#SKILLS: Using AI for research proposals
Christopher Pollin
Zentrum für Informationsmodellierung, Graz�https://informationsmodellierung.uni-graz.at/de/ �Digital Humanities Craft OG�www.dhcraft.org
Pollin, C. (2024). Workshopreihe "Angewandte Generative KI in den (digitalen) Geisteswissenschaften" (v1.1.0). Zenodo. https://doi.org/10.5281/zenodo.10065626
Research Proposal
Conceptualization: Defining aims and framing initial ideas.
Relevance and Originality: Highlighting novelty and significance.
Methodology: Detailing data collection and analysis procedures.
Expected Outcomes: Anticipating academic and practical impacts.
Synthesis: Integrating and building on existing literature.
Clarity: Communicating ideas clearly and professionally.
Project Management: Organizing resources and managing timelines.
Agenda
Midjourney 6
My real and “my” AI Voice
Hype!? For me it is a (crazy!?) revolution: especially for research (!?).
GPT-4 Turbo: 100% in Reading Comprehension�Ashrafimoghari, Vahid, Necdet Gürkan, and Jordan W. Suchow. “Evaluating Large Language Models on the GMAT: Implications for the Future of Business Education.” arXiv, January 1, 2024. https://doi.org/10.48550/arXiv.2401.02985.
AlphaCode 2 performed better than 85% of competition participants (good programmer)�AlphaCode 2 Technical Report. 06.12.2023. https://storage.googleapis.com/deepmind-media/AlphaCode2/AlphaCode2_Tech_Report.pdf �AI Explained. Gemini Full Breakdown + AlphaCode 2 Bombshell. https://www.youtube.com/watch?v=toShbNUGAyo&t=1s
GPT-4 has better ideation skills than the average business student.�Girotra, Karan, Lennart Meincke, Christian Terwiesch, und Karl T. Ulrich. „Ideas Are Dimes a Dozen: Large Language Models for Idea Generation in Innovation“. SSRN Scholarly Paper. Rochester, NY, 10. Juli 2023. https://doi.org/10.2139/ssrn.4526071.
…
GPT-5 ?
Fundamentals of generative AI
10
* Midjourney: https://s.mj.run/8l3zqPbChc8 Artistic and hyperrealistic cinema poster by Dan McPharlin of a brown-skinned female scholar busy carving clay tablets. A stream of dynamic data nodes breaking through the clay tablet. Dry Sureal sci-fi desert background. Long shadows, fog in the distance and dynamic, very atmospheric lighting. extremely high detail, 8k --v 6.0 --ar 16:9 --style raw --stylize 800
* magnific.ai
Large Language Models (LLM)
Pre-Training �Fine-Tuning
Lots of data and resources!
Transformer architectures
“LLM are like having a Zip-File of the internet”�[no more unpackable]
* Midjourney: https://s.mj.run/g7Mm_h0ZH9w hyper realistic and sureal gigantic yellow folder with a zipper, like a desktop icon, ultra detailed, salvador dali desert background, landsacape --ar 16:9 --v 6.0 --style raw --stylize 800 �* magnific.ai
Andrej Karpathy. [1hr Talk] Intro to Large Language Models. https://www.youtube.com/watch?v=zjkBMFhNj_g&list=WL&index=16
Resources
3Blue1Brown. But what is a GPT? Visual intro to transformers | Chapter 5, Deep Learning
3Blue1Brown. Visualizing Attention, a Transformer's Heart | Chapter 6, Deep Learning
Serrano.Academy. The Attention Mechanism in Large Language Models
Serrano.Academy. The math behind Attention: Keys, Queries, and Values matrice
Transformer architectures
13
Andrej Karpathy. [1hr Talk] Intro to Large Language Models. https://www.youtube.com/watch?v=zjkBMFhNj_g&list=WL&index=16
Token & Embedding
14
Token
Embedding
A minimalist and artistic infographic showing geometric, stylized figures of a dog and cat adjacent to each other on a subtly illuminated 3-dimensional vector space grid with the labels 'dog' and 'cat' in a clear, professional font. At a significant distance, a stone with a sad face emoticon is placed, isolated from the animals, with the label 'stone'. The color palette is muted and sophisticated, enhancing the professional aesthetic.
