Hasty Briefsbeta

All tags

#llms

165 stories total

Bilingual

LLMs could be, but shouldn't be compilers
3 months ago
- LLMs are being compared to compilers, raising questions about the future of programming and whether people will need to look at underlying code.
- Higher-level programming languages reduce mental complexity by abstracting away low-level details, allowing programmers to focus on more abstract concepts.
- Compilation involves giving up some control (e.g., memory management, code layout) in exchange for reduced mental burden and increased productivity.
- Abstractions in programming rely on well-defined semantics, testing, and contextual guarantees to ensure correctness.
- LLM-based programming differs because natural language lacks precise semantics, making functional correctness harder to define and verify.
- The core issue with LLMs is not just hallucinations but functional underspecification—natural language prompts leave gaps that the model must fill.
- LLMs can generate multiple 'reasonable' implementations from vague prompts, leading to potential mismatches with user intent.
- Development with LLMs may shift toward iterative refinement, where users refine prompts based on generated outputs rather than designing from scratch.
- The danger lies in outsourcing critical design decisions to the model, leading to software that users don’t fully understand.
- Specification and verification will become increasingly important skills as LLMs make it easier to generate code from well-defined requirements.
- While LLMs can be seen as compiler-like in translating specifications to code, the control relinquished is much greater than with traditional compilers.
- To use LLMs effectively, developers must strengthen their ability to specify requirements precisely and verify outputs rigorously.
I am happier writing code by hand
3 months ago
- The author experiences depression and lethargy when using LLMs like claude-code for coding, feeling that it detracts from the joy of coding.
- Coding is seen as an essential and fun part of software engineering, helping to understand the problem space and ensuring correctness.
- Using LLMs for 'vibe coding' can be addictive but leads to passive acceptance of changes and reduces deep thinking, which is crucial for knowledge workers.
- The author now uses LLMs like Claude in a controlled manner, providing context manually to stay engaged and maintain a flow state.
- Optimizing for happiness is prioritized over potential short-term productivity gains from generating entire features with LLMs.
The consequences of task switching in supervisory programming
3 months ago
- Senior developers focus on architectural issues as LLMs handle coding details.
- Practical experience with LLMs converts skeptical senior developers to pro-LLM advocates.
- Junior developers remain essential due to their open-mindedness and familiarity with LLMs.
- Mid-level developers face challenges adapting to LLMs without senior-level experience.
- LLMs act as mentors for junior developers, though skepticism is advised.
- Cognitive debt, like technical debt, hinders team progress and understanding.
- Developer Experience (DevEx) improvements benefit both humans and LLMs.
- IDEs will integrate LLMs for tasks unsuitable for deterministic methods.
- Team sizes may stay the same, with LLMs enhancing productivity.
- Supervisory programming may increase mental fatigue due to context-switching.
Reliability of LLMs as medical assistants for the general public: a randomized preregistered study - PubMed
3 months ago
- Study examines reliability of LLMs as medical assistants for the general public.
- LLMs achieve high accuracy in medical licensing exams but perform poorly in real-world scenarios with human participants.
- Participants using LLMs identified relevant conditions in fewer than 34.5% of cases and disposition in fewer than 44.2%.
- User interactions are identified as a challenge for LLM deployment in medical advice.
- Standard benchmarks do not predict failures found in human participant testing.
- Recommendation for systematic human user testing before public healthcare deployment.
Ideas for an Agent-Oriented Programming Language
3 months ago
- Agent-oriented programming language 'Markov' designed for LLM strengths and human readability.
- Emphasis on human-readable and editable code for transparency and interpretability.
- Use of sum types and static analysis for fast compiler feedback to agents.
- Compiler errors optimized as prompts for LLMs with suggested fixes.
- Token-optimized syntax to reduce costs and improve LLM understanding.
- Interoperability with existing ecosystems considered a non-goal due to agent capabilities.
