#Large-Language-Models
5 notas · última:
AI Coding Degrades: Silent Failures Emerge
The author says AI coding assistants have plateaued in 2025 and now fail more quietly, often producing code that runs but gives wrong results. He links this to training on user-accepted outputs.
7 Next-Generation Prompt Engineering Techniques
The article reviews seven advanced prompt engineering techniques, explaining how each one helps LLMs produce more accurate, structured, relevant, or verified outputs.
Hallucinations are NOT bugs
The post argues that LLM hallucinations are built into how these models work, with measurable error rates even on factual prompts, because generalization and precision are in tension.
How o3 and Grok 4 Accidentally Vindicated Neurosymbolic AI
The essay argues that pure scaling has hit diminishing returns, while LLMs improve when paired with symbolic tools. It presents this as a belated vindication of neurosymbolic AI.
GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.
Gary Marcus argues that GPT-5’s debut exposed the limits of hype around OpenAI, with public disappointment, lingering errors, weak generalization, and no evidence that pure scaling gets to AGI.