Tới nội dung

Hình thức

Hướng dẫn thực hành Mới nhất

Nội dung thực hành dùng được ngay: mẹo viết prompt, xây dựng quy trình, cách dùng công cụ và kinh nghiệm từ những lần hỏng.

21 tin chọn11 trong 30 ngày quaĐã thu thập 199

Cập nhật 12 tháng 11 08:00

Hướng dẫn thực hành

12 tháng 11Thứ NămMục 1–20 trong 21
  1. vLLM Blog72

    vLLM serves DeepSeek-V4.1-Flash on Vera Rubin NVL72 at 58x GB300 NVL72 throughput

    vLLM reports serving DeepSeek-V4.1-Flash on a single NVIDIA Vera Rubin NVL72 rack at 1.7M tokens/s per GPU with 387 tokens/s per user, 58x the GB300 NVL72, peaking at 2.52M tokens/s and 1,810 tokens/s for a single user. The gains come from kernels rebuilt for Rubin's split memory domains and FP4 tensor cores, a specialized TP8 decode path, and rack-wide KV and Engram pools shared over NVLink. With prefixes pooled across the rack and Dynamo workload-aware routing, the prefix hit rate holds at 96.5-97% even at high concurrency.

    Vì sao đáng chú ý: vLLM's third DeepSeek-V4.1-Flash post details how Rubin-specific kernels and rack-wide memory pooling lift throughput 58x over GB300 NVL72, offering a concrete reference for large-scale serving optimization.

11 tháng 11Thứ Tư
  1. OpenRouter Announcements65

    OpenRouter measures the context cost of MCP, CLI and skills on one GitHub task

    OpenRouter measured the same GitHub task across the GitHub MCP server, the gh CLI and a gh skill on Claude Sonnet 5.5, GPT-6 Luna and Gemini 3.8 Flash. On Sonnet 5.5 all three answered correctly, but the CLI used 1,421 input tokens, the skill 3,315 and the MCP server 59,417 with its 46 default tools, whose definitions alone add 19,436 tokens per request. Scoping the server to two tools cut the task to 23,228 tokens, capping results at 4,000 characters brought it to 5,357, and Tool Search cut the up-front cost of 46 tools to under 600 tokens, though it halved the whole-task cost on Sonnet 5.5 while raising it on Claude Haiku 5.5.

    Vì sao đáng chú ý: OpenRouter's measured comparison shows how much context an MCP server, a CLI and a skill each consume on the same GitHub task, and where trimming tool results or deferring definitions changes the bill.

  2. OpenRouter Announcements60

    OpenRouter tests what fp8 and fp4 quantization do to LLM quality and price

    OpenRouter ran 280 exact-answer tasks against six endpoints of Qwen3.8 27B on October 20, 2026 and did not detect an accuracy difference between fp16, fp8 and fp4, with all six scoring between 98.3% and 99.2%. Across 1,455 endpoints, 59.5% report no precision at all and fp8 is the most common labeled format at 23.6%. Lower precision was cheaper for most models but not all, and provider choice moved price more than precision did, with the priciest same-precision endpoint a median 1.72 times the cheapest.

    Vì sao đáng chú ý: OpenRouter measured fp8, fp4 and fp16 endpoints of one model across 240 exact-answer tasks and found no detectable accuracy gap, while price differences tracked the provider more than the precision.

  3. LangChain Blog60

    LangChain explains how Deep Agents manages context without keeping everything in the window

    LangChain details how its open source Deep Agents harness uses middleware to fight context rot on long agent runs. Tool results over roughly 20,000 tokens and user messages over 50,000 tokens are offloaded to files with only a preview shown, while summarization condenses older messages once usage hits 85% of the context window. Skills, subagents and a markdown long-term memory file keep the main agent's window small and let memory carry over across models and providers.

    Vì sao đáng chú ý: LangChain explains how its open source Deep Agents harness uses middleware to offload, summarize and isolate context, offering a reusable pattern for agents on long tasks.

10 tháng 11Thứ Ba
  1. Hugging Face Blog60

    Hugging Face explains decision models and their support in Transformers and llama.cpp

    Hugging Face published an explainer on decision models, which take a state plus typed questions and return probabilities over predefined answers instead of generating text. The post details the choice, score and noul primitives, the confidence formula, and new support through Transformers' DecisionPipeline and AutoModelForDecision, a shared /v1/systemone endpoint in Transformers and llama.cpp, a Hub tag for discovery, and fine-tuning with the Trainer API. It notes that the same endpoint also works with compatible autoregressive LLMs by scoring answer logits in a batched forward pass.