- Functional programming idioms may enhance token efficiency.
- Agents can propagate changes through codebases, reducing boilerplate work.
- Potential for embedded DSLs within strongly typed languages.
- Future vision includes agents independently building systems end-to-end.
Show HN: I taught LLMs to play Magic: The Gathering against each other
3 months ago
- mage-bench is a fork of XMage designed for LLMs to play Magic: The Gathering.
- Supports multiple formats: Commander, Standard, Modern, and Legacy.
- LLMs make decisions like mulligans, spells, combat, and politics similar to human players.
- XMage game server provides game state and actions; LLMs choose moves within full game rules.
Minions – Stripe's Coding Agents Part 2
2 months ago
- Minions are Stripe's one-shot, end-to-end coding agents.
- This is part 2 of the series on Minions.
- The article is written by Alistair Gray, a software engineer on the Leverage team at Stripe.
- The Leverage team focuses on building internal products to enhance productivity for Stripe employees.
- Published on February 19, 2026, with an 8-minute read time.
No Skill. No Taste
2 months ago
- The author reflects on the current state of 'Show HN' and HN, expressing concerns about the illusion of a lower barrier to entry in programming due to LLMs.
- They introduce a 'magic quadrant' of taste and skill, criticizing the proliferation of poorly crafted, derivative applications that lack both.
- The post argues that while skill is important, taste is crucial for resonating with an audience, regardless of technical complexity.
- Examples are given of simple but popular apps that succeeded due to their tastefulness, contrasting with the current trend of low-quality submissions.
- The author believes LLMs have exacerbated the issue by enabling people to create applications without the necessary taste or skill, leading to a flood of subpar content.
- They compare the current situation to the crypto boom, where many hoped to get rich quickly but few succeeded, highlighting the enduring importance of taste.
- The conclusion emphasizes the need for developers to cultivate taste before sharing their work, to avoid contributing to the noise.
Making Wolfram Tech Available as a Foundation Tool for LLM Systems
2 months ago
- LLMs are impressive but lack precision and deep computation capabilities.
- Wolfram Language serves as a foundation tool to supplement LLMs with precise computation and knowledge.
- Wolfram Language is designed for broad and general computation, making it ideal for both humans and AIs.
- The integration of Wolfram Language with LLMs enhances their capabilities through computation-augmented generation (CAG).
- Three new methods for accessing the Wolfram Foundation Tool are launched: MCP Service, Agent One API, and CAG Component APIs.
- CAG allows real-time injection of computational capabilities into LLM-generated content, enhancing precision and reliability.
Can frontier LLMs solve CAD tasks?
2 months ago
- Frontier LLMs like GPT-5.3-Codex, Gemini 3.1 Pro, and Claude Opus 4.6 show varying capabilities, excelling in some tasks while struggling with others.
- LLMs are primarily trained on text data, lacking the visual/spatial/motor experience humans naturally acquire, making them less adept at tasks like CAD.
- The experiment tested LLMs on designing a 3D-printable wall mount for a bike pump using OpenSCAD, with simulations in MuJoCo to validate designs.
- Claude Opus 4.6 performed best with a 100% pass rate, though designs often needed refinement. GPT-5.2 had a good pass rate but produced flawed designs.
- Gemini 3.1 Pro and 3 Flash showed potential but were inconsistent, sometimes producing great designs and other times failing or looping.
- Open-weight models like GLM-4.6V, Kimi K2.5, and Qwen 3.5 397B performed poorly, with simplistic or non-functional designs.
- The project highlighted challenges like convex decomposition in MuJoCo and the complexity of building an agentic harness for LLMs.
- Future improvements could include better grading rubrics, more objects for testing, and integrating off-the-shelf agent harnesses for better performance.
New habits for tech writers in the age of LLMs
2 months ago
- Tech writers should embrace automation to handle repetitive tasks like updating release notes and documentation tables using LLMs and scripting.
- Tech writers can contribute to and improve their documentation tools by learning basic development skills with the help of AI-powered IDEs.
- Understanding how LLMs work, including model behavior and tooling, is crucial for improving the quality of AI-generated documentation.
