2b:null 2c:["$","title",null,{"children":"Playing with GPT-3 summarization | Tanner Hobson | Observable"}] 31:Ta09,{ return EMBED(SUMMARY(` We propose to develop a cloud-based interactive visualization service that allows users to explore and analyze data in a more intuitive and interactive way. This would make it easier for them to find insights and make decisions. `, SUMMARY(` A community data repository is a repository of datasets that are created, curated, and shared by the community that uses them. We propose to extend cloud-based functionalities of a data repository beyond data services by providing interactive visualization services. This would allow users to explore and analyze data in a more intuitive and interactive way, making it easier for them to find insights and make decisions. `, SUMMARY(` A community data repository is a repository of datasets that are created, curated, and shared by the community that uses them. `, ` SCIENTISTS organize their research around data. When a research community starts to continuously accumulate, curate and share community datasets, community data repositories become a catalyst of future research. They are also a powerful source to engage the public, make science relevant, and deepen societal impact of science. Such continuously growing and open datasets are precious assets of the whole world. As an example, the data used in this work is the Climate Forecast System (CFS) ensemble data repository [52] from NOAA National Centers for Environmental Prediction (NOAA NCEP) that captures annual global atmospheric patterns at spatial 0.5×0.5 degree precision and a temporal resolution of 6 hours. `), SUMMARY(` We propose to extend cloud-based functionalities of a data repository beyond data services by providing interactive visualization services. This would allow users to explore and analyze data in a more intuitive and interactive way, making it easier for them to find insights and make decisions. `, ` While cloud has already become the leading solution for creating distributed systems to support large data repositories, in this work, we propose to extend cloud-based functionalities of a data repository beyond data services. Our primary proposal suggests interactive visualization services can become a component of cloud-based data repositories too. `)))) function EMBED(x) { return htl.html`