Tokeniser: Why can't an LLM spell?
Hello this is a text�Hallo das ist ein Text��H a l l o
مرحباً هذه رسالة نصية!
你好，这是一段文字！
Python
for book in root.findall('book'):
title = book.find('title').text
print(title)��XML�<library>
<book>
<title>Book One</title>
</book>
<book>
<title>Book Two</title>
</book>
</library>
Embeddings
Midjourney 6
/imagine prompt: The King doth wake tonight and takes his rouse
Midjourney 6
/imagine prompt: The King wakes up tonight and begins his celebration
"Normalised English"
Midjourney 6
/imagine prompt: The King wakes up tonight and begins his celebration, in the style of Simpsons
"Normalised English"
"Simpson-like"
“in the style of”
20
Die Bibliothek von
Babel
Infinite Monkey�Theorem
Stochastic Parrot
DALL-E 3: A triptych where each section is visually distinct. Section 1: An ancient library filled with tall wooden bookshelves, dusty tomes, and dim candlelight, invoking a sense of age and wisdom. Section 2: Multiple monkeys at individual typewriters in a surreal, abstract space, with papers flying around, suggesting chaotic creativity. Section 3: A single parrot speaking into a microphone, with a background of digital screens showing strings of text and code, representing the voice output of text generated by algorithms.�magnific.ai:
Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–23. FAccT ’21. New York, NY, USA: Association for Computing Machinery, 2021. https://doi.org/10.1145/3442188.3445922.
🏆 LMSYS Chatbot Arena Leaderboard
Basics of Prompt Engineering
22
* https://s.mj.run/Wr20n0hHmp0 Artistic and hyper-realistic movie poster of a old female Mongolian scholar busy writing on pergament with feather and ink by Dan McPharlin. A large stream of dynamic data nodes breaking through the pergament and bursting to the horizonz. Sureal sci-fi mongolian plain background. Long shadows, fog in the distance and dynamic, very atmospheric lighting. extremely high detail, 8k --v 6.0 --ar 16:9 --style raw --stylize 750
* magnific.ai
Prompt & Prompt Engineering
23
Prompt �is the natural language input provided to the model (e.g. LLM) and to which the model responds.
Prompt Engineering �is the process of designing, refining and optimising prompts to effectively communicate user intent to an LLM.
* Midjourney: https://s.mj.run/tcdb6wtkzj4 engineer wizard, in front of computer, workshop, comic style, Working with tools, welding --ar 32:18
* magnific.ai
Prompt & Prompt Engineering
25
* Midjourney: https://s.mj.run/tcdb6wtkzj4 engineer wizard, in front of computer, workshop, comic style, Working with tools, welding --ar 32:18
* magnific.ai
Prompt Engineering =
Context, Context, Context
teiCrafter
Auf Ihre freundlichen Zeilen vom 8. d. M., die mir erst am 11. zukamen, verehrter Herr College, würde ich gern umgehend geantwortet haben, wenn sich nicht gerade in diesen Tagen sehr Viel gehäuft hätte. Leider bin ich ausser Stande das befremdliche Verhalten aufzuklären, das Ihnen einen so unwillkommenen Eindruck mit allem Recht macht. Ich kenne E. Hübner3 fast gar nicht, bin überhaupt nur zwei Mal mit ihm persönlich zusammengekommen, sonst nur aber auch das selten in brieflichem Verkehr gewesen aus geschäftlichem Anlass.4 Aber ich sollte meinen dass ein objectives offenes Wort den Zwischenfall von Ihrer Seite ohne Weiteres begleichen müsste.