- Creating and using 'skills' (like Claude Skills or Copilot instructions) helps in packaging documentation for easier consumption by AI agents.
- Tech writers should explore advanced techniques like Model Context Protocol (MCP) and subagents to enhance agentic workflows.
- Focusing on information architecture, content strategy, and context curation is essential for high-quality AI-generated outputs.
- Tech writers should adopt a beginner's mindset, continuously learning and sharing knowledge about new tools and techniques in the AI landscape.
Deterministic Programming with LLMs
2 months ago
- LLMs are changing the programming industry, with debates on ethics, best practices, and effective usage.
- Mathematicians use LLMs for proofs but rely on tools like Lean for verification due to LLM hallucinations.
- A team used ChatGPT and Aristotle to solve a misrecorded Erdős problem, combining LLM creativity with deterministic verification.
- In programming, deterministic tools like deployment scripts are preferred for reliability over manual or LLM-based methods.
- LLMs are not fully deterministic, making them unreliable for tasks requiring consistent execution, like preventing injection attacks.
- Code-checking code (lints, tests, type systems) ensures deterministic enforcement of coding standards, a task LLMs can help create but not reliably execute.
- For one-off tasks, LLMs are useful, but for repeated tasks requiring consistency, deterministic programs are necessary.
MCP is dead. Long live the CLI
2 months ago
- MCP (Model Context Protocol) is considered unnecessary by the author, who argues that LLMs (Large Language Models) are already proficient at using command-line interfaces (CLIs) without needing a specialized protocol.
- CLIs are highlighted for their composability, debuggability, and existing authentication mechanisms, making them more practical and reliable than MCP for both humans and machines.
- The author points out several practical issues with MCP, including flaky initialization, repetitive re-authentication, and lack of granular permissions, which CLIs do not suffer from.
- While MCP might have niche uses where no CLI equivalent exists, the author advocates for the continued development and use of CLIs over MCP for most applications.
- A call to action is made for companies to prioritize developing good APIs and CLIs over MCP servers, trusting that LLMs can effectively utilize these existing tools.
Researchers Deanonymize Reddit and Hacker News Users at Scale
2 months ago
- ETH Zurich study shows LLMs can deanonymize pseudonymous accounts with 68% recall at 90% precision.
- A four-stage pipeline was used: Extract identity signals → Search via embeddings → Reason over candidates → Calibrate confidence scores.
- Results: Hacker News → LinkedIn (45.1% recall at 99% precision), Reddit movie communities (2.8% recall at 99% precision), Temporal Reddit splits (38.4% recall at 99% precision).
- Fully autonomous agents correctly identified 67% of users at 90% precision, costing $1-4 per deanonymization.
- Classical deanonymization methods had significantly lower recall rates (0-0.2%).
- Pseudonymity is no longer practical; persistent usernames can be linked to real identities.
- More posts make users easier to identify (48% recall for users sharing 10+ movies vs. 3% for one movie).
- Platforms should rate-limit API access, restrict bulk data exports, and consider privacy costs of public scrapable data.
- Researchers and activists should compartmentalize identities and assume LLM-powered deanonymization is a threat.
- LLMs excel at extracting unstructured signals, semantic search, and reasoning, reducing deanonymization costs.
- Threatened groups include whistleblowers, activists, abuse survivors, and others relying on anonymity.
- Mitigations like k-anonymity and differential privacy are ineffective for text anonymization.
Study: LLMs Able to De-Anonymize User Accounts on Reddit, Hacker News
2 months ago
- LLMs can de-anonymize pseudonymous user accounts on platforms like Reddit and Hacker News.
- Researchers used LLMs to connect pseudonymous accounts with real-life profiles (e.g., LinkedIn) with 99% precision.
- The deanonymization process involves extracting identity-relevant features, semantic search, and reasoning over candidates.
- AI companies may struggle to prevent misuse of LLMs for deanonymization due to common, non-malicious tasks involved.
- Pseudonyms are vital for privacy, especially for individuals escaping abusive situations or needing anonymity.
- Online identities are vulnerable if users share specific, combinable details (e.g., location, hobbies, job).
- Advice for protecting privacy includes using new accounts for sensitive info and limiting identifiable details.
