> 2021年08月26日信息消化 ### システム監視の始め方・続け方 origin: Software Design 2021 2 ##### Obeservablity 可観測性 リアリティのあるモニタリングデータを取得するための取り込むのことです。 ##### APM (Application Performance Monitoring) APPのパフォーマンス（処理の所要時間、応答時間）を計測・監視する取り組みや手法のことを指します。 ##### RUM (Real User Monitoring) サービス利用者の利用者側で観測を行う取り組み、あるいはその手法のことを指します。（利用環境・利用端末） 監視システムの変遷（せん） | Year | Software | Desc | | ------------ | :---------------------------------- | ------------------------------------------------------------ | | 2000年代 | Nagios, Zabbix, MRTG, Cacti | OK / Non-OKをチェック | | 2000年代後半 | Ganglia | データ取集をPUSH型で行うメトリクス取集システム | | 2010年代 | Datadog, Sensu, New Relic, Mackerel | チェック志向ではなく、メトリクス志向の監視システム。SaaS | | 2010年代中盤 | Prometheus | レイヤや掌管領域を越えた総合基盤として認知されてきた。Obeservablity、APM、RUMなど、APPの内部状態を深く知る方法が整備・普及されてきた。 | > Avoid Blame and Keep It Constructive. > > -- Googleの「SREサイトライアビリティエンジニアリング」 #### Other Concepts 外形監視であるSyntheticsの監視、ログ管理、インシデント管理。 ### Weak Foundations Make Weak Models origin: [Weak Foundations Make Weak Models](https://info.deeplearning.ai/the-batch-invasion-of-the-large-language-models-can-ai-recognize-opioid-addicts-better-coffee-through-ai-1?ecid=ACsprvvSywto09WVMLQcbSJM3JG4oxVZDBXuuTsfQuyZ3By0Szo5h1ZfTQc6Q1kRnSo2lm5JRTg_&utm_campaign=The%20Batch) A new study examines a major strain of recent research: huge models pretrained on immense quantities of uncurated, unlabeled data and then fine-tuned on a smaller, curated corpus. The sprawling 200-page document evaluates the benefits and risks. 一项新的研究考察了最近研究的一个主要方向：在大量未经整理的、无标签的数据上预训练的巨大模型，然后在较小的、经过整理的语料库上进行微调。这份长达200页的文件评估了其益处和风险。 **What’s new:** Researchers at Stanford’s Human AI Institute [proposed](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgX3QW7VSmd82LSccfW4Bj35b31pw64W3XQfXh7St_xQV52s-X1Yv748W74gZ1t12hwkfVHQP725QZF6yW5Vp4l64YJTtYVlMCYW6nWGgpW4nrPPG2DkF4tW3_TXMn6NFGK0W4hlWwY4wtyscW7BgB8H7Z_-qvW1fpLf_3XRbzhN6rs48bf55dpW3zpKF41l7CbrW1VG1562j6Zh9W18XN1H5kQy_BW2GXf4B9dslMNW7wL9jv5l43LLW6ZgTrL4G7r1x3n0-1) ways to prevent large language models like [BERT](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgRg_W18PcXR6Z-xGfW4GzMT54dg-ZKW29Ky3V1kn3b4W5SZ3Hm1N4_BDW81wPTs3V0ZJ3N1g09KYMZX9PW2XBS4D8TfKmvW1sjymk5xwRgTW5mnRJ24MkMCYW8d0V0R9f309DW9032Kq52NZ18W16DGVC359X_vW8pM7kT1cYQt0W1d4LQy56s8MTW8CcL553DsWwcW9dSNH98H31RnW7nfJNt4c2-7hW2hdDjc2ZKTQGW4KnhzG8tjmyBW5J4NF97C0DVg3m1D1), [CLIP](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgJH1W8XtyT75HcWhWW85NmSW50bz3fW6DzJMm4GsMk3W2lCXYM4g2D3FW9gYf4C9cpP4JW23_psH8JtjB_W7c1zhR6_-MS7N5YHZKRPr7cSW59YZhv9km8RtW6wdHT-50D1glVZ7XYs7qPlP7W4JmkgY27YJZ5W3NQVY-6lXWr-W7d_LpG4L8J09W8NxdG822GSgxW2YPJDK1kwl0sW7dlZ7y5B1w2SV62gP0647tX6W1cw7_D6vQG-GW6c0sPN32VMKz3hvM1), and [GPT-3](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgJ-hVQdC_18DJlKjW1gk4tS2Ht9xgW58qjNw98fpfvW4t7txP1TXxsZW7430xY6rvwsPN2TjrKMCryC0N50hz2hH2f4tW7sWBtc6Q5HVlW8r19Bb1HtSHmW5dlVnw3N7y9tW8TyPD15Jd4XWN6WkBnpvSZs-W4RMr1X3TXYSRVyxr6D7G2XCGV2F_7g6yJQ0mW7BVX873tqR8rW56MVfg8vG4MpW2tmLLR43q32NW75fYf96n3yvBW5v9NN42Ws_Vx3mzG1) — which they call *foundation models* for their ability to support a plethora of high-performance, fine-tuned variations — from manifesting hidden flaws after fine-tuning. **Key insight:** The very factors that make large language models so valuable — unsupervised training followed by adaptation to a wide variety of tasks (indeed, some outside the domain of natural language) — make them potential vectors for harm. Defects in the foundation, such as **biases learned from uncurated training data**, can emerge in fine-tuned versions as challenges to fairness, ethical use, and legal compliance. Moreover, this approach encourages a technological monoculture in which a limited number of architectures, despite their strengths, proliferate their weaknesses across various domains. **关键见解：**使大型语言模型如此有价值的因素--无监督的训练，然后适应各种各样的任务（事实上，一些在自然语言领域之外）--使它们成为潜在的伤害载体。基础中的缺陷，如从未经整理的训练数据中学到的偏见，可能会在微调版本中出现，成为对公平性、道德使用和法律合规的挑战。此外，这种方法鼓励了技术的单一化，在这种情况下，有限的架构尽管有其优势，但其弱点却在各个领域中扩散。 **Toward solid foundations:** The authors recommend ways to minimize unwelcome surprises such as unwitting contributions to social or economic inequality, unemployment, or disinformation: - Develop metrics that predict ways in which a model may instill harmful behavior in its fine-tuned offspring and standardized ways to document these metrics, for instance [data sheets](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgTJPN6rmY8JYXXp5W1161dh1dq1skW3Z6Dm72tJ4FMW5_YvS_8sFMJwW8WrgRh7YNM1jW1v58JS5Dwnt4W1T4J4d3fdhdzW8Pp1CZ1Qyrg9W43klP71pkdRFW3j2_Rc64SQXjW3ytcLY6VK-2lW9bZfXV8mT5bhW99G6Cs9jkxWVW9fZx_P6cLqDNW5-b6pH3q7v84W79BPhb69MwHMW5DKCLY3bjTBNW86s3lz8my0CFW7FB5KV58B0-xW1pfk0j19HHkm3h0x1). **迈向坚实的基础：**作者建议采用一些方法来尽量减少不受欢迎的意外，例如在不知不觉中对社会或经济不平等、失业或虚假信息的贡献。 - Create incentives for companies that develop large-scale, unsupervised models to publicly test and audit their work. Warn developers of follow-on systems to vet them thoroughly for undesired behaviors prior to deployment. 