BMI
Tổng CTCP Bảo Minh (BMI) đã công bố KQKD nửa đầu năm 2024 với tổng phí bảo hiểm gốc đạt 3,5 nghìn tỷ đồng (+17% YoY) và lợi nhuận trước thuế (LNTT) đ...
2019
Mar19: Phú YênJul19: Nghệ An - biển Quỳnh3/1/2019: pickup KC bring books. Read Facebook even news isn’t interesting because I think I won’t read late...
TV2
SongHau SH2PTV2 công bố KQKD quý 2/2024 với doanh thu đạt 362 tỷ đồng (+150% YoY) và LNST sau lợi ích CĐTS báo cáo đạt 12 tỷ đồng (-11% YoY). Trong n...
BMI
Tổng CTCP Bảo Minh (BMI) đã công bố KQKD nửa đầu năm 2024 với tổng phí bảo hiểm gốc đạt 3,5 nghìn tỷ đồng (+17% YoY) và lợi nhuận trước thuế (LNTT) đ...
2019
Mar19: Phú YênJul19: Nghệ An - biển Quỳnh3/1/2019: pickup KC bring books. Read Facebook even news isn’t interesting because I think I won’t read late...
TV2
SongHau SH2PTV2 công bố KQKD quý 2/2024 với doanh thu đạt 362 tỷ đồng (+150% YoY) và LNST sau lợi ích CĐTS báo cáo đạt 12 tỷ đồng (-11% YoY). Trong n...
Share Dialog
Share Dialog
Subscribe to RetoK
Subscribe to RetoK
<100 subscribers
<100 subscribers
I use ChatGPT for email, chat, how to response: both for English, structure, EQ…
KC use ChatGPT for her science homework
========================


Do you know what is the underlying reason that our AMF doesn’t support for ANR? In my opinion, ANR is important feature and all tier-1 operators request for it.
ChatGPT Do you know the underlying reason why our AMF doesn’t support ANR? In my opinion, ANR is an important feature, and all tier-1 operators are requesting it.
Rephrase: Hi George, Our customer wants to know why NSN does not support Event Exposure with bulk subscription? Can you provide some details? And for the RFP clause “The proposed system shall support to interwork with NEF for exposure network capability in comply with CAMARA project”, what should they change for Nokia to FC?
ChatGPT: Hi George, Our customer is asking why NSN does not support Event Exposure with bulk subscriptions. Could you provide some details? Additionally, regarding the RFP clause, "The proposed system shall support interworking with NEF for exposure network capability in compliance with the CAMARA project," what modifications should they make for Nokia to provide FC?
rephrase: Hi Magnus, I fell off motorbike last week and had an operation. Now although it finished, I still walk on a clutch. Therefore I mind going to the customer’s office with a clutch. Truong may accompany with you to the customer. In my opinion, there are some topics the customer may need us consult - Mobifone container terms, concepts, rfp - Viettel request for 5G call flow last time. Also recently they show interest for VoWiFi, FWA solutions: is it worth to invest, experiences from other operators, how to deploy…
ChatGPT: Hi Magnus, I had a motorbike accident last week and had to undergo surgery. Although the surgery is over, I'm still using crutches to walk. As a result, I don't feel comfortable visiting the customer's office. Truong can accompany you to the customer instead. In my opinion, there are a few topics the customer may need our consultation on: Mobifone container terms, concepts, and RFP. Viettel's request for 5G call flow information. Additionally, they have recently shown interest in VoWiFi and FWA solutions: whether it’s worth investing in, experiences from other operators, and how to deploy them.
rephrase: We are exploring the possibility of offering NCP, instead of NCS. Can you check in your team how it impacts on Delivery and Care services? => We are considering the option of providing NCP instead of NCS. Could you please inquire within your team about the potential impact on Delivery and Care services?
rephrase: Thank you team for your comments I consolidate it in the slide in the following links. Please take a look, provide your feedback and edition. => ChatGPT: "I appreciate your input, team. I've compiled it into the presentation linked below. Please review it, share your feedback, and suggest any necessary edits."
rephrase: "Next actions: Mobifone will provide the final scope based on the vendor’s recommendations. Nokia will do re-dimensioning and provide the updated BoQ Mobifone and Nokia work together for the Mobifone’s solution roadmap of next phases" => ChatGPT: "Upcoming steps: Mobifone is set to determine the definitive scope following the recommendations of the vendor. Nokia will proceed with re-dimensioning and furnish an updated Bill of Quantities (BoQ). Mobifone and Nokia will collaborate to formulate Mobifone's solution roadmap for the forthcoming phases."
rephrase: "Nokia propose to have site 2 in Hanoi for geo-redundancy. Site 2 is for critical network elements like SDM, PCF, 5G charging mediation nodes We should not have site 2 in HCM city because It makes a real geo-redundancy since both sites are at the same city If in the future, Mobifone want to deploy a new site in Hanoi for these nodes, we have to redeploy. So it is better to make a decision upfront" => ChatGPT: "Nokia suggests establishing Site 2 in Hanoi to ensure geo-redundancy. Site 2 is designated for housing critical network components such as SDM, PCF, and 5G charging mediation nodes. It is not advisable to position Site 2 in Ho Chi Minh City for the following reasons:
It does not provide genuine geo-redundancy since both sites are located in the same city.
In the event that Mobifone decides to deploy a new site in Hanoi for these nodes in the future, it would necessitate a redeployment. Therefore, it is more prudent to make this decision proactively."
rephrase: Accordign to the standard N 5G traffic profile, Average bit rate per UE/BH is 700 Kbps. So with the scope of 1mil subscribers, our optimistic estimation is the total throughput of 700 Kpbs*1mil = 700 Gbps. However with N’s experiences in other networks, the average bit rate per SAU from the early 5G SA deployment varies from 200 to 500kbps. So we can estimate that the total throughput is around 200 Gbps – 500 Gbps. Nokia recommend we shoud start small with 200 Gbps in the first phase and then expand later in the next phase (with other UPF nodes in site Hanoi). => ChatGPT: "Following the standard N 5G traffic profile, the average bit rate per UE/BH is 700 Kbps. Considering a subscriber base of 1 million, our optimistic estimate for the total throughput would be 700 Gbps. However, based on Nokia's experience in other networks, the average bit rate per SAU in early 5G SA deployments has ranged from 200 to 500 Kbps. Therefore, we can estimate the total throughput to be approximately in the range of 200 Gbps to 500 Gbps. In light of this, Nokia recommends that we initiate the first phase with a conservative estimate of 200 Gbps and consider expansion in subsequent phases, potentially with the addition of other UPF nodes in the Hanoi site."
rephrase: Need to check HW capacity, license in the existing Mobifone provisioning system for a new 1 millions 5G subscribers. The existing provisioning system need to integrate with new 5G network elements such as UDM, AUSF, PCF...=> ChatGPT
"We should conduct a capacity and licensing assessment within Mobifone's current provisioning system to accommodate a new influx of 1 million 5G subscribers. Furthermore, the existing provisioning system will require integration with the new 5G network elements, including UDM, AUSF, PCF, among others."

repharse: Mobifone can deploy SDM overlay solution during this intermediate period from 4G to 5G SA while waiting for a new modernization project of the existing SDM system Sim change is not mandatory
ChatGPT
"Mobifone has the option to implement an SDM overlay solution during the transitional phase between 4G and 5G SA, as we await the commencement of a forthcoming modernization project for the current SDM system. It's important to note that sim card replacement is not obligatory in this context."

rephrase: Except the CHF use case, Nokia recommend to have other optional use cases in the 2nd phase deployment
ChatGPT
"With the exception of the CHF use case, Nokia suggests incorporating additional optional use cases in the second phase of deployment."

