Skip to content

#教程/实践

4 today

Oct 9Friday

AI HOT · Tips & opinions

Hugging Face engineer builds 7 small models with ML Intern for about USD 103

A Hugging Face engineer used ML Intern to build seven small models, with CPU and GPU job costs totaling about USD 103. The projects cover citrus disease detection, image LoRA, prompt rewriting and distillation from an image model. Pocket Rewriter 0.8B runs on CPU with a 99.7% effective output rate.

Why it matters: The engineer wrote baseline evals, pre-training smoke tests and a budget cap into the ML Intern prompt, splitting training into checkable steps.

Computing Life · Share · Yage

GPT-6 guide: a build checklist for long-running agents, and what's still unsolved

A reading of the GPT-6 guide lays out how to build long-running agents, shifting the focus from homegrown memory management to collaboration and authorization while a task runs. Drawing on OpenAI documentation, it covers compaction and persistent reasoning, mid-turn steering, asynchronous tool calls and multi-agent delegation, plus steps for model choice, caching, skill rules and human approval. It also flags limits: correction instructions lost on disconnect, mutually exclusive API parameters, and encrypted compaction that is hard to audit.

Why it matters: It splits long-task engineering into managed memory and collaboration control, and assigns clear roles to async execution, mid-run correction and human approval.

Oct 8Thursday

claude.dev Blog

How Anthropic builds reliable agent automation with Claude Managed Agents

Anthropic published a reference implementation on Claude Managed Agents (beta) that polls Slack and GitHub for changes on a schedule and posts a digest back to Slack.

Why it matters: Anthropic splits the scheduled digest into bookmark-based resumption, send confirmation and permission isolation rules, giving reusable patterns for background agents to avoid missed reports, duplicate notifications and overreach.

Oct 7Wednesday

claude.dev Blog

Claude Code cloud sessions guide: parallel tasks, environment setup, local handoff

A Claude Code cloud sessions guide walks through four real runs covering parallel coding in separate VMs, environment setup, and how cloud sessions hand off to local ones. Three tasks started within 16 seconds of each other and all finished 87 seconds after the first launch. A cache fix passed 40 consecutive tests. Separate branches don't share fixes, so merge order has to be planned.

Why it matters: It shows how to split work by file boundary across parallel cloud tasks, and to build acceptance commands and branch merge order into the task design.

Oct 4Sunday

Hacker News front page

Alexey drops line-by-line agent code review for plan review and engineering guardrails

Alexey changed his agent development flow to reviewing plans, setting engineering guardrails and allowing auto-merge, and says 562 PRs were merged in two weeks. On his internal platform Sightline, a PR must pass unit tests, end-to-end tests, CI, Bugbot and Strix checks before Mergify lets it into the merge queue. He uses a coordinator agent to dispatch sub-agents that execute plans, controls risk with feature flags and small deployments, and tracks long-running goals through HTML progress reports.

Why it matters: Alexey moves code review to plan review and constrains agent auto-merge with mandatory tests, a merge queue and feature flags, putting quality control into engineering guardrails.

Oct 3Saturday

OpenAI News

OpenAI GPT-6 series guide: model choice, prompts and long-task management

OpenAI's GPT-6 series guide covers how to pick a model, tune prompts and skills, and manage the time and cost of long tasks. GPT-6 Astra targets hard reasoning, GPT-6.1 Sol targets complex coding, research and computer use, and GPT-6 Luna targets well-defined repetitive work at scale. The guide says to measure task success rate, latency and cost per successful task before deploying.

Why it matters: OpenAI folds model choice, reasoning effort and cache management into one production practice, with success rate, latency and cost per successful task as the basis for tuning.

Oct 2Friday

AI HOT · Models

NVIDIA details how Blackwell GPUs accelerate OpenAI's GPT-6 Astra Ultrafast

GPT-6 Astra Ultrafast runs on NVIDIA Blackwell GPUs and generates tokens up to 8x faster than Astra Standard mode. OpenAI used its own models to tune NVIDIA's GPU inference software, and faster responses cut the wait for coding agents that edit, test and debug in loops.

