![Input
[START]
Ich
bin
Student
[END]
[START]
GPT: Transformer
am
a
student
am
a
student
[END]
I
“Shifted“
Output
I
Output
Writes a new
sentence while
preserving content:
Signal when done.](https://image.slidesharecdn.com/chatgptlss-230921105524-74245a25/85/ChatGPT-in-academic-settings-H2-de-18-320.jpg)
ChatGPT in academic settings H2.de
- 1. Large Language Models and How to Tame Them. Using ChatGPT in academic settings. by David Döring AI.Analytics h2.de/ZAKKI
- 2. Poll ● Please answer the questions. ● What do you think about, when you hear ChatGPT? ● Have you used ChatGPT (or related tools) for studying or academics? ● How good were the Results in your opinion? ● Please use the Q&A Tab to collect upcoming questions. ● My Link
- 3. Outline ● What is this about? ● Technobabble ● Application Recommendations ● Risks and Side Effects ● Discussion and Questions
- 4. Outline
- 7. LLM – Large Language Model ● Are AI-Models that produce natural language ● AI-Models are purely statistical functions ● LLMs use Billions of “Parameters” = Large ● LLMs require more calculation power than a single computer can handle ● LLMs usually are Cloud-Services
- 8. ChatGPT creates human-like dialog by predicting weighted word tokens.
- 9. ChatGPT creates human-like dialog by predicting weighted word tokens. Try this out: https://platform.openai.com/tokenizer
- 10. Everyone knows „Chat“, but What does „GPT“ stand for? Generative Pretrained Transformer ???
- 11. GPT: Generative It generates Text. Quite similar to autocomplete.
- 12. GPT: Generative … just a bit more complicated.
- 13. GPT: Pretrained
- 14. GPT: Pretrained ● Pretrained on huge amount of text ● Fine-tuned to Chat ● Improved through rating system ● Adaptable to new applications
- 15. Fine-Tuning OpenAI employees created or gathered many example Dialogs to „teach“ human help-desk communication. Please explain the transformer achitecture. Yes, I could. Could you explain this in simple language? The transformer architecture is a deep learning model that has gained significant popularity and success in various natural language processing (NLP) tasks. It was introduced in the paper "Attention is All You Need" by Vaswani et al., published in 2017. The core idea behind the transformer is the use of self-attention mechanisms to weigh the importance of different words in a sentence, enabling the model to capture relationships and dependencies regardless of their positions within the input sequence.
- 16. Reinforcement Learning Please explain the transformer achitecture. 1: ... 2: ... 3: ... 4: ... 5: ... 6: ... Which Answer do humans prefer? Employees and Users rank and rate answers.
- 17. GPT: Transformer „Attention is All You Need“ – 2017 Google Brain/Research Attention-Layer f(x,y) = ( M × y ) × x Bildquellen: https://www.tensorflow.org/text/tutorials/transformer https://arxiv.org/abs/1706.03762 Self-Attention f(x) = ( M × x ) × x casual Self-Attention
- 18. Input [START] Ich bin Student [END] [START] GPT: Transformer am a student am a student [END] I “Shifted“ Output I Output Writes a new sentence while preserving content: Signal when done.
- 19. GPT – What does it do? ● handles text as tokens ● predict most likely token ● requires context ● generates text that look and feel correct
- 20. GPT – What does it NOT do? ● Use the internet for correct answers. ● Fact-Check ● obtain up-to-date information ● Actually perform given tasks like – „calculate“ – „research“ – „imagine“
- 21. How can we use it anyway? That doesn’t sound usefull for academics. So...
- 22. GPT as a product is an API … API: Application Programming Interface Bing Chat Microsoft Copilot
- 23. Use the correct tool for the job! ● ChatGPT on chat.openai.com is likely not the tool you need ● Many free tool use GPT-3.5 ● Many paid tools use GPT-4.0 ● Use tools that can reference sources. ● Prefer tools that protect your data rights.
- 24. What can we use LLMs for? Getting started / Beating Procrastination Research and Summaries Gathering Ideas Programmierung & Formatierung Improvement Text to Presentation Review Roleplay (e.g. Exam situation) Translation
- 25. What are we allowed to use?
- 26. What are we allowed to use? ● The use of AI-generated text in your work is permitted. ● Use the official Declaration of Originality from H2 (EN; DE) ● mark AI-generated or enhanced content ● reference with "supported by …” ● include section “overview of resources used” ● how you used these tools, including date and settings ● review AI-generated content. ● Mistakes made by using LLMs are your mistakes!
- 27. Known Problems Bias and Stereotypes lack of education loss of basic skills „Hallucinations“ confidently incorrect Traceability unclear Quoteability complicated Data Protection Inequality Financial Burden Lack of Transparency Rights and Regulations unclear Add more of your own context. Ask for Reasoning. Fact Check. Use tools with added context. Start quick but spend more time. Check the FAQs of Apps. Spend money for better results. Support Open-Source.
- 28. Better Results Repeat Stuff Work Interactively Add Knowledge and Context Improve your Prompt ● GPT-3.5 Context Size: 4,097 tokens ● “short memory” ● Important knowledge needs to be repeated ● Start with known good Prompts ● Github “Awesome ChatGPT Prompts” ● Roleplay. Start by describing it’s role. ● Provide information from citeable sources. ● Use Tools that Automate that. ● Perplexity.AI, ChatPDF.com, Bing Chat, Bard … ● Repeat and refine requests ● Use tools that allow reworking and editing ● Notion.so, Microsoft Office 365, …