制定指标，预测模型可能在其微调的后代中灌输有害行为的方式，并以标准化的方式记录这些指标，例如[数据表] - Counterbalance the power of deep-pocketed companies by making it easier for academic institutions and independent researchers to develop such models, for instance through a [National Research Cloud](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yr73q3nJV1-WJV7CgPJ8W6TZkDM6y9v5jW3_1smp28X1jLW154Sgs7fLBHqW2-jkHJ6cnQHQW2qCTy82_xCnnW325DVm727k0gW1fZRdD8GzgVmW29TJ-81NKfVNW7PHM2t15m-QDW5LvfLW2q6PLZW81l-h75sqHYsW1ghzf758lZtKN2Zk1hqs8jX5W8ThMF859cLkJW9kH2Ml3ZkKrkW5Yq-9b2mdMb-Vypb5l8XqrJlW89qXGb5w74RgW2j9jyj8N5ChDW3M2c_h8R7xFNW10xlYp26qjxwW8j63k_6xJDD2W3wYcz52bqVrVW3Sg8jK5pH13z3dZn1) and crowdsourced efforts to recreate GPT-style language models. 为开发大规模无监督模型的公司建立激励机制，公开测试和审计他们的工作。警告后续系统的开发者，在部署前要彻底审查其不希望出现的行为 **Behind the news:** The advent of BERT in 2018 accelerated adoption of unsupervised pretraining in natural language models and spawned ever-larger networks as researchers scaled up the concept and experimented with architectures. The approach has spun off fine-tuned models not only for language tasks like [conversation](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yrr3q3n_V1-WJV7Cg-8_W7vVvl-5mhxGdW4Z1NLX4PxzFWW7YDfc17jyKbyW3hVL7r3q7hSHW4NX-mH4GRr3tW6TDLR72_6BzmW6plnvN90-bLVW6NJxrN2MySSQW9896J31LH1xHV4zCyM2t0f4lW7WcvV-6PgL-rW6SYL-722DBvcW6vflT43JRsjTW36XZgd7C9YRKW78XwzK5LJrz1W1kJjnW1bX6HDW6-sC43173DZcVtfpd881b1yzW8jqZ4x78KPvRVPyMr98tBjtcVMVg082M3gnsW5yLcYD2Pf9nkW461p9C3745l4W7TLQDb5THscMW8HmpZW6vsBnSN5w1VMGdXGdJ3jg01), [image captioning](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgXvfW5s7JR97SF2q8W4FnMjT7zwgDDW7Hc8Dk804BlDW98s2Wz3LVS6VN1g_58F-cBkmW5GmH-q2B_W-rW6lYngS8FmN-PW4GC8kB9d3ShNW2HbZ_-3FWn25W5JTzfD5rGCSZW1K1ZPD3WkPrCW3c8Qbp2_gXfXW74zQwG4sYD4dW6q964s2WRVZdW3g9-wH7sdp-gW7gdY7N8tHgfQW5XmS_38n8VpfW8BcPMb89XgvyW6MwdMr8jMsg1W8Ybjr_43gvKF3cFm1), and [internet search](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yrr3q3n_V1-WJV7CgQ36W3rlvc35LPc-7W83Y8NG5NYT7QW3HFNh11V99L-W1hlRzB4v2m21W67b7463GsvC1W75g_5p8RX9J3W7QPpk63J_hK4W19s4QD2yGVjCW5MX8Jc89fzKtVYRWZz6MlxlhW3lgW_l65zrBRW7dmhmr3-lLn2W1ZHPkP4hJYBbW2NLN3H3QpL16N31n3gr7WFMBW7RCzW73q-MYBW8-H5ml671wJSW4KH-z_8Rp5k9N4V5zlTkML-7VYQRcq6J_byrW2YVH-x63WnknW5_PrK96ys367V6d1ML5SshxsW2BCjrr8lPVWtW7M6r638XSWl-N8bprl7P39W-38R71) but also far-flung applications including [modeling proteins](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yr73q3nJV1-WJV7CgLqNW4S_SNf6hshz6N3C_G1lF0303W5dj0Zz32MTZNW2kkDtp52D6NwW6YcwW82y14gLW67RMN794s05YW8Z9lyY3jHfbxW9d25WC49CwxGN6Cl5snFKkLLW1YrQB77ZTDNzW4gZrmY5hnf5TW35XpMy8LLrsXW5P22ZW2yPDPLW6TBsmp7SmVc6W6xnTTt3WDypzW586Mbg60KN_kW2rvk-b3gMGCgW6C43nM3XD09qW1B-K5D8-rdLLW90b48c4RPx91W4GGX0q8yv7bvVbXYPr5Rx7PMMPK5VFhZLdvW8nx36d2yY2QX3n8L1), [testing