics and software.
- LLMs are well-suited for searching the Overhang by memorizing and pattern-matching across large corpora, unlike humans who rely on intuitive leaps.
- Open source software represents an Overhang of canonized work that can be built upon for future development.
- Valuing only theorem proving or coding without canonization and knowledge transmission risks depleting the Overhang and undervaluing critical 'glue work' like cleanup and teaching.
DSLs Enable Reliable Use of LLMs
7 days ago
- Domain-Specific Languages (DSLs) enhance LLM reliability by providing clear boundaries and constraints, reducing variability in code generation.
- DSLs act as a source of truth for software systems, enabling LLMs to serve as natural language interfaces and iterative partners in design.
- Abstractions and semantic models, like Tickloom for distributed systems, ground LLM prompts in existing vocabulary, improving accuracy and reducing hallucinations.
- LLMs operate in two phases: as brainstorming partners during DSL design and as reliable generators once the DSL is established.
- DSLs support deterministic validation (e.g., parsers or compilers), allowing autonomous agents to self-correct errors without human intervention.
- Examples include using LLMs with PlantUML for presentations and internal DSLs for testing distributed system scenarios, showcasing their versatility.
- Well-factored DSLs produce maintainable, readable artifacts that capture intent, reducing reliance on prompts as the primary source of truth.
LLMs aren't remotely like compilers or power tools
a day ago
- LLMs are fundamentally different from predictable tools like compilers or power tools, which are designed for reliability and consistency.
- Delegation is a more accurate metaphor for LLMs because they involve stochastic, unpredictable outputs similar to human agents, leading to a principal-agent problem.
- Natural language interfaces of LLMs are inherently ambiguous and imprecise, making them less reliable than structured interfaces like programming languages or physical controls.
- LLMs mimic human-like tasks but are profoundly strange and not directly comparable to existing entities, as they operate based on complex neural network processes.
- While LLMs can be useful in specific, narrow situations, their current societal discourse is often misleading and thought-terminating.

Hasty Briefsbeta

#llms

LLM's as a Different Kind of Intelligence

Canonization and the Overhang

DSLs Enable Reliable Use of LLMs

LLMs aren't remotely like compilers or power tools