    Vì sao đáng chú ý: Hugging Face explains how decision models turn classification into a typed, non-generative interface and shows how to serve and fine-tune them with Transformers and llama.cpp.

  2. OpenRouter Announcements72

    OpenRouter benchmarks 15 vision models and Mistral OCR on scanned documents

    OpenRouter tested 15 vision models, Mistral OCR and Tesseract on eight self-made degraded scans, and Claude Opus 5.5 was most accurate with 262 of 264 key values correct at $16.42 per 1,000 pages, while Mistral OCR was best value at 258 of 264 for $2.02. On a faded receipt, GPT-6 Luna and GPT-6.1 Sol invented card digits, a server name and wifi passwords, errors that character error rates do not reveal. The post includes code to transcribe scans, turn reasoning off, cross-check two models from different labs and run Mistral OCR through the file-parser plugin.

    Vì sao đáng chú ý: A hands-on benchmark of 15 vision models plus Mistral OCR and Tesseract on eight degraded scans, with runnable code for cross-checking two models to catch fabricated values.

9 tháng 11Thứ Hai
  1. Anthropic · Claude.dev60

    Anthropic guide on choosing among Claude Fable 5.2, Opus 5.5, Sonnet 5.5 and Haiku 5.5

    Anthropic published a guide recommending developers start most coding and knowledge work on Claude Opus 5.5 at its default medium effort level, since it costs half as much per token as Fable 5.2. The guide maps Sonnet 5.5 to well-scoped everyday work, Haiku 5.5 to high-volume classification, routing and subagent tasks, and Fable 5.2 to the hardest coding, finance and science projects or long unattended runs. It also advises adjusting effort level before switching models, and notes that on SWE-bench Pro, Opus 5.5 at low with failures re-run at high passed about 97% of tasks for about $0.17 each, against 95.3% for $0.29 with every task at high.

    Vì sao đáng chú ý: Anthropic's own guide maps each Claude model and effort level to concrete task types, with per-task cost and benchmark comparisons for choosing between them.

6 tháng 11Thứ Sáu
  1. Google Developers Blog60

    Google study finds stale TODO bug references distract coding agents in 13% of sessions

    ) references. Agents acted on the bug ID in 6% of encounters, and 83% of those actions were unprompted by the developer; 41% of unprompted detours chased already closed tickets while 57% hit still-open bugs and expanded scope. Controlling for developer, codebase complexity, task type, model tier and other factors, unprompted bug detours added 70% more steps, 77% more tokens and 105% more wait time per session.

    Vì sao đáng chú ý: Google's large-scale trace analysis shows stale TODO bug references pull coding agents into unprompted detours, and its cleanup agents removed thousands of obsolete references with over 95% reviewer acceptance.

  2. Anthropic Research60

    Anthropic explains the components of Claude agent memory

    Anthropic outlined how it designs memory for Claude agents, defining agent memory as knowledge retained across sessions and separating it from weight changes, in-context learning and compaction. 5 matched it with memory. 7 wrote 413 files across 225 folders and earned one badge.

    Vì sao đáng chú ý: Anthropic details how Claude's agent memory works, comparing models' note-taking on CL-Bench and Pokémon Red and showing how offline dreaming corrects and organizes stored notes.

20 tháng 10Thứ Ba
  1. Google Developers Blog60

    Google 将 TorchTitan 接入 TorchTPU，并发布 TPU 模型优化手册

    Google released a TPU fork of Meta's TorchTitan that runs PyTorch training on Cloud TPUs through TorchTPU's native torch.device("tpu") backend, with TPU-native kernels and a step-by-step optimization playbook. On TPU v7x (Ironwood), Llama 3 405B holds about 57% MFU from 128 to 1,024 chips, Qwen3 235B reaches 32.42% MFU on one 1,024-chip slice, and FLUX.1 reaches 19.4% MFU on 1,024 chips at 93% weak-scaling efficiency. The fork is at https://github.com/google-pytorch/torchtitan/, and a Transformers modeling backend plus an agent onboarding skill extend the same trainer to other checkpoints.