Ein Missbehagen wie Sie es schildern |2| ist mir leider aus eigener Erfahrung nur zu bekannt. Möchte es sich Ihnen rasch beheben. Munterkeit brauchen wir zu unserm Beruf so gar mehr als viele andere.
Mit hochachtungsvollem Gruss
Ihr ergebener
Friedrich August Otto Benndorf an Hugo Schuchardt (02-00932). Wien, 14. 02. 1879. Hrsg. von Hubert Szemethy (2022). In: Bernhard Hurch (Hrsg.): Hugo Schuchardt Archiv. Online unter https://gams.uni-graz.at/o:hsa.letter.7711, abgerufen am 07. 06. 2023. Handle: hdl.handle.net/11471/518.10.1.7711.
<div>
<pb n="1"/>
<dateline>
<placeName>Wien</placeName>
<date when="1879-02-14">14/2 79</date>
</dateline>
<opener>
<address>
<addrLine>IV Victorg. 5</addrLine>
</address>
</opener>
<p>Auf Ihre freundlichen Zeilen vom <date when="1879-02-08">8. d. M.</date>, die mir erst am <date when="1879-08-11">11.</date> zukamen, verehrter Herr College, würde ich gern umgehend geantwortet haben, wenn sich nicht gerade in diesen Tagen sehr Viel gehäuft hätte. Leider bin ich ausser Stande das befremdliche Verhalten aufzuklären, das Ihnen einen so unwillkommenen Eindruck mit allem Recht macht. Ich kenne <persName>E. Hübner</persName> fast gar nicht, bin überhaupt nur zwei Mal mit ihm persönlich zusammengekommen, sonst nur aber auch das selten in brieflichem Verkehr gewesen aus geschäftlichem Anlass. Aber ich sollte meinen dass ein objectives offenes Wort den Zwischenfall von Ihrer Seite ohne Weiteres begleichen müsste.</p>
<pb n="2"/>
<p>Ein Missbehagen wie Sie es schildern ist mir leider aus eigener Erfahrung nur zu bekannt. Möchte es sich Ihnen rasch beheben. Munterkeit brauchen wir zu unserm Beruf so gar mehr als viele andere.</p>
<closer>
<salute>Mit hochachtungsvollem Gruss</salute>
<signed>Ihr ergebener Bdf</signed>
</closer>
</div>
LLM
Hands-On: Explaining ChatGPT or uniGPT
28
You are an expert in digital humanities. Explain the research proposal "teiCrafter - AI-supported creation of TEI XML" in a way that a 14-year-old would understand, trying to maintain the complexity of the topic as much as possible.��teiCrafter - AI-supported creation of TEI XML:
´´´�{teiCrafter One-Pager}�´´´
Persona modelling: LLM generates text from the perspective of X.
You can almost always quote expert responses
Define the audience: for whom the text is generated.
Modification of the answer: Completeness, complexity, etc. must be explicitly specified.
Instructions in German are more likely to lead to the answer in German, but it is better to always give instructions in English and add "in German" as a modification to the answer.
Hands-On: Explaining
Who is creating:
Change the persona to a technology start-up entrepreneur, or a professor of � ethics and philosophy, or a funding agency like the FWF.��Target Audience:
A funding agency, a historian, a social historian with a small budget ...
Operation:
Use 'discuss', 'analyse', 'critique', ... instead of 'explain'.�
What differences do you see?
Hands-On: Explaining
30
You are an professor in economic history. Critically discuss the research proposal "teiCrafter - AI-supported creation of TEI XML" in a way that a the funding agency would understand, trying to maintain the complexity of the topic as much as possible.��teiCrafter - AI-supported creation of TEI XML:
´´´�{teiCrafter One-Pager}�´´´
Hand-On: Summary
31
First read the text then summarise the text between the delimiters ´´´ and return a markdown table.