mathematical theorems](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yr73q3nJV1-WJV7CgQdqW3zJ4kD8sMy63W2lmgYB2tL53XW563sVg9hVwhYN2TCHZF4qtB2W8Mj2n_6RdkkHVm7wq54ZXWnqW2SS2d87f3_-5W3vKBDy5Pmlb0W2wKy4934vPpyW7HSCjx8cSdC_N628Zp0MKQ59N2CrQsyTXx2CW3b4RNp7gvZmqW4czVDP6Z_tXqW5B6kxq5CNMfGW7Mkldf5n39Y7W68P6sN8rjL5FW5GkXNm6bJ9SpW67kkdV2VH7BcW6qB7kf3MsrwlVrvfTW4YXP5LW8Bx_5z62TRh8Vj6Rtq3dgDbjW1c5yF231c0QM3hS01), [generating computer code](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgSsgW4MSy4z4GRjt9W1YY32_62Qp_jW186SB04nC5LVW2mgrXX3WPL2KVJ0gs33W0bSWW7nM9CQ33JDlkW3QB5Vz1hRrv5W3v6yvy3T-TpcW98W8vh27_DRsVZb1BW5gSRccW99LdLX1s0Y2WW7NQp1y7BY0sFW6zp6lk6Xvnx5W6XQkGM5dSdPVW7MNrRZ4nrk9dMlTmtq6cwBZW2s7gDd3sf8glW5KyhSX2kl3x4W7Wd6zg8cNZL2V75Mm47b00FZ3n951), [image recognition](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqS3q3npV1-WJV7CgDf4W95K6Yq39X6__W2XFPQl5SkDSgW8s8-G92yqphFW8X38D13b0fDsW3kPLBs3p2rbyW2gbZRK5-YyGYW2mn5j94Pb-7YVfwN9D5ZwL9zW46_4QT3Tfg8HW6ngBMS9gl9NKW97Q8T17g7kypW6MbBlK4HcpXhW7K6hXc4lhpB9W9khdf54kNwSZW5QsjjH2YlXGBMKkgHLH6vHRW1nSSGH8BtZ-WVzTWBh8SCkRhW8XK_hY58tKrZW8PbcmL8sjnScW4XQtmF8NMLylW5VwfJl75KL2D3chw1), [image generation](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqS3q3npV1-WJV7CgVXBW1c3g6g5W3wxBW27tnK93X4_9bN1TF4rjndtSsVFxGL92BRRqSW1plyTF72SQ0VW2y3vLT1QB055W16-Pnc2TvH0ZVVRy0X6zB4VXW7p77bB2RN2flW5l2Wrg8JDGdfW48sHx3619XpcW7F-XFB1FKsz6N34lHMd5xYcPW37nd--5MndQWW2WtHS52rCmxLW7vPH3b6w8kTPW37FDD62Kp_6TW5Kf4_61gpTRkMZPVyZKQZ6-W7dSPLj7n5ZQcW6XPdhT1-S70lW4cC3bW3lj1Ql3c_t1), and [reinforcement learning](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgLrTN1sxQN7cGpLMVW6zyr99Kf-4W5TF5dL3JXZpZW8vxbYW3h-SX3W5ynXVB6XRbG9W1zX3YH5c779bN8kzCGjl1thZW4WSwWw6mqKpkW6C8sMr4f851jW1j1KC71VCBWwW5Ss0_d4bLfhmW8f_7Xf3nMP2jW37z9p27bl_JSW3FDlnZ1TrLkdW56ysyt2v8-j0W767rp-2gK2ZsW2wdsYW3gYFLMW4sJ5Bt6PlhJmW1QDyc47qmDTbW8BxQkv3MlvNc3l2q1). **Why it matters:** Such models can cause harm due to intrinsic flaws by, say, propagating data-driven biases against members of particular [religions](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yqy3q3n5V1-WJV7CgGM9W1sqWbc728bHTVKpjY_1CGSZ5W6CFkN26q1kDKW5vzT5c4M_Zg7W94MqNj8wJfRPW1BXgPg8yYmGYW7S_Vhl1B22J8W7G0wNP62hSZ4W77JcsY3Jdk77W5NvfBn5QlGc3N5CrQwhD010NW39b6qx2L9mJ8W46yf8d67h5YtW1qwGFg7CKGksVDF8JK8lVH-SV9wNv_71KKR_W2-MFnx4g74y_W4SXGtZ79tgG7W2BlW6B41-3t8V1NHqL63YJvP395z1) or other groups) and extrinsic flaws, such as [energy-intensive