
Model routing: send different tasks to different systems
Routing classifies a request and sends it to a suitable processing path.
Independent AI news & analysis
Stories 1045–1080 of 1108.

Routing classifies a request and sends it to a suitable processing path.

Parallel workflows run independent subtasks at the same time and combine their results.

An evaluator–optimizer workflow separates generation from feedback and revision.

An agent tool needs a clear purpose, precise inputs, and a result the model can interpret.

A tool error should tell the agent what failed and whether an action completed.

A sandbox restricts the environment in which an agent or its code runs.

Agent memory preserves information beyond the current model context.

Context compaction condenses a long working history into the information needed to continue.

MCP resources provide contextual data, while tools expose executable functions.

MCP prompts expose reusable message templates through a server.

A computer-use agent, in Anthropic’s tool docs, gets screenshots plus mouse and keyboard actions. Your application runs those calls, and a person should confirm actions with real-world consequences.

Multiple agents can divide work, but they also add communication, duplicate context, and integration overhead.

Human-in-the-loop AI, in the MCP tools specification, means a person can deny a tool call. The useful check shows the exact action, not a generic yes.

An agent evaluation suite tests complete task outcomes, not only the final text.

The 2017 paper replaced recurrence and convolution with attention alone. On WMT 2014 it reported 28.4 BLEU for English-to-German and 41.8 BLEU for English-to-French.

Hoffmann and colleagues found that, for a fixed training budget, model size and the number of training tokens should grow together. Chinchilla was their 70-billion-parameter test of that claim.

Reinforcement learning from human feedback uses preference information to improve model behavior.

Direct Preference Optimization trains on pairs of preferred and rejected answers with a classification loss. The 2023 paper uses that loss instead of a separate reward model plus reinforcement learning.

Constitutional AI studies using a set of principles to guide critique, revision, and AI-generated feedback.

LoRA freezes a pretrained model and trains small low-rank matrices instead. On the paper’s GPT-3 175B comparison, that cut trainable parameters by 10,000 times versus full fine-tuning.

QLoRA combines a quantized frozen base model with trainable low-rank adapters.

Lost in the Middle studies how models use relevant information placed at different positions in long inputs.

Self-consistency samples multiple reasoning paths and combines their final answers.

SWE-bench evaluates software tasks based on real repository issues and associated fixes.

HumanEval is a set of programming problems introduced with research on code-trained language models.

HELM stands for Holistic Evaluation of Language Models.

TruthfulQA evaluates whether models produce truthful answers to questions designed around common misconceptions.

ReAct combines reasoning and actions in a task-solving process.

Toolformer studies teaching a model when and how to use external tools, including interpreting their outputs.

FlashAttention is an attention algorithm designed around GPU memory movement.

Speculative decoding uses a faster draft process to propose tokens, then checks them with a target model.

Evaluate an AI vendor against your actual workflow, data requirements, and failure tolerance.

Total cost of ownership includes operating the complete AI system.

Provider lock-in can come from APIs, prompt formats, tools, stored state, and operational dependencies.

Open weights means the parameters are available under some terms. The Open Source AI Definition 1.0 also asks for training-data information and the code used to train and run the system.

A model’s license governs permitted use under its stated terms.