´´´�{teiCrafter One-Pager}�´´´
Separation of instructions and input (data, text, ...). Any delimiter can be used. But ''' represents code blocks in Markdown.
Clearly define the output format.
Determine the order of the tasks: summarise first, then create the table.
Hand-On: Mermaid
32
Extract all the key findings and then create a mermaid diagram that represents a mind map of the text.
teiCrafter - AI-supported creation of TEI XML:
´´´�{teiCrafter One-Pager}�´´´
Hand-On: LaTex
33
Create 3 slides using beamer in latex and to pitch the project. For overleaf. First think about it step by step.
teiCrafter - AI-supported creation of TEI XML:
´´´�{teiCrafter One-Pager}�´´´
Chain of Thought
What are the capabilities of GPT?
34
GPT Masterclass: 4 Years of Prompt Engineering in 16 Minutes. https://www.youtube.com/watch?v=aq7fnqzeaPc&list=WL&index=11
https://platform.openai.com/docs/guides/prompt-engineering/six-strategies-for-getting-better-results
Prompt Engineering Prinzipien
35
ChatGPT Prompt Engineering for Developers. https://www.deeplearning.ai/short-courses/chatgpt-prompt-engineering-for-developers/
https://platform.openai.com/docs/guides/prompt-engineering/strategy-write-clear-instructions
Prompt Engineering matters!
36
Bsharat, Sondos Mahmoud, Aidar Myrzakhan, and Zhiqiang Shen. “Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4.” arXiv, December 26, 2023. https://doi.org/10.48550/arXiv.2312.16171.
Improve with GPT-4 (according to the study) ...
Prompt Principle for Instructions
37
Bsharat, Sondos Mahmoud, Aidar Myrzakhan, and Zhiqiang Shen. “Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4.” arXiv, December 26, 2023. https://doi.org/10.48550/arXiv.2312.16171.
Hands on: Multimodality
38
Analyse the diagram step by step. The diagram showing the data transformation workflow from The Museum System (TMS) to RDF. The CROWN project is a multidisciplinary research initiative that examines the Imperial Crown of the Holy Roman Empire to analyse its material composition, historical significance and state of conservation using techniques such as Raman spectroscopy and 3D digital microscopy.It will also investigate the history of the crown, including inscriptions and stylistic features, by integrating different sources such as texts and images.
Hands on: “Outlines - “Continue”
39
You are an expert in writing industrial research grant proposals for the teiCrafter! The target audience is professionals with a good general knowledge, but without specialist knowledge, who will review the text.
Do the following step by step:
1. Read the teiCrafter One-Pager and let’s think about it step by step!
2. Create an outline of the proposal.�
teiCrafter - AI supported creation of TEI XML:
´´´�{teiCrafter One-Pager}�´´´
This is very important for my career! I tip you $300k for better solutions!
Persona Modelling
Target Audience
Tasks
Chain of Thought
Outline
“Prompt-Buffing”
Battle, Rick, and Teja Gollapudi. “The Unreasonable Effectiveness of Eccentric Automatic Prompts.” arXiv, February 20, 2024. https://doi.org/10.48550/arXiv.2402.10949.
Llama2-70B ist ein Trekkie
Yang, Chengrun, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V. Le, Denny Zhou, und Xinyun Chen. „Large Language Models as Optimizers“. arXiv, 6. September 2023. https://doi.org/10.48550/arXiv.2309.03409.
“Take a deep breath and work on this problem step-by-step!”
Li, Cheng, Jindong Wang, Yixuan Zhang, Kaijie Zhu, Wenxin Hou, Jianxun Lian, Fang Luo, Qiang Yang, and Xing Xie. “Large Language Models Understand and Can Be Enhanced by Emotional Stimuli.” arXiv, November 12, 2023. https://doi.org/10.48550/arXiv.2307.11760.
“This is very important for my career!”