${x}`; } function SUMMARY(output, ...inputs) { return htl.html.fragment`

${output}

${inputs.map((d, i) => htl.html.fragment`

${i+1}:${d}`)}

`; } }32:T144b,$0 = { const $0 = `\ SCIENTISTS organize their research around data. When a research community starts to continuously accumulate, curate and share community datasets, community data repositories become a catalyst of future research. They are also a powerful source to engage the public, make science relevant, and deepen societal impact of science. Such continuously growing and open datasets are precious assets of the whole world. As an example, the data used in this work is the Climate Forecast System (CFS) ensemble data repository [52] from NOAA National Centers for Environmental Prediction (NOAA NCEP) that captures annual global atmospheric patterns at spatial 0.5×0.5 degree precision and a temporal resolution of 6 hours. While cloud has already become the leading solution for creating distributed systems to support large data repositories, in this work, we propose to extend cloud-based functionalities of a data repository beyond data services. Our primary proposal suggests interactive visualization services can become a component of cloud-based data repositories too. Such a broadened scope of data repositories can help lower many barriers of adoption by diverse user communities. For example, researchers can accelerate their work by being able to always see the latest data to test and refine their hypotheses without needing to maintain an up-to-date local replica of the entire dataset. They can also collaborate with more people, by being able to share their findings with others who do not have, or cannot afford to have, an entire local copy of the data. Recently, interactive volume visualization as a service has been shown to be feasible [49], [50]. In this work, we focus on creating interactive parallel flow visualization as a service, because flow visualization is important to many disciplines, including atmosphere, ocean, fusion, petroleum, aerodynamics, and cardiovascular biophysics, where “seeing” the flow is often the first step of scientific research. In addition, interactive parallel flow visualization offers a unique opportunity to study how to use heterogeneous cloud resources to achieve consistent parallel accelerations in support of synchronous user interactions. Cloud resources are a promising alternative to traditional HPC computing and visualization resources that require scientists to have a working relationship with supercomputing centers. Due to this reason, our work focuses on using cloud platforms to make leading-edge scientific datasets interactively usable at an incremental cost. Our prototype system is called Visualization Cloud Instances (VCIs). VCIs work collaboratively as a self-organizing swarm for parallel computing. Each swarm appears as a single cloud service, i.e. a VCI Service, which can be used locally on an institutional cluster, or remotely on a public cloud such as Amazon AWS. Our results show that the VCI approach is able to support large flow data and ensembles of data and still maintain crucial cloud-based characteristics: (i) built for large flow data and ensembles of flow data; (ii) achieves fast interactivity; (iii) instantaneously available; (iv) serves multiple users concurrently; (v) serves users locally and remotely; (vi) supports a variety of user devices, and (vii) lightweight to integrate into applications. Desktop applications can use VCI Service through a JavaScript library, vci.js, which transparently manages parallelism, performance, and fault-tolerance. By hosting NOAA NCEP CFS data in the cloud, we show (1) an application that can provide interactive visualization of 3D global atmospheric flow field to students (Figure 1); (2) a comparative visualization for scientists to analyze deviations between forecast vs. observed ground truth (Figure 8); and (3) a public-awareness application that integrates a VCI Service into the popular D3 [1] library. This last example enables any citizen to investigate how pollution from nuclear power plants in the United States, on any particular day in the year, can impact the entire globe (Figure 9). All applications use a year-long observation dataset of the Earth’s 3D atmospheric flow (1,463 timesteps, 720 × 361 × 36 spatial resolution, 150 GB) from the NCEP CFS repository [52]. The second application adds a corresponding forecast dataset of the same dimensionality from the same CFS repository. In all cases, the AWS setup required is less than $0.70/hour. Our results suggest cloud services, like the VCI Service, can conduct parallel computing using heterogeneous resources and support flexible, interactive, general use of large datasets on a desktop. These interactive use cases, coupled with efforts to quantify the exact cost-performance benefits of the cloud [22], expand the application potential of cloud hosted data resources. All software of VCI will be open-sourced upon publication of the paper. The remainder of this paper is organized as background in Section 2, system architecture in Section 3, and application development in Section 4. We show results in Section 5 and discuss conclusion, and future works in Section 6.`; return $0.split(/\n\n/).map((s) => s.replaceAll(/\s+/g, ' ')); }33:[] 2f:{"id":"26ca77e46aba6d64","slug":null,"title":"Playing with GPT-3 summarization","thumbnail":"82d16480b5f61f94ae055bed0a7a47a1c5b905c09491546565572c9ff5b7e6df","access_level":"public","create_time":"$D2023-02-20T19:36:54.060Z","update_time":"$D2023-02-21T22:30:45.565Z","view_time":null,"first_public_time":"$D2023-02-20T19:36:54.060Z","fork":false,"stars":0,"starred":false,"owner":{"type":"team","id":"7343f790779e6735","login":"player1537","name":"Tanner Hobson","avatar_url":"https://avatars-next.observableusercontent.com/16170a1a32e364e1ff6054c6045de6814a032d01331aba01288f82afceeaaec2","tier":"starter_2024"},"version":42,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"e0a627f6b6f3beb0","login":"player1537","name":"Tanner Hobson","avatar_url":"https://avatars-next.observableusercontent.com/16170a1a32e364e1ff6054c6045de6814a032d01331aba01288f82afceeaaec2","tier":null}],"members":[{"type":"individual","id":"e0a627f6b6f3beb0","login":"player1537","name":"Tanner Hobson","avatar_url":"https://avatars-next.observableusercontent.com/16170a1a32e364e1ff6054c6045de6814a032d01331aba01288f82afceeaaec2","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"lapsed":null,"subscribed":null,"forks":1,"collection":null,"type":"notebook","body":{"title":"Playing with GPT-3 summarization","cells":[{"id":0,"value":"# Playing with GPT-3 summarization","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":4,"value":"$31","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":24,"value":"$32","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":28,"value":"$$1 = {\n let $1 = '';\n for (const paragraph of $0) {\n const summary = await summarize(`${paragraph}\\n1 sentence summary:`);\n $1 += `\\n${STRIP(summary)}`;\n }\n\n return STRIP($1).split(/\\n\\n/).map((s) => s.replaceAll(/\\s+/g, ' '));\n \n function STRIP(s) {\n return s.replace(/^\\s+/, '').replace(/\\s+$/, '');\n }\n}","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":38,"value":"summarize(`${1[0]}\\n1 sentence summary:`);","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":18,"value":"{\n return htl.html`\n

${SUMMARIZE(\n summarize(`SCIENTISTS organize their research around data. When a research community starts to continuously accumulate, curate and share community datasets, community data repositories become a catalyst of future research. They are also a powerful source to engage the public, make science relevant, and deepen societal impact of science. Such continuously growing and open datasets are precious assets of the whole world. As an example, the data used in this work is the Climate Forecast System (CFS) ensemble data repository [52] from NOAA National Centers for Environmental Prediction (NOAA NCEP) that captures annual global atmospheric patterns at spatial 0.5×0.5 degree precision and a temporal resolution of 6 hours.`);\n function SUMMARIZE(...inputs) {\n return inputs.map((input) => htl.html.fragment`\n

${STRIP(summarize(input))}\n }\n\n function STRIP(s) {\n return s.replace(/^\\s+/, '').replace(/\\s+$/, '');\n }\n}","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":15,"value":"summarize = {\n return;\n const history = new Map();\n return async function summarize(prompt) {\n if (history.has(prompt)) {\n return Promise.resolve(history.get(prompt));\n }\n\n const response = await fetch(`https://api.openai.com/v1/completions`, {\n method: 'POST',\n headers: {\n 'Content-Type': 'application/json',\n 'Authorization': 'Bearer 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',\n },\n body: JSON.stringify({\n prompt: `${prompt}\\nTL;DR:`,\n model: 'text-curie-001',\n max_tokens: 256,\n temperature: 0.2,\n }),\n });\n\n const json = await response.json();\n console.log({ json });\n\n const completion = json.choices[0].text;\n history.set(prompt, completion);\n\n return completion;\n };\n}","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W33","stdlib":"1","titling":"auto"}} 30:"$2f:body"