    Vì sao đáng chú ý: Google 与 Meta 合作把 TorchTitan 移植到 TPU，并给出从 FLUX.1 到 Qwen3 235B 的逐级调优路径与 MFU 数字，可作为 PyTorch 团队评估 TPU 训练可行性的参考。

15 tháng 10Thứ Năm
  1. Anthropic · Claude.dev62

    How the Claude Code team does code review

    The Claude Code team describes its layered code review flow, in which most reading is done by machines and human attention is reserved for risk-based approval. From June to September, merged pull requests to the Claude Code codebase more than doubled while the share that had to be reverted fell by more than 40 percent. The process stacks verification before the PR, required CI, Claude review, security review, risk-tiered human approval and a merge queue with flags and fast rollback, and the team turned past incidents into a daily eval set for the reviewer.

    Vì sao đáng chú ý: Anthropic's Claude Code team details how it moved most code reading to automated layers while merged PRs doubled and reverts fell over 40 percent, offering a transferable review model for agent-heavy teams.

8 tháng 10Thứ Năm
  1. Anthropic · Claude.dev62

    Anthropic walks through building scheduled agent automations with Claude Managed Agents

    Anthropic published a reference implementation for Claude Managed Agents (beta) that reads Slack channels and GitHub pull requests on a cron schedule, tracks changes since the last run via per-source bookmarks, and posts a brief to Slack. The walkthrough covers vault-held credentials, read-only access and a per-run spending cap, plus rules for reporting a failed source as unreadable rather than a quiet day and confirming a Slack post before updating bookmarks and the ledger. A Claude Code command with the claude-api skill configures the agent from the article's guidance.

    Vì sao đáng chú ý: Anthropic shares a reference implementation for scheduled Claude Managed Agents that read Slack and GitHub, with concrete rules for bookmarks, failed reads and post confirmation.

6 tháng 10Thứ Ba
  1. Anthropic · Claude.dev72

    Claude Code cloud sessions: a field guide to running tasks on separate VMs

    Addy Osmani walks through Claude Code cloud sessions, where each task runs on a fresh Anthropic-managed VM with the repository cloned onto its own branch, and cloud sessions are included in Pro, Max, Team and Enterprise plans at no extra compute charge. In a four-session test on a sample Node repository, three sessions started within 16 seconds of each other finished 87 seconds after the first began, and the flaky-test session ran the suite 40 times with zero failures. The guide covers setup scripts, network levels, GitHub connection paths and seven workflows suited to cloud sessions, while noting that local sessions remain the better fit for tasks needing local data, VPN services, GPUs or Zero Data Retention.

    Vì sao đáng chú ý: A hands-on field guide to Claude Code cloud sessions, showing how per-task VMs enable parallel work and where local sessions still fit better.

1 tháng 10Thứ Năm
  1. Anthropic · Claude.dev60

    Getting started with Claude Code mods: building a context-window forecast

    Claude Code mods are small JavaScript or TypeScript files that run inside a session as hooks shipped in plugins, and can observe events, rewrite behavior or draw custom UI in the terminal or desktop app. The guide builds Token Weather, an roughly 80-line mod that reads $.session.usage() after each turn and draws a weather-style context-window band above the prompt, then tours Blast Radius, which holds risky Bash commands, and Replay Theater, which steps through a turn's edits. Mods require Claude Code 2.1.287 or later and are on by default; finished code for all three is in anthropics/claude-code-playground.

    Vì sao đáng chú ý: A hands-on walkthrough of Claude Code mods that builds a working context-window meter in about 80 lines, showing how hooks, state and custom UI fit together.

28 tháng 9Thứ Hai
  1. Anthropic · Claude.dev72

    Anthropic adds build-eval and hillclimb commands to the claude-api skill in Claude Code

    Anthropic added two commands to the claude-api skill inside Claude Code: /claude-api build-eval builds an evaluation set in your codebase, and /claude-api hillclimb iterates on an application against that evaluation with a held-out set to catch overfitting.

    Vì sao đáng chú ý: Anthropic adds build-eval and hillclimb commands to its claude-api skill, showing how Claude Code can design evaluations and iterate against them while guarding against overfitting.