rephrase: For SMS in 5G, there are 2 options:SMS via IP-SM-GW and SMS via SMSF. Nokia recommend to have SMSF in the 2nd phase, when mMTC use cases are mature enough. EIR is optional. Nokia recommend to deploy it in the next phase SCP: mandatory for implementing SBA Model C and D. Nokia strongly recommend to deploy it in the 1st phase because deploying in the next phases require changing in network architecture, signaling rerouting, then create many opex expenses for reconfiguration and reintergration NRF: mandatory for implementing SBA Model B, C and D. Nokia strongly recommend to deploy it in the 1st phase. Model A is not recommended for production BSF: optional. Nokia recommend to have SEPP in the 2nd phase when having multiple PCF instances which require session binding SEPP: optional if no roaming. Nokia recommend to have SEPP in the 2nd phase when Mobifone have 5G roaming agreement with roaming partners and also when the customers demand high 5G data throughput. In the first phase, 5G subscriber can use VoLTE and 4G data roaming instead. NEF: optional. It is for API gateway/portal and plugins for 3rd application and developers. Nokia recommend to have NEF in the 2nd phase with use cases such as Monitoring Location_Report Event
ChatGPT
Here is the information rephrased:
Regarding SMS in the 5G context, there are two options: SMS via IP-SM-GW and SMS via SMSF. Nokia suggests implementing SMSF in the second phase, once the mMTC use cases are sufficiently developed.
EIR is an optional component, and Nokia recommends deploying it in a subsequent phase.
SCP is mandatory for implementing SBA Models C and D. Nokia strongly advises deploying it in the first phase since introducing it in later phases would necessitate changes in network architecture, signaling rerouting, resulting in increased operational expenses for reconfiguration and reintegration.
NRF is mandatory for implementing SBA Models B, C, and D. Nokia highly recommends deploying it in the first phase. Model A is not recommended for production.
BSF is optional. Nokia suggests considering SEPP in the second phase, especially when multiple PCF instances requiring session binding are in use.
SEPP is optional unless there is roaming involved. Nokia recommends introducing SEPP in the second phase when Mobifone has 5G roaming agreements with partners and when customers demand high 5G data throughput. In the first phase, 5G subscribers can use VoLTE and 4G data roaming.
NEF is optional and serves as an API gateway/portal with plugins for third-party applications and developers. Nokia recommends implementing NEF in the second phase, particularly for use cases like Monitoring Location_Report Event.
rephrase: We will have a Mbf 5G SA trial workshop with customer on Wed 1Nov2023. It is an important meeting because their depute director will lead the meeting. So please help to join to support me in case I need for technical questions => ChatGPT: On Wednesday, November 1st, 2023, we are scheduled to conduct a trial workshop on Mobifone's 5G SA with the customer. This meeting holds significant importance as it will be led by their Deputy Director. Therefore, I kindly request your presence to assist me in addressing any technical queries that may arise.
rephrase: There is a new update of scope from the customer. Can you help to update the NPT? => ChatGPT: The customer has provided a new scope update. Can you assist in revising the NPT accordingly?
rephrase: In my opinion, optimized solution 1 is most likely accepted by the customer => ChatGPT: From my perspective, it's highly probable that the customer will accept optimized solution 1.
rephrase: To prepare for the customer discussion of scope optimization, the customer may want to know how much cost/price they can save if they take out this scope. So can you help to prepare some number for it: => ChatGPT: In anticipation of the customer's discussion regarding scope optimization, they might be interested in understanding the potential cost savings associated with removing certain elements from the scope. Could you assist in generating some cost estimates for this purpose?
rephrase: Career: I should not only focus on knowledge (try to read more) but skills: know how to do (working order, what need to do…), process, networking (who to ask). In the past, I was used to only follow knowledge, only try to read, to become expert, know more or better than people, then show off. But starting from TPM, I moved to high level approach. When I moved to PM role, I focused less on knowledge, spent some time for knowledge. => ChatGPT: Career-wise, it's important for me to not solely emphasize acquiring knowledge (such as reading more), but also to develop skills encompassing practical know-how (how to execute tasks effectively, understanding processes), building a network (connecting with people who can provide guidance), and cultivating a holistic approach. In the past, my approach was primarily centered on accumulating knowledge. I aimed to read extensively, become an expert, and outshine others in terms of knowledge, often showcasing my expertise. However, my perspective shifted towards a more high-level approach when I transitioned into a TPM role. As I moved into a PM role, I began to place less emphasis on accumulating knowledge and allocated more time to skill development and knowledge maintenance.
encompass /ɪnˈkʌm.pəs/: to include different types of things:
The festival is to encompass everything from music, theatre, and ballet to literature, cinema, and the visual arts.
rephrase: Hi, We are preparing to submit the quotation. In our solution proposal, we proposed 12 failover servers . However, to compete with our main competitor, we are considering to optimize the number of servers. I would like to reduce the number of failover server (to 6 servers or even 2 failover server only). Please help to share your view. => ChatGPT: Hello, We are in the process of finalizing our quotation. In our solution proposal, we initially suggested using 12 failover servers. However, in order to stay competitive with our main rival, we are contemplating the optimization of server numbers. I am interested in reducing the failover server count, possibly down to 6 servers or even as low as 2. I would appreciate your view on this matter.
E// already has 5G SA trial phase 2 with Viettel. Viettel has no plan for 5G trial with other vendor which make us to have to propose new thing for trial => ChatGPT: "Ericsson (E//) has already conducted the second phase of 5G Standalone (SA) trials with Viettel. As Viettel has no plans for 5G trials with other vendors, we need to propose new ideas for further trial phases."
Huawei SDM’s swapping bid is delayed. The customer decided to expand Huawei SDM’s capacity instead => ChatGPT. The swapping bid for Huawei SDM has been postponed. Instead, the customer has chosen to increase the capacity of Huawei SDM.
rephrase: Regarding to RFP/SOC, we complied the following. But it only mentions about 05 year of life cycle and repair/return; not mentions about the right of 06 year of supply of spare parts => ChatGPT: In reference to the RFP/SOC, we have compiled the following information. However, it solely addresses a five-year lifecycle and repair/return aspects, without specifying the entitlement to a six-year supply of spare parts.
rephrase: "Our ATCA products are reaching limited HW manufacturing stage we cannot commit for the request of supply of equivalent spare parts for a period of 5 years. However Nokia commit to provide repair services during 5 years services under warranty and/or separate care agreement” => ChatGPT: "While our ATCA products are progressing towards the limited hardware manufacturing phase, we are unable to guarantee the availability of equivalent spare parts for a period of five years. Nonetheless, Nokia is committed to offering repair services throughout this five-year period, both under warranty and through separate care agreements."
dịch sang tiếng Anh: thuê bao 5G có thể nằm phân tán trên các dải đầu số khác nhau; trong khi DRA chỉ có thể định tuyến nông theo dải đầu số. DRA không thể định tuyến sâu đến mức nguyên cả số msisdn => ChatGPT: "5G subscribers can be distributed across different number ranges, while the DRA (Diameter Routing Agent) can only route traffic based on number ranges. The DRA cannot route down to the level of individual MSISDN numbers."
rephrase: DRA route base on msisdn/imsi range. But 5G subscribers are random, may belong to different ranges. So how DRA is able to route 5G traffic to New SDM system? DRA cannot route for a specific msisdn/imsi number. It has to route based on a specific range: such as 09031 => ChatGPT: The DRA (Diameter Routing Agent) traditionally routes traffic based on the MSISDN/IMSI range. However, with 5G subscribers, the challenge arises because they can belong to various ranges, making it difficult to route based on specific individual numbers (MSISDN/IMSI). Instead, the DRA must adapt to the randomness of 5G subscriber assignments and route traffic effectively to the new SDM (Subscriber Data Management) system. In this context, the DRA cannot route traffic for a particular MSISDN/IMSI number; it must operate based on broader criteria, such as specific range identifiers like "09031."
Rain so could not go out to take photo or visit. Had a cold breakfast. Worried for my trip, may it completely ruin. We may have to stay at homestay the whole day. People’s mood isn’t good. But my plan was still good when I decided to stay in Trung Luong cafe shop in the morning. When it stopped raining, I went out to take photo so encourage people to take photos too. Then had a good decision to visit Trung Luong camping area where children can play and we had some nice photos => Due to the rain, we couldn't venture out to take photos or go sightseeing. We settled for a cold breakfast and I grew anxious about how the weather might disrupt my trip. There was a possibility that we'd be confined to our homestay for the entire day, and everyone's mood was rather somber. However, my initial plan to spend the morning at Trung Luong cafe proved fortuitous. As soon as the rain subsided, I seized the opportunity to head outside and capture some photos, then inspire others to do the same. Eventually, I made the wise decision to explore the Trung Luong camping area, where children could enjoy themselves, and we could snap some beautiful pictures.
rephrase: Minh’s iphone doesn’t have high resolution so “nước ảnh không trong” => Minh's iPhone lacks high resolution, so the photos don't come out clear.
rephrase: This picture is funny when there is a contrary between a lamp’s face and Minh Khanh’s face => ChatGPT: "This picture becomes humorous due to the contrast between the expression on the lamp's face and Minh Khanh's expression."
rephrase: Look at Khanh An’s face: always funny and happy => ChatGPT: "Observe Khanh An's expression; it's consistently cheerful and full of joy."
rephrase: Leaving airport for Trung Luong. It rained drizzling. Reaching to homestay in a small fishing village. It is not comfortable. I felt worried. Only satisfied that it locates near the beach with nice sea view. Put luggage at homestay and move to Trung Luong’s cafe. => ChatGPT: "Departing from the airport en route to Trung Luong, we encountered a light drizzle. Upon arriving at a homestay in a quaint fishing village, we found it to be less than comfortable, which left me feeling somewhat concerned. The only consolation was its proximity to the beach, offering a beautiful sea view. After dropping off our luggage at the homestay, we made our way to Trung Luong's cafe."
rephrase: After leaving Trung Luong’s cafe, we headed to Cau Ut’s restaurant for lunch. Delicious lunh with hop pot of crab. People were satisfied and happy. Reaching to Nha cua ba’s homestay. After a long nap, I woke up and urged people for swimming. Trung Luong beach is beautiful with flat sand beach and calm wave. The first day of swimming at sea is always the best because people are eager to swim and enjoy. Later days, people are lazy to swim because they have to wear swimsuit and bath after that. One good thing in this trip is I paid attention on meals. I planned it well with flexibility so during this trip, we had chance to enjoy a variety of delicious food. For dinner, I ordered Bánh xèo tôm and mực. And it proved that it was a perfect decision. People were satisfied, excited about it. => ChatGPT: "After leaving Trung Luong's cafe, we made our way to Cau Ut's restaurant for lunch. We indulged in a delightful meal, highlighted by a crab hotpot, leaving everyone satisfied and in high spirits. Upon arriving at Nha cua ba's homestay, I took a lengthy nap and later rallied everyone for a swim. Trung Luong beach, with its serene sandy shores and gentle waves, provided the perfect setting. The first day of swimming in the sea is always a highlight, as everyone is eager to embrace the experience. As the trip unfolded, the enthusiasm for swimming waned a bit due to the effort of putting on swimsuits and bathing afterward.
One exceptional aspect of this journey was my attention to meals. I had planned our dining experiences meticulously, offering flexibility for us to savor a diverse range of delicious dishes. For dinner, I ordered Bánh xèo with shrimp and squid, a decision that proved to be impeccable. The satisfaction and excitement it brought to everyone were evident."
rephrase: In the evening, it continued raining so we had no activity except sleeping. Next day, starting with good breakfast with local food and restaurant. It still rained but we go around the homestay to snap beautiful photos. People were lazy to swim but we made an effort to push children to swim. Leaving Nha Cua Ba’s homestay at 12h to head to Nhon Hai’s village. => ChatGPT:"During the evening, the rain persisted, leaving us with no option but to retire early and rest. The following day began with a satisfying breakfast at a local restaurant, where we indulged in regional cuisine. Despite the ongoing rain, we decided to explore the vicinity of the homestay and capture some picturesque photos. Although most of us were reluctant to swim, we encouraged the children to take a dip. Around noon, we bid farewell to Nha Cua Ba's homestay and set off for Nhon Hai's village.
In the past, I was shy to contact or talk, communicate with people. Now I’m more comfortable because I approach with mutual benefit (picking up children, funeral together, just cafe to enjoy weather and chat…). If people don’t response, I don’t have to wonder why, wait for them. I just ignore and move next people, do other tasks. They are just one of my task. Their ignore doesn’t affect my life and my plan => In the past, I used to be shy to contact or talk, communicate with other people. Nowadays, I've been comfortable in communication as I approach interactions with a focus on mutual benefit. If someone doesn't respond, I no longer dwell on why or wait for them. I simply move on, viewing it as just one of the tasks in my daily life. Their lack of response doesn't have a significant impact on my life or my plans.
rephrase: In the past, I identify some target people and plan to be close with them. So I try to show off, do together. Then I will be disappointed, upset, confused if they don’t response or keep distance with me. I felt hurt, think that they disregard me so I try to attract, show off more (about stock, foreign company, travel, children’s achievements, my knowledge, elegance…). Now I’m independent with it. I communicate with large of people (strangers, neighbors…); keep contact with many people. So if someone ignore (Toanu, HienChu, VanLTV…), I just bypass and communicate with other people. I’m busy and self-fulfilled. I can enjoy alone, be lonely so I don’t need any relationships, beg for them. I communicate with them just for mutual benefit or for relaxing (cafe, chat…) so if they don’t response, I don’t have to insist, ask twice, wonder why, be vulnerable, be hurt. I don’t communicate with anyone for any purpose (to be close: customer, Quyen EVN, LeHung, Quyet, VietKaist….) so I can be dignified, know my values, contribution, independent: just come to query them, or cafe, chat for fun, weather… => ChatGPT: In the past, I used to target specific individuals and make efforts to build close relationships with them (by visiting house, spending time together, showing off my achievements or my children’s achievements…). However, I'd become disappointed, upset, and confused if they didn't respond or maintained a distance. It would make me feel hurt, as if they were disregarding me, leading me to try to showcase my achievements and knowledge, earning from stocks, working foreign companies, traveling, or my children's accomplishments. Now, I've become more self-reliant in my interactions. I engage with a wide range of people, including strangers and neighbors, and maintain connections with many. So, if someone like Toanu, HienChu, or VanLTV… ignores me, I simply move on and engage with other individuals. I lead a busy and self-fulfilled life, capable of enjoying my own life without relying on relationships or seeking validation. My interactions are now based on mutual benefit or for relaxation purposes, like meeting for coffee or
chatGPT nói riêng hay mô hình ngôn ngữ lớn (LLM) nói chung, đã và đang đạt được những thành tựu cực kỳ quan trọng lĩnh vực ngôn ngữ tự nhiên, chúng dần dần xuất hiện ở mọi nơi và cho thấy một khả năng đáng ngạc nhiên trong việc khiến máy tính có thể giao tiếp với con người.
Kể từ thời điểm chatGPT công bố phát hành miễn phí, nó đã tạo nên một cơn sốt không chỉ trong lĩnh vực AI mà còn được chú ý trên toàn cầu. Nó không chỉ có khả năng trò chuyện, trả lời giống con người mà còn có thể làm được rất nhiều thứ khác mà bạn chưa hề biết.
Trong bài viết này, mình sẽ giới thiệu cho các bạn về Prompt Engineering, một kỹ thuật có thể giúp bạn tối ưu sức mạnh của chatGPT hay LLMs.
Mô hình ngôn ngữ lớn (Large Language Models) là các hệ thống trí tuệ nhân tạo được xây dựng để hiểu và tạo ra ngôn ngữ tự nhiên. Chúng được huấn luyện trên các tập dữ liệu lớn và đa dạng để học cách dự đoán từ tiếp theo trong văn bản. Một mô hình ngôn ngữ lớn có thể hiểu ngữ cảnh, xử lý ngôn ngữ tự nhiên và trả lời các câu hỏi.

