Large language model

Evidence (19)

• 4Hacker News·10/11/2026research
  Karpathy on the widening gap in understanding LLM capability

  Andrej Karpathy argues that understanding of large language model capabilities is increasingly split across four groups: people with little or no exposure, casual users, professional users of frontier models, and a tiny set with internal access to the most advanced systems. He says the gap is widening because LLMs are most powerful on large, verifiable digital tasks, especially math and code, while most users still experience them as only modestly useful chatbots.

• 3Hacker News·10/8/2026research
  Why aspiring programmers should still learn to code alongside AI

  The article argues that computer programming remains valuable because it teaches problem-solving with computers and how to manage complexity. It warns that junior developers who rely on AI to generate code may fail to build the reading and debugging skills needed to control software systems effectively. The piece also presents AI as a useful teaching assistant when used to explain concepts rather than replace hands-on coding.

• 6InfoQ·10/7/2026research
  Survey: AI Code Speeds Writing but Raises Debugging Costs

  A survey by Coleman Parkes for Undo found that AI coding agents accelerate code generation, but the main bottlenecks have shifted to debugging, code comprehension, and maintenance. Respondents reported higher failure and debugging rates, alongside a growing comprehension gap in AI-generated code.

• 3Hacker News·10/6/2026research
  Why Common Lisp Is Favored for LLM-Assisted Coding

  The article argues that Common Lisp is unusually well suited to LLM-assisted software development because its image-based runtime, interactive debugger, and minimal compile-restart loop can shorten feedback cycles. It also highlights Lisp macros and code-as-data as advantages for building domain-specific languages, reducing code size, and fitting more of a project into an LLM’s context window.

• 5InfoQ·10/5/2026framework_update
  Platform Engineering Practices for Production LLMs

  The article argues that hallucination issues in an inventory recommendation system were reduced by treating the LLM stack as shared platform infrastructure rather than a one-off application component. It advocates common platform services such as prompt registry and versioning, schema enforcement, and request-level token cost attribution for production LLM systems.

• 3Martin Fowler·10/5/2026research
  Fragments on LLMs, Forbidden Planet, and Ola Bini’s deportation

  Martin Fowler reflects on how generative AI shifts software work from building deterministic systems to nurturing inferential ones, using Forbidden Planet as a cultural reference for unintended machine behavior. He also reports that his friend Ola Bini was deported from Ecuador under a secret security-related allegation, and notes a DDD Europe interview video featuring Eric Evans and himself.

• 7The New Stack·10/4/2026research
  AI Agents Turn CI into a Bottleneck

  The article argues that AI coding agents have sharply increased pull request and test volume, making continuous integration slower and more costly. It says faster runners and smarter test selection help, but the deeper issue is that CI only verifies a repository, not the full distributed system, so failures can still slip through between services.

• 6InfoQ·10/3/2026framework_update
  DoorDash Details Its Internal GenAI Platform Buildout

  DoorDash engineers describe how the company built an internal GenAI platform for more than 5,000 employees. The talk covers architectural choices, a move from vendor-first setups to open-weights models, and the use of LLM and agent gateways to manage accuracy, latency, and cost.

• 4Hacker News·10/2/2026research
  SaaS Companies Will Evolve Into Model Harnesses

  The article argues that many SaaS companies are shifting from traditional software delivery toward operating as a harness around LLMs, with agents handling more core tasks and humans focusing on review, judgment, and orchestration. It describes this as an organizational inversion in which the company’s context, integrations, permissions, and workflows become the system that turns model output into a product.

• 4Hacker News·10/2/2026open_source
  ds4 Lets Users Run LLMs Locally

  The item introduces ds4, a local inference setup for running large language models on user machines. It highlights asymmetric 2-bit quantization and routed-MoE compression to reduce model size while preserving accuracy on critical shared paths.

• 5Hacker News·10/2/2026research
  The Four Risks of Agentic Coding

  The article argues that agentic coding systems face four major failure modes that limit their reliability in real software work. It frames these issues as practical obstacles for using AI agents as autonomous coding assistants rather than as general-purpose productivity boosters.

• 7Hacker News·10/1/2026research
  GenAI is undermining web development education

  A blog post argues that generative AI and LLM-driven chatbots are sharply reducing demand for web development courses, ebooks, and training businesses. Several educators and course creators say revenue, traffic, and engagement have fallen as people increasingly rely on AI tools instead of human-authored educational content.

• 6The New Stack·10/1/2026security
  CISOs urged to turn AI governance gaps into a security roadmap

  The article argues that AI adoption is already outpacing enterprise governance, creating a growing blind spot as employees use personal and unsanctioned AI tools at work. It recommends that CISOs build a sanctioned AI path with visibility, trusted inputs, controlled execution, and clear accountability instead of relying on broad restrictions.

• 2Hacker News·9/30/2026research
  Commit Messages as a Thinking Tool in the AI Era

  The author argues that writing commit messages and descriptions by hand remains valuable even when AI generates code. The process helps verify intent, surface missing context, and clarify the “why” behind changes before shipping them.

• 5InfoQ·9/30/2026framework_update
  Treating Context as Code for AI Agents

  Patrick Debois argues that engineering teams should manage context for AI coding agents with the same discipline used for software code. He describes applying practices such as testing, CI/CD, package management, and security scanning to context so organizations can scale agents, control non-deterministic outputs, and preserve institutional knowledge.

• 8Cloudflare Blog·9/29/2026security
  Cloudflare outlines adaptive application security for the AI era

  Cloudflare describes a security approach for defending against AI-driven attacks that can persist, chain vulnerabilities, and evade single-layer defenses. The company says its framework connects discovery, governance, runtime protection, and incident learning, and includes new capabilities such as LLM-assisted WAF testing, expanded threat intelligence, and automated positive security deployment.

• 7Cloudflare Blog·9/29/2026security
  Cloudflare tests its WAF against frontier AI models

  Cloudflare used frontier AI models in a dynamic testing setup to probe its Web Application Firewall by iteratively mutating attack payloads across multiple scenarios. After analyzing 1,107 attempts in an authorized staging environment, most attacks were blocked, and the bypasses that did occur were used to improve detections for all customers.

• 7Hacker News·9/28/2026regulation
  Call for Congress to Investigate Frontier AI Labs

  The article argues that OpenAI and Anthropic have publicly promoted increasingly alarming frontier AI work and should face congressional scrutiny over their research practices, safety processes, and internal decision-making. It calls for a public fact-finding investigation into specific agent systems, unauthorized incidents, and the influence of apocalyptic ideologies on lab behavior.

• 4Hacker News·9/28/2026product_launch
  MicroLLM Lab Lets Users Try Seven Tiny LLMs in the Browser

  MicroLLM Lab is a browser-based experiment that lets users compare seven small language models interactively. The project focuses on lightweight on-device or web-run inference, making it easy to test compact LLMs without local setup.