Why it matters: OpenAI tuned NVIDIA's GPU inference software with its own models, and the speedup targets the wait time coding agents hit when they edit, test and debug repeatedly.

claude.dev Blog

Claude Code mods intro: build a context monitor plugin

This Claude Code mods guide builds a roughly 80-line Token Weather plugin from scratch that shows live context window usage above the prompt. It targets Claude Code 2.1.287 or later and covers event hooks, hot reload, state persistence, verification tests and plugin distribution.

Why it matters: A hands-on walkthrough that ties event hooks, state persistence and testing into reusable steps for customizing the Claude Code interface.

AI HOT · Tips & opinions

LangChain explains how to build a model router in an Agent Harness

LangChain described how it built a model router inside the Agent Harness of Open SWE. In A/B tests, median cost per thread dropped 64%, with no measurable change in quality.

Why it matters: LangChain combines task analysis, model tiers and result tracking in agent routing, a reusable way to control model cost by task complexity.

Computing Life · Share · Yage

How OpenAI's AI worm was produced in sandbox training and used for safety training

OpenAI's self-replicating prompt worm was produced inside controlled sandbox training, and the report says no effects were seen beyond simulated tool calls during training and evaluation. The article explains how GPT-Red trains attackers by rewarding unauthorized actions and forwarding of attack instructions, generating injections that spread through email and other channels, with bulk attack samples fed into GPT-5.6 training.

Why it matters: OpenAI uses automated red teaming to mass-produce attack samples and fold them into model training, moving defense from pre-release manual testing to continuous adversarial work during training.

Sep 30Wednesday

Hacker News front page

Text classification from bag-of-words to Jev: a language model that classifies without per-task fine-tuning

Jev turns text classification into a general decision service that needs no per-task fine-tuning. The article traces the history of classification techniques and tests Jev's limits. On the 25,000-review IMDb test set, the Choice API hit 96.47% accuracy at a total cost of $0.6492. The article notes it cannot confirm whether the test set was part of the training data.

Why it matters: It compares Jev with bag-of-words models, ModernBERT and general LLMs, and lays out the cost and fit of direct cross-task classification versus fine-tuning a dedicated model.

Sep 29Tuesday

OpenAI News

OpenAI drafts a safety case framework for frontier AI training

OpenAI is building a safety case framework for frontier reinforcement learning training, arguing that structured, evidence-based risk arguments should come before training continues. The proposal covers alignment training, containment and monitoring, plus cross-team challenge, an executive veto, pause mechanisms, incident investigation and public disclosure. It is being rolled out inside OpenAI, and the safety case stays a goal built up over time. The document does not cover the broader alignment requirements for internal and external deployment.

Why it matters: OpenAI splits training safety into technical safeguards, operational approval and incident investigation, folding monitoring thresholds, pause mechanisms and accountability into how it manages training risk.

Sep 28Monday

claude.dev Blog

How to design evals and hillclimb with Claude

Claude's claude-api skill added guidance on eval design and hillclimb optimization. In Claude Code, /claude-api build-eval builds an eval and /claude-api hillclimb improves an app step by step.

Why it matters: Claude organizes eval design and step-by-step optimization into workflows that need human confirmation, using held-out test sets and noise checks to cut overfitting and wasted changes.

Sep 25Friday

claude.dev Blog

How Opus 5.5 task cost in Claude Code is calculated

Task cost for Opus 5.5 in Claude Code comes from turns, cache reads, output tokens and model price, and cutting usage can raise retry cost. At API list prices, the same sample session falls from $3.50 on Opus 5 to $2.40 on Opus 5.5, about 31% lower, with cache writes excluded from the sample. The estimate of 40% lower cost on typical workloads also includes fewer task tokens under default settings.

Why it matters: The Claude Code cost guide bills turns, cache and thinking output separately, and uses the same task to separate a lower unit price from lower task consumption.

claude.dev Blog

Claude Code guide: matching effort level to the task

Claude Code's effort setting controls how much compute the model spends on a task. Higher levels add autonomous judgment, verification and edge-case testing; lower levels suit fast interaction and iteration. Drawing on Opus 5.5 tests and Terminal-Bench 3.0 results, the author suggests having the model interview you to clarify requirements first, then implement and iterate at low effort, and verify at high effort. More effort usually cuts failures from missed edge cases, but it cannot fix a wrong approach.

Why it matters: A tested workflow that splits fast iteration and edge-case testing across different stages of a Claude Code task.