Thực chất, LLM cũng như những Language model trước kia như Bert hay XLNet, chỉ khác là qua thời gian, nhờ vào sự phát triển của công nghệ tính toán, ta có thể triển khai được những mô hình có kích thước gấp trăm/ ngàn lần trước kia, và đó là cách mà mô hình ngôn ngữ lớn ra đời -- những mô hình ngôn ngữ có trên 10 tỉ tham số.
Các bạn có thể tìm hiểu thêm về Large Language Model tại bài viết trước của mình:
Kể từ khi Google công bố kiến trúc transformer, đánh dấu thời đại mới cho lĩnh vực xử lí ngôn ngữ tự nhiên, kéo theo hàng loạt các mô hình NLP tiên tiến qua các năm, có thể kể đến BERT 2019, GPT 2020, PaLM 2023,... và kèm theo đó là số lượng tham số trong mỗi mô hình cũng tăng theo cấp số nhân (hình bên phải)

Và nếu như bạn truy cập vào trang web futurepedia.io, bạn sẽ dễ dàng nhìn thấy hàng trăm ứng dụng/website AI hỗ trợ các tác vụ liên quan đến NLP, và rất nhiều trong đó, đều dựa trên nền tảng LLM của openAI. Có thể nói, những ứng dụng từ LLM đang trở nên tràn ngập và giải quyết rất nhiều vấn đề cho người dùng.
Và không thể không đề cập tới, đứng phía sau việc xây dựng những ứng dụng này, chính là một kĩ thuật gọi là Prompt Engineering, một kĩ thuật không quá phức tạp nhưng lại mang về nhữung kết quả không ngờ với LLMs

Như các bạn đã biết, LLM hoạt động như một hộp đen đã trược train trên một tập dữ liệu cực kỳ lớn, nó nhận vào một đoạn text và trả về một đoạn text khác dưới dạng phản hồi. Có nghĩa là, kết quả mà LLM hay chatGPT trả về, hoàn toàn phụ thuộc vào input mà ta đưa vào cho nó.
Prompt Engineering chính là một kĩ thuật giúp ra điều chỉnh các yếu tố trong đoạn text input để tối ưu hóa đầu vào LLM và khai thác tối đa hiệu quả của LLM mà ta nhận được qua output.
Một số kĩ thuật prompt engineering đơn giản và dễ thấy nhất, chính là cunng cấp ngữ cảnh trong input, hay thêm một số điều kiện và ràng buộc cho output mà ta muốn nhận được. mình sẽ nói rõ hơn về điều này ở phần sau.
Bằng việc hiểu rõ và sử dụng các kĩ thuật prompt engineering, ta có thể định hình output mong muốn, chọn lọc các thông tin và tùy biến LLM để thực hiện nhiều loại nhiệm vụ nhất định.

Hãy nhìn vào hai ví dụ trên, đây là những thứ mình thường làm khi sử dụng chatGPT, hâu hết là để giải đáp những việc lặt vặt thôi. Nhưng nếu chỉ có vậy thì cần gì prompt engineering nhỉ? Đúng vậy, những gì chatGPT có thể làm còn nhiều hơn thế, và đây là các để bạn khai thác điều đó.
Hãy lấy một ví dụ, mình đang muốn xây dựng một ứng dụng quản lý tài chính cá nhân cho một lượng lớn người dùng, và mình muốn chatGPT gợi ý cho mình những framework có thể sử dụng để đáp ứng nhu cầu cho ứng dụng của mình.

