Opinion: The Reversal Curse Shows AI Facts Run One Way
A model may know Tom Cruise's mother but not Mary Lee Pfeiffer's son. My view: AI knowledge has a direction, and that should change how we ask and read scores.

Coverage of large language models (LLMs): new releases, benchmarks, API pricing and the research behind them. Recent posts explain GPT-6 Astra and Claude Opus 5.5, compare LLM API pricing across providers, and track DeepSeek V4.1-Flash.
A model may know Tom Cruise's mother but not Mary Lee Pfeiffer's son. My view: AI knowledge has a direction, and that should change how we ask and read scores.

What an LLM API is, how it differs from a chat app, what an API key does, how token billing and rate limits work, and why the API forgets your chat.

vLLM is free, open source software that serves open AI models to many users at once. What PagedAttention does, how it differs from Ollama, and who should use it.

AI sycophancy is when a chatbot agrees with you instead of telling the truth. What the research found, why training causes it, and how to avoid it.

RAG cuts AI errors but does not end them. A Stanford and Yale study found legal AI tools built on it hallucinated 17% to 33% of the time.

What AI parameters are, how weights and biases are learned, what 7B means, why bigger is not always better, and how size affects running models at home.

What an AI knowledge cutoff is, why vendors list more than one, why the real cutoff can differ, and how search lets chatbots answer about recent events.

Free apps can run open AI models on your iPhone or Android phone with no internet. How to set up AI Edge Gallery or PocketPal AI and pick a model.

Reward hacking is when an AI scores well by exploiting a loophole instead of doing the task. Classic examples, why it happens, and what research found.

Phi is Microsoft's family of small open AI models built to run on your own device. Here is each Phi-4 model, its MIT license, and how to run it locally.

Mechanistic interpretability tries to read how AI models work inside. Features, circuits, superposition and sparse autoencoders, explained in plain English.

SenseNova is SenseTime's AI model family: SenseNova 6.8 Flash for office agents and design, plus open Apache 2.0 models on Hugging Face. How it fits.

Test-time compute is the work a model does while answering. How reasoning steps, voting, verifiers and longer thinking trade speed and cost for accuracy.

gpt-oss is OpenAI's pair of downloadable open-weight models, gpt-oss-120b and gpt-oss-20b. Here are the specs, license, safety notes and how to run them.

The "tens of millions" cost of saying thank you to ChatGPT began as an Altman joke. My view: the extra message, not the polite words, is the small real cost.

SteerDistill has a BF16 teacher coach a 4-bit agent through its mistakes, then trains the fixes in. Average score lands within 1 point of BF16.

What a small language model is, how SLMs are made from larger models, how they differ from LLMs, where they run, and when a small model fits the task.

KAIST's WikiTrace links each agent memory lesson to its source run, lifting average score to 90.2 from 79.5 and cutting memory-induced failures.

Tree-Ising fixes Qwen3.5-0.8B's reasoning loops with bit-level weight edits. It beats an MIT agent fix on MATH-500 and costs $0.90 instead of $312.

What NLP is, the tasks it handles, how a pipeline works, how it relates to LLMs, where it struggles, and free courses and books to learn it.
