#Llms
17 notas · última:
How AI does, and does not, change the way I do math
Rachel Webb argues that LLMs will change how she executes math research, but not what makes it interesting. She frames math as both personal enjoyment and a community practice.
How I use LLMs as a staff engineer
The author uses LLMs as smart autocomplete, a tutor for new domains, a fallback for bugs, and a proofreading tool, while avoiding them for full drafts and familiar production work.
The fate of “small” open source
The post argues that LLMs make small utility libraries less necessary, which weakens both their reason to exist and their role in teaching developers. Open source still matters, but the author thinks its value is shifting toward larger, more inventive, or niche work.
LLMs bring new nature of abstraction
The article argues that LLMs change software development as much as the move from assembler to high-level languages, but also add non-determinism, so programming becomes a different kind of interaction with the machine.
LLMs are mirrors of operator skill
The author argues that AI is a matter of skill: LLMs reflect the operator’s level, which makes it harder to identify qualified operators and shows that the interview process is broken.
LLMs will lie forever
The text presents an article that argues that hallucinations in LLMs are inevitable, based on mathematical arguments, and that this should change how we think about reliability, fact-checking, and the use of AI.
LLMs reward expertise
The post argues that the main skill in prompting is domain knowledge. Expert users can steer LLMs harder, spot odd answers, and get more useful results than people who only know general prompting tips.
My LLM coding workflow going into 2026
Addy Osmani lays out the workflow he converged on after a year of AI-assisted coding: spec before code, small chunks, heavy context packing, and treating every AI commit as his own responsibility.
News Feature: What are the limits of deep learning?
PNAS profile of 2019 surveys where deep learning breaks: adversarial spoofing, data inefficiency, opacity, and the missing common sense that symbolic AI once tried to hand-code.
A Primer on Large Language Models and their Limitations
Johnson and Hyland-Wood's LLM primer argues hallucination is the wrong word: LLMs produce Frankfurtian bullshit, indifferent to truth, alongside catastrophic forgetting, model collapse, and jailbreak risk.
After two years of vibecoding, I'm back to writing by hand
Mo Bitar traces two years of AI coding through excitement, spec docs, and what he calls slop, and explains why he went back to writing most of his code by hand.
Code is cheap. Show me the talk.
Kailash Nadh argues LLMs made code abundant and cheap, so accountability, provenance, and the ability to think and articulate now matter more than the code itself.
I don't like LLMs
Fowler feels both fascination and fear about AI, but talking to LLMs directly leaves him with one dominant reaction: dislike, traced mainly to the Silicon Valley culture that shaped these models.
How I use AI (Oct 2025) - Ben Stolovitz
Ben Stolovitz describes where AI helps him most: coding autocomplete, research for “pub facts,” summaries, transcription, and editing. He is skeptical of AI search, writing, and image generation.
LLMs are not like you and me—and never will be.
The post argues that LLMs are not human-like minds but autocomplete systems that lack proper world models, so they keep making obvious errors in time, history, economics, and factual reasoning.
Why Your AI Never Works on the First Try
The article argues that AI output often needs many iterations because LLMs only move step by step through possibility space. That makes AI a cognitive exoskeleton that needs human navigation.
The case against conversational interfaces
The essay argues that conversational interfaces are too slow and lossy to replace existing computing tools, but can work as a complementary layer that adds voice-driven actions across apps and workflows.