Việc đầu tiên, cũng chính là việc đơn giản nhất, hãy viết điều bạn muốn, đây được xem như là một "hướng dẫn" cho chatGPT. Nguyên tắc của hướng dẫn là: càng chi tiết càng tốt. Vì vậy, ngoài việc yêu cầu thiết kế kiến trúc cho ứng dụng, mình còn mô tả thêm về chi tiết của ứng dụng mà mình định xây dựng

chatGPT là một mô hình ngôn ngữ đa năng, nhưng đôi khi việc đa năng này là không cần thiết khi hỏi về một ĩnh vực nhất định. Vì vậy, ta cần cho nó một vai trò để giới hạn lĩnh vực của nó.
Điều này cũng tương tự như việc khi ta muốn hỏi về sức khỏe ta sẽ tìm đến bác sĩ hay khi muốn hỏi về luật ta sẽ tìm đến luật sư. Trong trường hợp này, mình sẽ gán cho chatGPT với vai trò là một kĩ sư có nhiều kinh nghiệm về lĩnh vực kiến trúc hệ thống phần mềm.

Đây là phần mà ta sẽ cung cấp nhiều thông tin hơn cho chatGPT, tại phần này mình sẽ cho nó những thông tin về câu trả lời mà nó cần tạo ra, cách suy nghĩ và phân tích,...

Và cuối cùng, hãy định hình cho kết quả mà bạn muốn, bạn có thể mô tả hình thức, cách trả lời, độ dài, mức độ chi tiết,.... Ở đây, mình yêu cầu kết quả phải ở dạng một báo cáo ngắn và mang tính thực tế.

Thông qua một prompt được xây dựng từ các bước như trên, ta có thể dễ dàng nhận được một kết quả như hình dưới, trong đó:
chatGPT liệt kê tất cả những thành phần liên quan đến phần mềm mình cần xây dựng
Với mỗi thành phần, chatGPT cho mình 2-3 lựa chọn cho những framework mà mình có thể sử dụng
Đồng thời, chatGPT cũng phân tích lợi và hại khi mình sử dụng mỗi framework

Nhưng kế sau đó, mình đã đưa ra thêm một yêu cầu, đó là lựa chọn framework tốt nhất cho mỗi thành phần và đưa ra kết quả dưới dạng JSON, và thu được một bảng tóm tắt cực kỳ rõ ràng cho ứng dụng của mình như hình bên phải.
Rõ ràng, mình như từng đề cập gì về JSON hay mô tả JSON có cú pháp như thế nào nhưng chatGPT vẫn biết và đưa ra được kết quả phù hợp, đây chính là một trong những khả năng quan trọng của chatGPT hay LLM gọi là IN-context Learning, điều này giúp cho LLM có thể năng hiểu và thích nghi với những yêu càu tùy vào bối cảnh.
Chi tiết hơn về in-context learning, mình sẽ trình bày rõ hơn trong mục các kĩ thuật Prompt Engineering nâng cao.

Trong ví dụ trên, mình đưa ra cùng một input trong 2 lần yêu cầu, chỉ khác ở chỗ mình để hờ một từ gợi ý ở cuối input để hướng dẫn cho chatGPT về output. Như bạn đã thấy, mình hoàn toàn không hề mô tả hay kì vọng chatGPT sẽ đưa ra "Positive" hay "Negative" trong câu đầu, đó hoàn toàn là chatGPT tự hiểu và thực hiện. Đây chính là một ví dụ cho khả năng thích ứng với bối cảnh: in-context learning.
Kĩ thuật zero-shot là kĩ thuật cơ bản và dễ dàng nhất mà ta có thể sử dụng trong nhiều trường hợp để khiến chatGPT đưa ra output ta muốn

Phát triển từ kĩ thuật zero-shot, nhưng lần này, ta không để hờ yêu cầu nữa, mà ta sẽ thực sự tạo ra một mẫu phản hồi để chatGPT bắt chước, trong ví dụ trên, mình đưa ra một câu hỏi mẫu và liệt kê những câu trả lời theo format mình muốn. Sau khi kết thúc câu hỏi đầu tiên và đưa ra câu trả lời làm ví dụ, mình đặt ra câu hỏi thứ 2 và để chatGPT thực hiện phần còn lại.
Tương tự thế, kĩ thuật few-shot cũng yêu cầu bạn đưa ra mẫu ví dụ, nhưng chỉ khác là bạn có thể đưa ra nhiều ví dụ hơn thay vì 1, điều này sẽ chắc chắn hơn việc chatGPT sẽ tạo ra một câu htrả lời có format tương tự như những ví dụ mà bạn đưa ra cho nó.
Nhưng bạn sẽ thấy rằng, one-shot hay few-shot không phải lúc nào cũng hoạt động đúng, đặc biệt là với những câu hỏi hóc búa và yêu cầu tư duy như con người như ví dụ bên dưới.

Giải pháp cho vấn đề này chính là chain-of-thoughts, để khắc phục vấn đề của ví dụ trên, mình đã giải thích cách suy nghĩ từng bước để tìm ra đáp án. Và như bạn thấy, chatGPT cũng đã thay đổi cách nó đưa ra câu trả lời, và nó cố gắng suy nghĩ và giải thích y hệt như ví dụ mình đưa ra.

Một phương pháp ngắn gọn hơn gọi là zero-shot chain-of-thoughts (sự kết hợp giữa zero-shot và chain-of-thoughts), thay vì mô tả cách tư duy một cách cụ thể và rõ ràng, mình chỉ yêu cầu chattGPT: "Hãy suy nghĩ từng bước". và kết quả nhận được cũng tương tự. Tuy nhiên, cách này chỉ hoạt động trong một số trường hợp đơn giản không cần tư duy quá phức tạp.

Một ví dụ khác cho chain of-thoughts là bạn có thể hướng dẫn cách chatGPT tạo ra một nội dung, trong ví dụ bên dưới, mình đã yêu cầu chatGPT tạo ra một câu chuyện với một chủ đề, nhưng đồng thời cũng cho nó một vài tình huống và thứ tự diễn ra trong câu chuyện, và nhận được một câu chuyện đúng như yêu cầu đưa ra.

Nếu bạn là lập trình viên, bạn sẽ khá quen thuộc với khái niệm function: một hàm chức năng nhập đầu vào, thực hiện những nhiệm vụ nhất định, và đưa ra output. Prompt function cũng tương tự, bạn sẽ hướng dẫn để chatGPt hiểu rằng nó sẽ đóng vai làm một hàm thực hiện chức năng, kèm theo đó, hãy mô tả rõ ràng input, output và parameter mà bạn đưa vào cho chatGPT

Ngoài những kỹ thuật cơ bản đã nêu ở trên, cũng có những kĩ thuật nâng cao hơn nhằm khắc phục nhược điểm và tạo cho LLM những tính năng ưu việc
Cơ chế ngẫu nhiên của chatGPT đôi khi là một vấn đề, nó thường sẽ đưa 10 đáp án khác nhau cho cùng 1 câu hỏi, vì thế ta không thể nào chắc được đáp án đầu tiên ta nhận được là đáp án chính xác. Generated Knowledge Prompting sẽ giải quyết điều đó bằng cách yêu cầu chatGPT tạo ra nhiều câu hỏi lời cho cùng 1 câu hỏi, sau đó sẽ đưa 10 đáp án nhận được vào context và yêu cầu chatGPT lựa chọn phương án phù hợp và đúng đắn nhất.

Một kĩ thuật phát triển từ chain-of-thoughts, nhưng thay vì chỉ có 1 luồn suy nghĩ, bạn sẽ yêu cầu chatGPT phải suy luật theo nhiều luồn suy nghĩ và loại bỏ những tư duy sai lầm. Điều này sẽ khắc phục được vấn đề khi suy luận của chatGPT sai trong chain-of-thoughts sẽ khiến toàn bộ kết quả bị sai.
Phương pháp triển khai kĩ thuật này cũng rất đơn giản, bạn sẽ nói GPT tưởng tượng ra một cuộc hợp có 10 chuyên gia trong lĩnh vực và yêu cầu họ suy luận và đưa ra cách giải quyết, bất cứ người nào nhận ra sai lầm trong suy luận của họ sẽ phải tự giác rời đi. Cuối cùng, người ở lại cuối cùng sẽ đưa ra đáp áp của họ.

Một kĩ thuật được sử dụng cực kì phổ biến trong các ứng dụng LLM-based, Bằng việc nhúng các tài liệu bên ngoài thành dạng embedding vector, ta có thể cho phép LLM truy cập vào một lượng lớn dữ liệu bên ngoài mà chatGPT chưa từng được học trước đó. Với mỗi câu hỏi được đưa vào chatGPT, câu hỏi sẽ được nhungs về dạng vector và so sánh với tất cả những tài liệu khác trong cơ sở dữ liệu để xác định đâu là nguồn tài liệu có khả năng chứa câu trả lời. Sau đó, tài liệu này sẽ được đưa vào prompt dưới dạng context kèm với câu hỏi để đưa được ra câu trả lời cho câu hỏi đó như mong muốn
Bạn có thể dễ dàng bắt gặp phương pháp này trong rất nhiều ứng dụng, tiêu biểu nhất là các ứng dụng tra cứu hoặc trả lời câu hỏi trên các lĩnh vực nhất định.