training](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yr73q3nJV1-WJV7CgFZXW6V1rld6tX77wW5TCMGJ6yQbm0W2SsBJW1NGcR4W8xz6dz1B24qkW8xG8T197-rvmW8SbcBz4W0MsxW5gz4QN88y53YW89DVfl7FnztZW6rm4SL3NkLZsW8tRtk82LWdRVW8QzZdQ5TfbDkW2Kc2wn8CDtZxW85f4l718C9CrW1mv61T1RN9LLN8rSnJ3Sq21bW5TmPl41hcgNPVKfN_F2xCGHgW90nLf57rypwTW8wwFhZ1Xg6x4W2rj0cL5QW3CjW8C3Cwz74pt3gW27PhTT5j63bBW4fgTQH3bgm_ZW6PLr0n3JNnPq3ntq1) that leaves a large carbon footprint and misuse such as [propagating disinformation](https://info.deeplearning.ai/e3t/Btc/LX+113/cJhC404/VXjMWX2Y5ySCN55k5w2T0l6bW8mS6hr4wH5v7N1h3yrr3q3n_V1-WJV7CgRHjW6-2ljS6tt9m5VNrQb68jQSkTW4KbCv_6PhNy4W4VBj862Y-2LMN42qX-fJn7YlW6sHx6D3RdR4CW1g5SCS2NYLtbW6Vx5RH97KPmgW6Spqrd4t3_hVW2htKCL9l304pW16qTFm44d1-CW8Xq1LL5nbD4DN17twv8jyRTDTJwX-6j9mbxVK4d6C1BwjMvW1wh2nJ12vDR7W5VPK-T4Hrl9NW8fWKcj7sS8fmW3-9c4s65ZbxCW5GSJDC1LkHJjW5Lq8M51Rs4CRW12qcc41nJbSbW3CFBv91Fqwp7W4mbYS_5mLqmJW64BFyL8SMw0GVdDhy95yB4sr3dzn1). Deep learning systems developed without foresight run the risk of becoming a burden rather than a boon. **它为什么重要：**这种模型可能由于内在的缺陷而造成伤害，例如，传播数据驱动的对特定[宗教]成员的偏见或其他团体）和外在的缺陷。如[能量密集型训练]留下大量的碳足迹和滥用，如[传播虚假信息] 在没有远见的情况下开发的深度学习系统有可能成为一种负担，而不是一种福音。 **We’re thinking:** The future of AI may well be built on a limited variety of foundation models. In any case, the painstaking work of checking models for flaws beats cleaning up messes caused by neglecting to do so. **我们在想：**人工智能的未来很可能是建立在有限的各种基础模型之上。在任何情况下，检查模型缺陷的艰苦工作胜过清理因疏忽而造成的混乱。 ### 5 Visual Mathematical Proofs origin: [5 Visual Mathematical Proofs](https://medium.com/math-simplified/5-visual-mathematical-proofs-7e085397c01d)   Moving on to another equation representing the sum of first *n* natural numbers     We are going to shift gears from algebra to calculus now, but don’t worry it still remains quite simple. Let’s take a lot at the formula of “An Integral of a sum of reciprocal powers”. An example of reciprocal powers might be *x^(3/2)* and *x^(2/3) —* Basically staying with the same base and inverting the power. The statement suggests that   The last proof I wanted to showcase is Young’s Inequality which states that if two functions *phi* and *psi* are continuous, vanishing at the origin, strictly increasing and inverse to each other then the following inequality holds for **a,b greater or equal to 0**  The above statement would be equality if and only if b=g(a). And yet again, with a simple illustration, we get an elegant proof  ### Misc - 脆弱性：第三者が悪用可能なバグ - 「CSE-2019-5428」は2019 年にcURLで発見・修正された：ヒープバッファオーバーフローの脆弱性に当たります。