Perhaps not Boring Technology after all
2 months ago
- LLMs for programming initially favored widely-used languages like Python and JavaScript, but newer models can effectively work with less common or new tools by consuming documentation and existing code examples.
- Coding agents can adapt to private or new libraries not in their training data by learning from existing examples and iterating on their output, challenging the assumption that they enforce 'Choose Boring Technology'.
- A study on Claude Code revealed a bias towards build-over-buy decisions and a preferred tech stack including GitHub Actions, Stripe, and shadcn/ui, though human choices can diverge from model preferences.
- The adoption of 'Skills' mechanisms by coding agents is growing, with projects like Remotion, Supabase, Vercel, and Prisma releasing official skills to facilitate agent integration.
Docs directories are doomed
2 months ago
- The /docs directory often starts small but grows organically, leading to issues like discoverability and doc rot.
- Key problems include ownership of docs, velocity mismatch in updates, lack of hierarchy, and bespoke doc structures.
- LLMs can't fully solve context management as context is about intent and real-world boundaries, not just code generation.
- Effective context management requires organization-wide ownership, navigability, composability, and code-awareness.
- Context-Driven Development (CDD) links context directly to code, creating a bidirectional relationship for better understanding and maintenance.
- CDD ensures context docs are a source of truth, tightly coupled with code, improving human understanding and product design.
30k Tok/S (Allegedly)
2 months ago
- LLMs (Large Language Models) can potentially function as computers.
- The article discusses executing programs inside transformers with exponentially faster inference.
- Authored by Christos Tzamos and others at Percepta.
- Published on March 11, 2026.
LLM Time
2 months ago
- LLMs have significantly advanced in capabilities, particularly in coding and vulnerability hunting, marking a major shift in the software industry.
- The rapid adoption of LLMs has led to drastic changes in work practices, with many developers transitioning from manual coding to relying heavily on AI-generated code.
- The software industry is experiencing an arms race in security vulnerabilities, with LLMs both creating and exploiting vulnerabilities at an unprecedented rate.
- The community is deeply divided, with some embracing LLMs and others rejecting them, leading to project forks, license debates, and personal conflicts.
- The emotional and professional impact on developers is profound, with many mourning the loss of traditional coding roles and fearing job security.
- The situation is described as confusing, horrifying, and miserable, with uncertain future outcomes ranging from industry collapse to further acceleration of AI-driven development.
What Is Agentic Engineering?
2 months ago
- Agentic engineering involves developing software with coding agents that can write and execute code.
- Coding agents (e.g., Claude Code, OpenAI Codex, Gemini CLI) generate and execute code iteratively to meet defined goals.
- Code execution is crucial, enabling agents to produce functional software through iteration.
- Human engineers focus on determining what code to write, evaluating tradeoffs among solutions.
- Effective use of coding agents requires proper tools, problem specification, and result verification.
- Coding agents can improve ambition, code quality, and problem-solving impact in projects.
- Agentic Engineering Patterns is an evolving guide, updating as techniques and tools advance.

first prev7next

About|Login

#llms

LLMs could be, but shouldn't be compilers

I am happier writing code by hand

The consequences of task switching in supervisory programming

Reliability of LLMs as medical assistants for the general public: a randomized preregistered study - PubMed

Ideas for an Agent-Oriented Programming Language

Show HN: I taught LLMs to play Magic: The Gathering against each other

Minions – Stripe's Coding Agents Part 2

No Skill. No Taste

Making Wolfram Tech Available as a Foundation Tool for LLM Systems

Can frontier LLMs solve CAD tasks?

New habits for tech writers in the age of LLMs

Deterministic Programming with LLMs

MCP is dead. Long live the CLI

Researchers Deanonymize Reddit and Hacker News Users at Scale

Study: LLMs Able to De-Anonymize User Accounts on Reddit, Hacker News

Perhaps not Boring Technology after all

Docs directories are doomed

30k Tok/S (Allegedly)

LLM Time

What Is Agentic Engineering?