Mình xin chia sẻ một vài thành phần trong prmopt có thẻ hữu ích trong quá trình bạn sử dụng hay xây dựng ứng dụng với LLM. Cám ơn đã theo dõi
Nếu bạn muốn xem nhiều bài viết về AI của mình hơn, bạn có thể ghé thăm blog của Lưu
Nguồn: https://luu.name.vn/tat-tan-tat-nhung-ki-thuat-prompt-engineering-huu-ich-nhat-cho-chatgpt/
I use ChatGPT for email, chat, how to response: both for English, structure, EQ…
KC use ChatGPT for her science homework
========================


Do you know what is the underlying reason that our AMF doesn’t support for ANR? In my opinion, ANR is important feature and all tier-1 operators request for it.
ChatGPT Do you know the underlying reason why our AMF doesn’t support ANR? In my opinion, ANR is an important feature, and all tier-1 operators are requesting it.
Rephrase: Hi George, Our customer wants to know why NSN does not support Event Exposure with bulk subscription? Can you provide some details? And for the RFP clause “The proposed system shall support to interwork with NEF for exposure network capability in comply with CAMARA project”, what should they change for Nokia to FC?
ChatGPT: Hi George, Our customer is asking why NSN does not support Event Exposure with bulk subscriptions. Could you provide some details? Additionally, regarding the RFP clause, "The proposed system shall support interworking with NEF for exposure network capability in compliance with the CAMARA project," what modifications should they make for Nokia to provide FC?
rephrase: Hi Magnus, I fell off motorbike last week and had an operation. Now although it finished, I still walk on a clutch. Therefore I mind going to the customer’s office with a clutch. Truong may accompany with you to the customer. In my opinion, there are some topics the customer may need us consult - Mobifone container terms, concepts, rfp - Viettel request for 5G call flow last time. Also recently they show interest for VoWiFi, FWA solutions: is it worth to invest, experiences from other operators, how to deploy…
ChatGPT: Hi Magnus, I had a motorbike accident last week and had to undergo surgery. Although the surgery is over, I'm still using crutches to walk. As a result, I don't feel comfortable visiting the customer's office. Truong can accompany you to the customer instead. In my opinion, there are a few topics the customer may need our consultation on: Mobifone container terms, concepts, and RFP. Viettel's request for 5G call flow information. Additionally, they have recently shown interest in VoWiFi and FWA solutions: whether it’s worth investing in, experiences from other operators, and how to deploy them.
rephrase: We are exploring the possibility of offering NCP, instead of NCS. Can you check in your team how it impacts on Delivery and Care services? => We are considering the option of providing NCP instead of NCS. Could you please inquire within your team about the potential impact on Delivery and Care services?
rephrase: Thank you team for your comments I consolidate it in the slide in the following links. Please take a look, provide your feedback and edition. => ChatGPT: "I appreciate your input, team. I've compiled it into the presentation linked below. Please review it, share your feedback, and suggest any necessary edits."
rephrase: "Next actions: Mobifone will provide the final scope based on the vendor’s recommendations. Nokia will do re-dimensioning and provide the updated BoQ Mobifone and Nokia work together for the Mobifone’s solution roadmap of next phases" => ChatGPT: "Upcoming steps: Mobifone is set to determine the definitive scope following the recommendations of the vendor. Nokia will proceed with re-dimensioning and furnish an updated Bill of Quantities (BoQ). Mobifone and Nokia will collaborate to formulate Mobifone's solution roadmap for the forthcoming phases."
rephrase: "Nokia propose to have site 2 in Hanoi for geo-redundancy. Site 2 is for critical network elements like SDM, PCF, 5G charging mediation nodes We should not have site 2 in HCM city because It makes a real geo-redundancy since both sites are at the same city If in the future, Mobifone want to deploy a new site in Hanoi for these nodes, we have to redeploy. So it is better to make a decision upfront" => ChatGPT: "Nokia suggests establishing Site 2 in Hanoi to ensure geo-redundancy. Site 2 is designated for housing critical network components such as SDM, PCF, and 5G charging mediation nodes. It is not advisable to position Site 2 in Ho Chi Minh City for the following reasons:
It does not provide genuine geo-redundancy since both sites are located in the same city.
In the event that Mobifone decides to deploy a new site in Hanoi for these nodes in the future, it would necessitate a redeployment. Therefore, it is more prudent to make this decision proactively."
rephrase: Accordign to the standard N 5G traffic profile, Average bit rate per UE/BH is 700 Kbps. So with the scope of 1mil subscribers, our optimistic estimation is the total throughput of 700 Kpbs*1mil = 700 Gbps. However with N’s experiences in other networks, the average bit rate per SAU from the early 5G SA deployment varies from 200 to 500kbps. So we can estimate that the total throughput is around 200 Gbps – 500 Gbps. Nokia recommend we shoud start small with 200 Gbps in the first phase and then expand later in the next phase (with other UPF nodes in site Hanoi). => ChatGPT: "Following the standard N 5G traffic profile, the average bit rate per UE/BH is 700 Kbps. Considering a subscriber base of 1 million, our optimistic estimate for the total throughput would be 700 Gbps. However, based on Nokia's experience in other networks, the average bit rate per SAU in early 5G SA deployments has ranged from 200 to 500 Kbps. Therefore, we can estimate the total throughput to be approximately in the range of 200 Gbps to 500 Gbps. In light of this, Nokia recommends that we initiate the first phase with a conservative estimate of 200 Gbps and consider expansion in subsequent phases, potentially with the addition of other UPF nodes in the Hanoi site."
rephrase: Need to check HW capacity, license in the existing Mobifone provisioning system for a new 1 millions 5G subscribers. The existing provisioning system need to integrate with new 5G network elements such as UDM, AUSF, PCF...=> ChatGPT
"We should conduct a capacity and licensing assessment within Mobifone's current provisioning system to accommodate a new influx of 1 million 5G subscribers. Furthermore, the existing provisioning system will require integration with the new 5G network elements, including UDM, AUSF, PCF, among others."

repharse: Mobifone can deploy SDM overlay solution during this intermediate period from 4G to 5G SA while waiting for a new modernization project of the existing SDM system Sim change is not mandatory
ChatGPT
"Mobifone has the option to implement an SDM overlay solution during the transitional phase between 4G and 5G SA, as we await the commencement of a forthcoming modernization project for the current SDM system. It's important to note that sim card replacement is not obligatory in this context."

rephrase: Except the CHF use case, Nokia recommend to have other optional use cases in the 2nd phase deployment
ChatGPT
"With the exception of the CHF use case, Nokia suggests incorporating additional optional use cases in the second phase of deployment."