Sep 22Tuesday

claude.dev Blog

How to get the most out of Opus 5.5 in Claude and Claude Code

Anthropic's official guide covers using Opus 5.5 in Claude and Claude Code, spanning prompts, long-task control and result checks. It recommends stating clear completion criteria, setting in CLAUDE.md the conditions for continuing versus stopping to ask, keeping task lists in files, and checking the evidence sub-agents submit.

Why it matters: Anthropic's guide breaks long tasks into completion criteria, stop conditions and result checks, using file lists and sub-agent evidence to cut missed steps.

Jul 24Friday

claude.dev Blog

New context engineering rules for Claude 5 generation models

Anthropic describes its context engineering approach for the Claude 5 generation models, cutting the Claude Code system prompt for Claude Opus 5 and Claude Fable 5 by more than 80% with no measurable loss on coding evals.

Why it matters: Anthropic splits context management into lightweight repo notes, on-demand skills and specific reference files, a transferable way to cut repeated instructions and over-constraint.

Jun 3Wednesday

claude.dev Blog

Building Claude Code: how Anthropic uses skills

Anthropic shared lessons from hundreds of active skills inside Claude Code, grouping them into nine uses and covering how to write, distribute and measure them. A skill is a folder of instructions, scripts and resources. Write down common errors, load reference files on demand, reuse helper scripts, and phrase the description as a trigger condition. Small teams can share skills through a repo; at larger scale an internal plugin marketplace works, with a PreToolUse hook logging usage to find skills that under-trigger.

Why it matters: Anthropic sorts internal skill practice into nine uses and distills a team skill library from error logs, on-demand file reads and script reuse.

Jun 2Tuesday

claude.dev Blog

How Claude Code dynamic workflows tailor an execution framework per task

Claude Code shipped dynamic workflows last week. Claude writes an execution framework for the current task on the spot, and workflows can be saved, shared and reused. They coordinate subagents through agent(), parallel() and pipeline() in JavaScript, can pick a model and a separate worktree, and resume a session after an interruption.

Why it matters: Claude Code splits execution and adversarial verification across subagents with separate contexts, a concrete orchestration method for cutting self-preference and goal drift on complex tasks.

May 20Wednesday

claude.dev Blog

Using Claude Code: why HTML beats Markdown for complex output

Thariq Shihipar, a member of the Claude Code team, shares a method for generating complex output in HTML instead of Markdown, to improve reading, visualization, sharing and interaction. He exports interface tweaks as prompts so manual edits feed back into the agent workflow.

Why it matters: A Claude Code team member generates interactive documents in HTML and turns interface tweaks into prompts, feeding manual edits back into the agent workflow.

Apr 30Thursday

claude.dev Blog

Building Claude Code: designing agents around prompt caching

The Claude Code team shared engineering lessons on prompt caching, keeping the request prefix stable to cut latency and cost in long agent sessions. Put static content first, update state through messages, avoid switching models or adding and removing tools mid-session, and use defer_loading to load tools late. When compacting context, reuse the original session's system prompt, tool definitions and history, then append the compaction request so it hits the cache. The team also tracks cache hit rate and treats an abnormal drop as an incident.

Why it matters: The Claude Code team applies cache prefix stability to prompt ordering, tool loading and context compaction, a transferable caching method for long agent sessions.

Apr 10Friday

claude.dev Blog

Tools from the agent's point of view: how Claude Code designs tools

The Claude Code team reviewed its tool design, stressing that tools keep changing with model capability, agent goals and runtime environment. AskUserQuestion became a standalone tool for structured questions after two failed attempts, and Task replaced TodoWrite with task dependencies and shared updates across subagents. Context retrieval moved from pre-indexed RAG to Grep search and progressive disclosure; the Claude Code Guide subagent searches docs in its own context and returns only the answer to the main agent.

Why it matters: The Claude Code team uses question asking, task collaboration and context search as examples of treating observed model behavior as the input to tool design.

Mar 11Wednesday

Mistral AI

Mistral builds an agent on Vibe that writes Rails tests automatically

Mistral built an agent on its open-source coding assistant Vibe that writes Rails RSpec tests on its own. It reads source code, generates or improves tests, checks them against style rules and coverage targets, and runs unattended in CI/CD.

Why it matters: Mistral published its full method for building an auto-RSpec-test agent on Vibe, including transferable details on context engineering, skill files and custom tools.