17 tháng 4Thứ Sáu
  1. OpenAI Developers Blog60

    OpenAI publishes API deployment checklist for GPT-6 workloads

    OpenAI published an API deployment checklist that walks developers through moving workloads onto the Responses API and the GPT-6 model family, including gpt-6-astra, gpt-6.1-sol, gpt-6 Sol and gpt-6 Luna. The guide covers reasoning.effort levels from low to max, text.verbosity, tool_search with deferred tool loading, Programmatic Tool Calling, Multi-agent, compaction, prompt caching and WebSocket mode, which OpenAI says has shown up to roughly 40% faster end-to-end execution for workflows with 20 or more tool calls. It also details reliability and safety steps such as handling 429 and 503 errors, background=True for long jobs, safety_identifier and misalignment monitoring.

    Vì sao đáng chú ý: OpenAI's official deployment checklist consolidates model selection, reasoning effort, tool search, compaction and WebSocket mode into one migration guide for GPT-6 workloads.

12 tháng 3Thứ Năm
  1. OpenAI Developers Blog67

    OpenAI updates Sora 2 prompting guide with character references, 20s clips and video extension

    OpenAI updated its Sora 2 prompting guide to cover new API capabilities: reusable character references for objects and animals, 1920×1080 and 1080×1920 exports, a maximum clip length raised from 12s to 20s, video extension that uses the full original clip as context, and a Batch API for asynchronous video jobs. The guide also walks through prompt structure for shot framing, camera motion, lighting and dialogue, and notes that resolution, duration and character IDs must be set as API parameters rather than described in prose. Extensions of up to 20 seconds are supported, and a video can be extended up to 6 times for a total of 120s.

    Vì sao đáng chú ý: OpenAI's updated Sora 2 prompting guide details new API capabilities such as character references, 20s clips and video extension, plus reusable prompt templates for controlling shots.

25 tháng 2Thứ Tư
  1. OpenAI Developers Blog62

    OpenAI publishes Codex prompting guide for gpt-5.3-codex

    OpenAI released a Codex prompting guide for developers calling the Codex-tuned gpt-5.3-codex model directly through the API, recommending medium reasoning effort for interactive coding and high or xhigh for the hardest tasks. The guide covers migrating an existing harness to the open-source codex-cli, including the standard Codex-Max prompt, the apply_patch, shell_command and update_plan tools, and first-class compaction via /compact in the Responses API for multi-hour sessions. It also describes the new phase field, supported only on gpt-5.3-codex, which marks assistant items as commentary or final_answer and must be preserved in history to avoid performance degradation.

    Vì sao đáng chú ý: OpenAI's official guide details how to tune prompts, tools and compaction for gpt-5.3-codex, giving teams a concrete migration path for their own coding agents.

26 tháng 9, 2025Thứ Sáu
  1. Modal Blog60

    Modal engineers reverse-engineer how Flash Attention 4 works

    Modal engineers read the released Flash Attention 4 source code and reverse-engineered how the Blackwell-optimized attention kernel works, since no official technical report exists yet. The write-up traces a tile through the kernel's warp-specialized asynchronous pipeline and details two math tricks: a cubic-polynomial software exponential that can replace SFU-based exponentiation, and a smarter online softmax rescaling rule that Tri Dao said cuts corrections by a factor of 10. FA4 is reported to be roughly 20% faster than Nvidia's closed-source cudnn attention kernels, and a backwards version was not yet available at the time of writing.

    Vì sao đáng chú ý: Modal engineers read the released Flash Attention 4 source and explain its warp-specialized pipeline and two math tricks, offering a transferable view of where CUDA kernel design is heading.

5 tháng 8, 2024Thứ Hai
  1. Modal Blog62

    Modal replicates Large Language Monkeys: LLaMA 3.1 8B matches GPT-4o on HumanEval via search

    Modal's Charles Frye replicated the core claim of the Large Language Monkeys paper on a new dataset, showing that repeated sampling lifts LLaMA 3.1 8B from 66.4% pass@1 to 90.5% pass@100 and 95.1% pass@1000 on HumanEval, versus GPT-4o's reported 90.2% pass@1. The runs used vLLM on Modal serverless GPUs and Modal Sandboxes for parallel test execution, cost well under $50, and showed smooth, predictable scaling across three orders of magnitude.

    Vì sao đáng chú ý: A hands-on replication shows how repeated sampling with a small open model can approach GPT-4o on HumanEval, offering a concrete, low-cost recipe for trading compute for accuracy.