rephrase: For SMS in 5G, there are 2 options:SMS via IP-SM-GW and SMS via SMSF. Nokia recommend to have SMSF in the 2nd phase, when mMTC use cases are mature enough. EIR is optional. Nokia recommend to deploy it in the next phase SCP: mandatory for implementing SBA Model C and D. Nokia strongly recommend to deploy it in the 1st phase because deploying in the next phases require changing in network architecture, signaling rerouting, then create many opex expenses for reconfiguration and reintergration NRF: mandatory for implementing SBA Model B, C and D. Nokia strongly recommend to deploy it in the 1st phase. Model A is not recommended for production BSF: optional. Nokia recommend to have SEPP in the 2nd phase when having multiple PCF instances which require session binding SEPP: optional if no roaming. Nokia recommend to have SEPP in the 2nd phase when Mobifone have 5G roaming agreement with roaming partners and also when the customers demand high 5G data throughput. In the first phase, 5G subscriber can use VoLTE and 4G data roaming instead. NEF: optional. It is for API gateway/portal and plugins for 3rd application and developers. Nokia recommend to have NEF in the 2nd phase with use cases such as Monitoring Location_Report Event
ChatGPT
Here is the information rephrased:
Regarding SMS in the 5G context, there are two options: SMS via IP-SM-GW and SMS via SMSF. Nokia suggests implementing SMSF in the second phase, once the mMTC use cases are sufficiently developed.
EIR is an optional component, and Nokia recommends deploying it in a subsequent phase.
SCP is mandatory for implementing SBA Models C and D. Nokia strongly advises deploying it in the first phase since introducing it in later phases would necessitate changes in network architecture, signaling rerouting, resulting in increased operational expenses for reconfiguration and reintegration.
NRF is mandatory for implementing SBA Models B, C, and D. Nokia highly recommends deploying it in the first phase. Model A is not recommended for production.
BSF is optional. Nokia suggests considering SEPP in the second phase, especially when multiple PCF instances requiring session binding are in use.
SEPP is optional unless there is roaming involved. Nokia recommends introducing SEPP in the second phase when Mobifone has 5G roaming agreements with partners and when customers demand high 5G data throughput. In the first phase, 5G subscribers can use VoLTE and 4G data roaming.
NEF is optional and serves as an API gateway/portal with plugins for third-party applications and developers. Nokia recommends implementing NEF in the second phase, particularly for use cases like Monitoring Location_Report Event.
rephrase: We will have a Mbf 5G SA trial workshop with customer on Wed 1Nov2023. It is an important meeting because their depute director will lead the meeting. So please help to join to support me in case I need for technical questions => ChatGPT: On Wednesday, November 1st, 2023, we are scheduled to conduct a trial workshop on Mobifone's 5G SA with the customer. This meeting holds significant importance as it will be led by their Deputy Director. Therefore, I kindly request your presence to assist me in addressing any technical queries that may arise.
rephrase: There is a new update of scope from the customer. Can you help to update the NPT? => ChatGPT: The customer has provided a new scope update. Can you assist in revising the NPT accordingly?
rephrase: In my opinion, optimized solution 1 is most likely accepted by the customer => ChatGPT: From my perspective, it's highly probable that the customer will accept optimized solution 1.
rephrase: To prepare for the customer discussion of scope optimization, the customer may want to know how much cost/price they can save if they take out this scope. So can you help to prepare some number for it: => ChatGPT: In anticipation of the customer's discussion regarding scope optimization, they might be interested in understanding the potential cost savings associated with removing certain elements from the scope. Could you assist in generating some cost estimates for this purpose?
rephrase: Career: I should not only focus on knowledge (try to read more) but skills: know how to do (working order, what need to do…), process, networking (who to ask). In the past, I was used to only follow knowledge, only try to read, to become expert, know more or better than people, then show off. But starting from TPM, I moved to high level approach. When I moved to PM role, I focused less on knowledge, spent some time for knowledge. => ChatGPT: Career-wise, it's important for me to not solely emphasize acquiring knowledge (such as reading more), but also to develop skills encompassing practical know-how (how to execute tasks effectively, understanding processes), building a network (connecting with people who can provide guidance), and cultivating a holistic approach. In the past, my approach was primarily centered on accumulating knowledge. I aimed to read extensively, become an expert, and outshine others in terms of knowledge, often showcasing my expertise. However, my perspective shifted towards a more high-level approach when I transitioned into a TPM role. As I moved into a PM role, I began to place less emphasis on accumulating knowledge and allocated more time to skill development and knowledge maintenance.
encompass /ɪnˈkʌm.pəs/: to include different types of things:
The festival is to encompass everything from music, theatre, and ballet to literature, cinema, and the visual arts.
rephrase: Hi, We are preparing to submit the quotation. In our solution proposal, we proposed 12 failover servers . However, to compete with our main competitor, we are considering to optimize the number of servers. I would like to reduce the number of failover server (to 6 servers or even 2 failover server only). Please help to share your view. => ChatGPT: Hello, We are in the process of finalizing our quotation. In our solution proposal, we initially suggested using 12 failover servers. However, in order to stay competitive with our main rival, we are contemplating the optimization of server numbers. I am interested in reducing the failover server count, possibly down to 6 servers or even as low as 2. I would appreciate your view on this matter.
E// already has 5G SA trial phase 2 with Viettel. Viettel has no plan for 5G trial with other vendor which make us to have to propose new thing for trial => ChatGPT: "Ericsson (E//) has already conducted the second phase of 5G Standalone (SA) trials with Viettel. As Viettel has no plans for 5G trials with other vendors, we need to propose new ideas for further trial phases."
Huawei SDM’s swapping bid is delayed. The customer decided to expand Huawei SDM’s capacity instead => ChatGPT. The swapping bid for Huawei SDM has been postponed. Instead, the customer has chosen to increase the capacity of Huawei SDM.
rephrase: Regarding to RFP/SOC, we complied the following. But it only mentions about 05 year of life cycle and repair/return; not mentions about the right of 06 year of supply of spare parts => ChatGPT: In reference to the RFP/SOC, we have compiled the following information. However, it solely addresses a five-year lifecycle and repair/return aspects, without specifying the entitlement to a six-year supply of spare parts.
rephrase: "Our ATCA products are reaching limited HW manufacturing stage we cannot commit for the request of supply of equivalent spare parts for a period of 5 years. However Nokia commit to provide repair services during 5 years services under warranty and/or separate care agreement” => ChatGPT: "While our ATCA products are progressing towards the limited hardware manufacturing phase, we are unable to guarantee the availability of equivalent spare parts for a period of five years. Nonetheless, Nokia is committed to offering repair services throughout this five-year period, both under warranty and through separate care agreements."
dịch sang tiếng Anh: thuê bao 5G có thể nằm phân tán trên các dải đầu số khác nhau; trong khi DRA chỉ có thể định tuyến nông theo dải đầu số. DRA không thể định tuyến sâu đến mức nguyên cả số msisdn => ChatGPT: "5G subscribers can be distributed across different number ranges, while the DRA (Diameter Routing Agent) can only route traffic based on number ranges. The DRA cannot route down to the level of individual MSISDN numbers."
rephrase: DRA route base on msisdn/imsi range. But 5G subscribers are random, may belong to different ranges. So how DRA is able to route 5G traffic to New SDM system? DRA cannot route for a specific msisdn/imsi number. It has to route based on a specific range: such as 09031 => ChatGPT: The DRA (Diameter Routing Agent) traditionally routes traffic based on the MSISDN/IMSI range. However, with 5G subscribers, the challenge arises because they can belong to various ranges, making it difficult to route based on specific individual numbers (MSISDN/IMSI). Instead, the DRA must adapt to the randomness of 5G subscriber assignments and route traffic effectively to the new SDM (Subscriber Data Management) system. In this context, the DRA cannot route traffic for a particular MSISDN/IMSI number; it must operate based on broader criteria, such as specific range identifiers like "09031."
Rain so could not go out to take photo or visit. Had a cold breakfast. Worried for my trip, may it completely ruin. We may have to stay at homestay the whole day. People’s mood isn’t good. But my plan was still good when I decided to stay in Trung Luong cafe shop in the morning. When it stopped raining, I went out to take photo so encourage people to take photos too. Then had a good decision to visit Trung Luong camping area where children can play and we had some nice photos => Due to the rain, we couldn't venture out to take photos or go sightseeing. We settled for a cold breakfast and I grew anxious about how the weather might disrupt my trip. There was a possibility that we'd be confined to our homestay for the entire day, and everyone's mood was rather somber. However, my initial plan to spend the morning at Trung Luong cafe proved fortuitous. As soon as the rain subsided, I seized the opportunity to head outside and capture some photos, then inspire others to do the same. Eventually, I made the wise decision to explore the Trung Luong camping area, where children could enjoy themselves, and we could snap some beautiful pictures.
rephrase: Minh’s iphone doesn’t have high resolution so “nước ảnh không trong” => Minh's iPhone lacks high resolution, so the photos don't come out clear.
rephrase: This picture is funny when there is a contrary between a lamp’s face and Minh Khanh’s face => ChatGPT: "This picture becomes humorous due to the contrast between the expression on the lamp's face and Minh Khanh's expression."
rephrase: Look at Khanh An’s face: always funny and happy => ChatGPT: "Observe Khanh An's expression; it's consistently cheerful and full of joy."
rephrase: Leaving airport for Trung Luong. It rained drizzling. Reaching to homestay in a small fishing village. It is not comfortable. I felt worried. Only satisfied that it locates near the beach with nice sea view. Put luggage at homestay and move to Trung Luong’s cafe. => ChatGPT: "Departing from the airport en route to Trung Luong, we encountered a light drizzle. Upon arriving at a homestay in a quaint fishing village, we found it to be less than comfortable, which left me feeling somewhat concerned. The only consolation was its proximity to the beach, offering a beautiful sea view. After dropping off our luggage at the homestay, we made our way to Trung Luong's cafe."
rephrase: After leaving Trung Luong’s cafe, we headed to Cau Ut’s restaurant for lunch. Delicious lunh with hop pot of crab. People were satisfied and happy. Reaching to Nha cua ba’s homestay. After a long nap, I woke up and urged people for swimming. Trung Luong beach is beautiful with flat sand beach and calm wave. The first day of swimming at sea is always the best because people are eager to swim and enjoy. Later days, people are lazy to swim because they have to wear swimsuit and bath after that. One good thing in this trip is I paid attention on meals. I planned it well with flexibility so during this trip, we had chance to enjoy a variety of delicious food. For dinner, I ordered Bánh xèo tôm and mực. And it proved that it was a perfect decision. People were satisfied, excited about it. => ChatGPT: "After leaving Trung Luong's cafe, we made our way to Cau Ut's restaurant for lunch. We indulged in a delightful meal, highlighted by a crab hotpot, leaving everyone satisfied and in high spirits. Upon arriving at Nha cua ba's homestay, I took a lengthy nap and later rallied everyone for a swim. Trung Luong beach, with its serene sandy shores and gentle waves, provided the perfect setting. The first day of swimming in the sea is always a highlight, as everyone is eager to embrace the experience. As the trip unfolded, the enthusiasm for swimming waned a bit due to the effort of putting on swimsuits and bathing afterward.
One exceptional aspect of this journey was my attention to meals. I had planned our dining experiences meticulously, offering flexibility for us to savor a diverse range of delicious dishes. For dinner, I ordered Bánh xèo with shrimp and squid, a decision that proved to be impeccable. The satisfaction and excitement it brought to everyone were evident."
rephrase: In the evening, it continued raining so we had no activity except sleeping. Next day, starting with good breakfast with local food and restaurant. It still rained but we go around the homestay to snap beautiful photos. People were lazy to swim but we made an effort to push children to swim. Leaving Nha Cua Ba’s homestay at 12h to head to Nhon Hai’s village. => ChatGPT:"During the evening, the rain persisted, leaving us with no option but to retire early and rest. The following day began with a satisfying breakfast at a local restaurant, where we indulged in regional cuisine. Despite the ongoing rain, we decided to explore the vicinity of the homestay and capture some picturesque photos. Although most of us were reluctant to swim, we encouraged the children to take a dip. Around noon, we bid farewell to Nha Cua Ba's homestay and set off for Nhon Hai's village.
In the past, I was shy to contact or talk, communicate with people. Now I’m more comfortable because I approach with mutual benefit (picking up children, funeral together, just cafe to enjoy weather and chat…). If people don’t response, I don’t have to wonder why, wait for them. I just ignore and move next people, do other tasks. They are just one of my task. Their ignore doesn’t affect my life and my plan => In the past, I used to be shy to contact or talk, communicate with other people. Nowadays, I've been comfortable in communication as I approach interactions with a focus on mutual benefit. If someone doesn't respond, I no longer dwell on why or wait for them. I simply move on, viewing it as just one of the tasks in my daily life. Their lack of response doesn't have a significant impact on my life or my plans.
rephrase: In the past, I identify some target people and plan to be close with them. So I try to show off, do together. Then I will be disappointed, upset, confused if they don’t response or keep distance with me. I felt hurt, think that they disregard me so I try to attract, show off more (about stock, foreign company, travel, children’s achievements, my knowledge, elegance…). Now I’m independent with it. I communicate with large of people (strangers, neighbors…); keep contact with many people. So if someone ignore (Toanu, HienChu, VanLTV…), I just bypass and communicate with other people. I’m busy and self-fulfilled. I can enjoy alone, be lonely so I don’t need any relationships, beg for them. I communicate with them just for mutual benefit or for relaxing (cafe, chat…) so if they don’t response, I don’t have to insist, ask twice, wonder why, be vulnerable, be hurt. I don’t communicate with anyone for any purpose (to be close: customer, Quyen EVN, LeHung, Quyet, VietKaist….) so I can be dignified, know my values, contribution, independent: just come to query them, or cafe, chat for fun, weather… => ChatGPT: In the past, I used to target specific individuals and make efforts to build close relationships with them (by visiting house, spending time together, showing off my achievements or my children’s achievements…). However, I'd become disappointed, upset, and confused if they didn't respond or maintained a distance. It would make me feel hurt, as if they were disregarding me, leading me to try to showcase my achievements and knowledge, earning from stocks, working foreign companies, traveling, or my children's accomplishments. Now, I've become more self-reliant in my interactions. I engage with a wide range of people, including strangers and neighbors, and maintain connections with many. So, if someone like Toanu, HienChu, or VanLTV… ignores me, I simply move on and engage with other individuals. I lead a busy and self-fulfilled life, capable of enjoying my own life without relying on relationships or seeking validation. My interactions are now based on mutual benefit or for relaxation purposes, like meeting for coffee or
chatGPT nói riêng hay mô hình ngôn ngữ lớn (LLM) nói chung, đã và đang đạt được những thành tựu cực kỳ quan trọng lĩnh vực ngôn ngữ tự nhiên, chúng dần dần xuất hiện ở mọi nơi và cho thấy một khả năng đáng ngạc nhiên trong việc khiến máy tính có thể giao tiếp với con người.
Kể từ thời điểm chatGPT công bố phát hành miễn phí, nó đã tạo nên một cơn sốt không chỉ trong lĩnh vực AI mà còn được chú ý trên toàn cầu. Nó không chỉ có khả năng trò chuyện, trả lời giống con người mà còn có thể làm được rất nhiều thứ khác mà bạn chưa hề biết.
Trong bài viết này, mình sẽ giới thiệu cho các bạn về Prompt Engineering, một kỹ thuật có thể giúp bạn tối ưu sức mạnh của chatGPT hay LLMs.
Mô hình ngôn ngữ lớn (Large Language Models) là các hệ thống trí tuệ nhân tạo được xây dựng để hiểu và tạo ra ngôn ngữ tự nhiên. Chúng được huấn luyện trên các tập dữ liệu lớn và đa dạng để học cách dự đoán từ tiếp theo trong văn bản. Một mô hình ngôn ngữ lớn có thể hiểu ngữ cảnh, xử lý ngôn ngữ tự nhiên và trả lời các câu hỏi.