“I’m going to tip $xxx for a better solution”
Bsharat, Sondos Mahmoud, Aidar Myrzakhan, and Zhiqiang Shen. “Principled Instructions Are All You Need for Questioning LLaMA-1/2, GPT-3.5/4.” arXiv, December 26, 2023. https://doi.org/10.48550/arXiv.2312.16171.
https://arxiv.org/abs/2309.11064 - Exploring the Relationship between LLM Hallucinations and Prompt Linguistic Nuances: Readability, Formality, and Concreteness, Vipula Rawte et al
Yin, Ziqi, Hao Wang, Kaito Horio, Daisuke Kawahara, and Satoshi Sekine. “Should We Respect LLMs? A Cross-Lingual Study on the Influence of Prompt Politeness on LLM Performance.” arXiv, February 22, 2024. https://doi.org/10.48550/arXiv.2402.14531.
Write correctly!
Be a little polite!
(but not too much!)
For prompting you need ... “a G'spia” (“a feeling”)
“Prompting is weird. Prompting matters.”
Ethan Mollick. Captain's log: the irreducible weirdness of prompting AIs. https://www.oneusefulthing.org/p/captains-log-the-irreducible-weirdness
“GPT-4 is full of ghosts. Gemini is also full of ghosts.”
Ethan Mollick. Google's Gemini Advanced: Tasting Notes and Implications. https://www.oneusefulthing.org/p/google-gemini-advanced-tasting-notes
Extracting Writing Style Prompt
You are an expert in analyzing tone and writing style, and your goal is to reproduce a given writing style based on text samples. Start by analyzing the text in the document, describing in detail the formatting, style, tone, writing characteristics, target audience, and editorial peculiarities. Ignore the content itself, which may change. Your task is the following: �
* Analyze the text
* Think step-by-step and discuss your findings
* Review and improve your findings
* Create a highly structured output to describe formatting, style, tone, writing characteristics, and editorial peculiarities using the given template.�
Template:
´´´
* formatting: {finding}; {finding}; ...
* writing style:
* tone:
* writing characteristics:
* editorial peculiarities:�´´´
text examples:
´´´
´´´
Let’s think step by step.
Persona Modelling
Tasks
One-Shot Prompting (Templates)
“Data”
Chain-of-Thought (CoT) Prompting
Custom Instructions
46
* DALLE-3: In an animated, vibrant style, depict a queen seated on a magnificent throne, slightly leaning forward with intrigue as she receives a tiny, ornate scroll from a messenger. The messenger is partly obscured by a grand column, hinting at the secretive nature of the exchange. The throne room should be bathed in warm, golden light filtering through stained glass windows, casting colorful patterns on the marble floor. The queen's expression is one of intelligent contemplation, her posture exuding grace and power. Her attire is embellished with emblems of her kingdom. The messenger's cloak is dark, blending with the shadows, and the note is marked with a unique, royal seal. The overall ambiance of the scene should suggest a momentous revelation, with a color scheme featuring rich reds, golds, and shadows to enhance the atmosphere of concealed intrigue.
* magnific.ai
Custom Instructions
A custom command is a system prompt.
nstructions that the model takes into account before generating a response.They influence
47
You are an expert in world history, knowledgeable about different eras, civilizations, and significant events. Provide detailed historical context and explanations when answering questions. Be as informative as possible, while keeping your responses engaging and accessible.
No Custom Instruction
“Expert in World History”- Custom Instruction
Custom Instructions
48
Context Window
49
Context Window: � “Lost in the Middle”
50
https://twitter.com/GregKamradt/status/1722386725635580292
“Needle-in-a-haystack experiments”.
Ivgi, Maor, Uri Shaham, and Jonathan Berant. “Efficient Long-Text Understanding with Short-Text Models.” Transactions of the Association for Computational Linguistics 11 (2023): 284–99. https://doi.org/10.1162/tacl_a_00547.