Thực chất, LLM cũng như những Language model trước kia như Bert hay XLNet, chỉ khác là qua thời gian, nhờ vào sự phát triển của công nghệ tính toán, ta có thể triển khai được những mô hình có kích thước gấp trăm/ ngàn lần trước kia, và đó là cách mà mô hình ngôn ngữ lớn ra đời -- những mô hình ngôn ngữ có trên 10 tỉ tham số.
Các bạn có thể tìm hiểu thêm về Large Language Model tại bài viết trước của mình:
Kể từ khi Google công bố kiến trúc transformer, đánh dấu thời đại mới cho lĩnh vực xử lí ngôn ngữ tự nhiên, kéo theo hàng loạt các mô hình NLP tiên tiến qua các năm, có thể kể đến BERT 2019, GPT 2020, PaLM 2023,... và kèm theo đó là số lượng tham số trong mỗi mô hình cũng tăng theo cấp số nhân (hình bên phải)

Và nếu như bạn truy cập vào trang web futurepedia.io, bạn sẽ dễ dàng nhìn thấy hàng trăm ứng dụng/website AI hỗ trợ các tác vụ liên quan đến NLP, và rất nhiều trong đó, đều dựa trên nền tảng LLM của openAI. Có thể nói, những ứng dụng từ LLM đang trở nên tràn ngập và giải quyết rất nhiều vấn đề cho người dùng.
Và không thể không đề cập tới, đứng phía sau việc xây dựng những ứng dụng này, chính là một kĩ thuật gọi là Prompt Engineering, một kĩ thuật không quá phức tạp nhưng lại mang về nhữung kết quả không ngờ với LLMs

Như các bạn đã biết, LLM hoạt động như một hộp đen đã trược train trên một tập dữ liệu cực kỳ lớn, nó nhận vào một đoạn text và trả về một đoạn text khác dưới dạng phản hồi. Có nghĩa là, kết quả mà LLM hay chatGPT trả về, hoàn toàn phụ thuộc vào input mà ta đưa vào cho nó.
Prompt Engineering chính là một kĩ thuật giúp ra điều chỉnh các yếu tố trong đoạn text input để tối ưu hóa đầu vào LLM và khai thác tối đa hiệu quả của LLM mà ta nhận được qua output.
Một số kĩ thuật prompt engineering đơn giản và dễ thấy nhất, chính là cunng cấp ngữ cảnh trong input, hay thêm một số điều kiện và ràng buộc cho output mà ta muốn nhận được. mình sẽ nói rõ hơn về điều này ở phần sau.
Bằng việc hiểu rõ và sử dụng các kĩ thuật prompt engineering, ta có thể định hình output mong muốn, chọn lọc các thông tin và tùy biến LLM để thực hiện nhiều loại nhiệm vụ nhất định.

Hãy nhìn vào hai ví dụ trên, đây là những thứ mình thường làm khi sử dụng chatGPT, hâu hết là để giải đáp những việc lặt vặt thôi. Nhưng nếu chỉ có vậy thì cần gì prompt engineering nhỉ? Đúng vậy, những gì chatGPT có thể làm còn nhiều hơn thế, và đây là các để bạn khai thác điều đó.
Hãy lấy một ví dụ, mình đang muốn xây dựng một ứng dụng quản lý tài chính cá nhân cho một lượng lớn người dùng, và mình muốn chatGPT gợi ý cho mình những framework có thể sử dụng để đáp ứng nhu cầu cho ứng dụng của mình.

Việc đầu tiên, cũng chính là việc đơn giản nhất, hãy viết điều bạn muốn, đây được xem như là một "hướng dẫn" cho chatGPT. Nguyên tắc của hướng dẫn là: càng chi tiết càng tốt. Vì vậy, ngoài việc yêu cầu thiết kế kiến trúc cho ứng dụng, mình còn mô tả thêm về chi tiết của ứng dụng mà mình định xây dựng

chatGPT là một mô hình ngôn ngữ đa năng, nhưng đôi khi việc đa năng này là không cần thiết khi hỏi về một ĩnh vực nhất định. Vì vậy, ta cần cho nó một vai trò để giới hạn lĩnh vực của nó.
Điều này cũng tương tự như việc khi ta muốn hỏi về sức khỏe ta sẽ tìm đến bác sĩ hay khi muốn hỏi về luật ta sẽ tìm đến luật sư. Trong trường hợp này, mình sẽ gán cho chatGPT với vai trò là một kĩ sư có nhiều kinh nghiệm về lĩnh vực kiến trúc hệ thống phần mềm.

Đây là phần mà ta sẽ cung cấp nhiều thông tin hơn cho chatGPT, tại phần này mình sẽ cho nó những thông tin về câu trả lời mà nó cần tạo ra, cách suy nghĩ và phân tích,...

Và cuối cùng, hãy định hình cho kết quả mà bạn muốn, bạn có thể mô tả hình thức, cách trả lời, độ dài, mức độ chi tiết,.... Ở đây, mình yêu cầu kết quả phải ở dạng một báo cáo ngắn và mang tính thực tế.

Thông qua một prompt được xây dựng từ các bước như trên, ta có thể dễ dàng nhận được một kết quả như hình dưới, trong đó:
chatGPT liệt kê tất cả những thành phần liên quan đến phần mềm mình cần xây dựng
Với mỗi thành phần, chatGPT cho mình 2-3 lựa chọn cho những framework mà mình có thể sử dụng
Đồng thời, chatGPT cũng phân tích lợi và hại khi mình sử dụng mỗi framework

Nhưng kế sau đó, mình đã đưa ra thêm một yêu cầu, đó là lựa chọn framework tốt nhất cho mỗi thành phần và đưa ra kết quả dưới dạng JSON, và thu được một bảng tóm tắt cực kỳ rõ ràng cho ứng dụng của mình như hình bên phải.
Rõ ràng, mình như từng đề cập gì về JSON hay mô tả JSON có cú pháp như thế nào nhưng chatGPT vẫn biết và đưa ra được kết quả phù hợp, đây chính là một trong những khả năng quan trọng của chatGPT hay LLM gọi là IN-context Learning, điều này giúp cho LLM có thể năng hiểu và thích nghi với những yêu càu tùy vào bối cảnh.
Chi tiết hơn về in-context learning, mình sẽ trình bày rõ hơn trong mục các kĩ thuật Prompt Engineering nâng cao.

Trong ví dụ trên, mình đưa ra cùng một input trong 2 lần yêu cầu, chỉ khác ở chỗ mình để hờ một từ gợi ý ở cuối input để hướng dẫn cho chatGPT về output. Như bạn đã thấy, mình hoàn toàn không hề mô tả hay kì vọng chatGPT sẽ đưa ra "Positive" hay "Negative" trong câu đầu, đó hoàn toàn là chatGPT tự hiểu và thực hiện. Đây chính là một ví dụ cho khả năng thích ứng với bối cảnh: in-context learning.
Kĩ thuật zero-shot là kĩ thuật cơ bản và dễ dàng nhất mà ta có thể sử dụng trong nhiều trường hợp để khiến chatGPT đưa ra output ta muốn

Phát triển từ kĩ thuật zero-shot, nhưng lần này, ta không để hờ yêu cầu nữa, mà ta sẽ thực sự tạo ra một mẫu phản hồi để chatGPT bắt chước, trong ví dụ trên, mình đưa ra một câu hỏi mẫu và liệt kê những câu trả lời theo format mình muốn. Sau khi kết thúc câu hỏi đầu tiên và đưa ra câu trả lời làm ví dụ, mình đặt ra câu hỏi thứ 2 và để chatGPT thực hiện phần còn lại.
Tương tự thế, kĩ thuật few-shot cũng yêu cầu bạn đưa ra mẫu ví dụ, nhưng chỉ khác là bạn có thể đưa ra nhiều ví dụ hơn thay vì 1, điều này sẽ chắc chắn hơn việc chatGPT sẽ tạo ra một câu htrả lời có format tương tự như những ví dụ mà bạn đưa ra cho nó.
Nhưng bạn sẽ thấy rằng, one-shot hay few-shot không phải lúc nào cũng hoạt động đúng, đặc biệt là với những câu hỏi hóc búa và yêu cầu tư duy như con người như ví dụ bên dưới.

Giải pháp cho vấn đề này chính là chain-of-thoughts, để khắc phục vấn đề của ví dụ trên, mình đã giải thích cách suy nghĩ từng bước để tìm ra đáp án. Và như bạn thấy, chatGPT cũng đã thay đổi cách nó đưa ra câu trả lời, và nó cố gắng suy nghĩ và giải thích y hệt như ví dụ mình đưa ra.

Một phương pháp ngắn gọn hơn gọi là zero-shot chain-of-thoughts (sự kết hợp giữa zero-shot và chain-of-thoughts), thay vì mô tả cách tư duy một cách cụ thể và rõ ràng, mình chỉ yêu cầu chattGPT: "Hãy suy nghĩ từng bước". và kết quả nhận được cũng tương tự. Tuy nhiên, cách này chỉ hoạt động trong một số trường hợp đơn giản không cần tư duy quá phức tạp.

Một ví dụ khác cho chain of-thoughts là bạn có thể hướng dẫn cách chatGPT tạo ra một nội dung, trong ví dụ bên dưới, mình đã yêu cầu chatGPT tạo ra một câu chuyện với một chủ đề, nhưng đồng thời cũng cho nó một vài tình huống và thứ tự diễn ra trong câu chuyện, và nhận được một câu chuyện đúng như yêu cầu đưa ra.

Nếu bạn là lập trình viên, bạn sẽ khá quen thuộc với khái niệm function: một hàm chức năng nhập đầu vào, thực hiện những nhiệm vụ nhất định, và đưa ra output. Prompt function cũng tương tự, bạn sẽ hướng dẫn để chatGPt hiểu rằng nó sẽ đóng vai làm một hàm thực hiện chức năng, kèm theo đó, hãy mô tả rõ ràng input, output và parameter mà bạn đưa vào cho chatGPT

Ngoài những kỹ thuật cơ bản đã nêu ở trên, cũng có những kĩ thuật nâng cao hơn nhằm khắc phục nhược điểm và tạo cho LLM những tính năng ưu việc
Cơ chế ngẫu nhiên của chatGPT đôi khi là một vấn đề, nó thường sẽ đưa 10 đáp án khác nhau cho cùng 1 câu hỏi, vì thế ta không thể nào chắc được đáp án đầu tiên ta nhận được là đáp án chính xác. Generated Knowledge Prompting sẽ giải quyết điều đó bằng cách yêu cầu chatGPT tạo ra nhiều câu hỏi lời cho cùng 1 câu hỏi, sau đó sẽ đưa 10 đáp án nhận được vào context và yêu cầu chatGPT lựa chọn phương án phù hợp và đúng đắn nhất.

Một kĩ thuật phát triển từ chain-of-thoughts, nhưng thay vì chỉ có 1 luồn suy nghĩ, bạn sẽ yêu cầu chatGPT phải suy luật theo nhiều luồn suy nghĩ và loại bỏ những tư duy sai lầm. Điều này sẽ khắc phục được vấn đề khi suy luận của chatGPT sai trong chain-of-thoughts sẽ khiến toàn bộ kết quả bị sai.
Phương pháp triển khai kĩ thuật này cũng rất đơn giản, bạn sẽ nói GPT tưởng tượng ra một cuộc hợp có 10 chuyên gia trong lĩnh vực và yêu cầu họ suy luận và đưa ra cách giải quyết, bất cứ người nào nhận ra sai lầm trong suy luận của họ sẽ phải tự giác rời đi. Cuối cùng, người ở lại cuối cùng sẽ đưa ra đáp áp của họ.

Một kĩ thuật được sử dụng cực kì phổ biến trong các ứng dụng LLM-based, Bằng việc nhúng các tài liệu bên ngoài thành dạng embedding vector, ta có thể cho phép LLM truy cập vào một lượng lớn dữ liệu bên ngoài mà chatGPT chưa từng được học trước đó. Với mỗi câu hỏi được đưa vào chatGPT, câu hỏi sẽ được nhungs về dạng vector và so sánh với tất cả những tài liệu khác trong cơ sở dữ liệu để xác định đâu là nguồn tài liệu có khả năng chứa câu trả lời. Sau đó, tài liệu này sẽ được đưa vào prompt dưới dạng context kèm với câu hỏi để đưa được ra câu trả lời cho câu hỏi đó như mong muốn
Bạn có thể dễ dàng bắt gặp phương pháp này trong rất nhiều ứng dụng, tiêu biểu nhất là các ứng dụng tra cứu hoặc trả lời câu hỏi trên các lĩnh vực nhất định.

Mình xin chia sẻ một vài thành phần trong prmopt có thẻ hữu ích trong quá trình bạn sử dụng hay xây dựng ứng dụng với LLM. Cám ơn đã theo dõi
Nếu bạn muốn xem nhiều bài viết về AI của mình hơn, bạn có thể ghé thăm blog của Lưu
Nguồn: https://luu.name.vn/tat-tan-tat-nhung-ki-thuat-prompt-engineering-huu-ich-nhat-cho-chatgpt/
RetoK
RetoK
No activity yet