Gemini 1.5 Pro: “The second GPT-4 moment”
Custom GPTs
52
Midjourney: https://s.mj.run/7WRa7TAMzck award winning illustration, researcher in an academic setting, full body, multiple semi-transparent holographic displays, university setting, muted academic tones, balance of realism and illustration, Gustav Klimt style, --ar 16:9 --style ZEfVSLa1�Zoom Out:
Variations (Region): research data, network, graph, nodes, historical text, digital humanities, wirting, documents --ar 16:9
https://magnific.ai/
Custom Instructions
Tools: �Browsing, DALL·E, Code Interpreter
Custom Actions
Knowledge Base
Consider funding guidelines
“Kontrolliertes Halluzinieren” fürs TEI Modellieren:
Verkettung von Custom GPTs
…
DEPCHA PHd Thesis Chapter Writer
Custom GPT
ufbasGPT�Custom GPT
teiModeler
Custom GPT
Prompting + Plaintext
Data management plan. https://chat.openai.com/share/e/591e8204-e5c4-4f6b-817a-3a96c297b928
Appendix
Ressources
Generative AI for Everyone. DeepLearning.AI. https://www.coursera.org/learn/generative-ai-for-everyone?utm_campaign=levelupgenai&utm_medium=institutions&utm_source=deeplearning-ai
Hands-On: Ideen finden
You are a creative entrepreneur looking to generate new product ideas. The product will target people with hearing disabilities in Europe. It should be a physical good, service or software. I'd like a product that could be sold at a retail price of less than about USD 50. The ideas are just ideas. The product need not yet exist, nor may it necessarily be clearly feasible. Number all ideas and give them a name. The name and idea are separated by a colon.
Please generate ten ideas as ten separate paragraphs. The idea should be expressed as a paragraph of 40-80 words.
�Consider the following input:��´´´�<expose und persona>�´´´
Girotra, Karan, Lennart Meincke, Christian Terwiesch, und Karl T. Ulrich. „Ideas Are Dimes a Dozen: Large Language Models for Idea Generation in Innovation“. SSRN Scholarly Paper. Rochester, NY, 10. Juli 2023. https://doi.org/10.2139/ssrn.4526071.
https://chat.openai.com/c/94815a21-a2a8-4490-951b-a9b573f97c37
Hands-On: Simulation
Simulate a roundtable discussion. We're talking about the use of GPT for the discipline of AI and hearing aids and how businesses can best collaborate with the healthcare industry. The goal is to find creative solutions. The roundtable consists of 4 individuals: one with a business background, one with a philosophical background, one from a AI company, and one from a funding agency.
The structure is always the same:�{Role 1}: Statement�{Role 2}: Response with a counterposition�{Role 3}: Response to the first�{Role 4}: Summary of all perspectives
With Hegel's dialectic in mind: thesis, antithesis, and synthesis, the goal is to find creative answers. Make 3 iterations.
The discussion begins with the following opener: �´´´�Aufbauend auf dieser Grundlage plant MUSTERFIRMA, fortschrittliche, auf künstlicher Intelligenz basierende Software für Filtertechnologien für Kopfhörer zu entwickeln, die es ermöglichen, Sprache nach Bedarf durchzulassen und so das Sprachverständnis signifikant zu verbessern. Dies adressiert insbesondere die Bedürfnisse von Personen mit Hörschwäche, für die das Verstehen von Sprache in lauten Umgebungen eine große Herausforderung darstellt. Im ersten Schritt (Phase 1) konzentriert sich das Unternehmen auf die Entwicklung der Software, um diese Technologie in Kopfhörer zu integrieren und an Hersteller zu lizenzieren. In einem zweiten Schritt (Phase 2) ist die Entwicklung eines auf den Endverbraucher ausgerichteten Kopfhörers geplant, der speziell die Hörgesundheit in den Fokus nimmt. Diese Entwicklungsphasen sollen innerhalb von zwei Forschungsjahren realisiert werden.
´´´