| A Series of Distinctions |
| Steps: Visit https://ollama.com/ Download for your platform (Windows, macOS, Linux) Open Command Line (Windows) or Terminal (macOS, Linux) Enter: ollama run llama3.2:3b ollama run deepseek-r1:8b ollama run phi4 Wait! Language models take time to download. Then use your local LLM… |
| Artificial Intelligence |
Natural Intelligence? Inspirations: Philosophy of Mind, Cybernetics, Psychoanalysis, Childhood Psychology, Neural & Cognitive Science, Literature, Drama, Game Theory… But is there consensus on what intelligence is? |
| Artificial Intelligence |
Artificial Stupidity? “It is claimed that the interrogator could distinguish the machine from the man simply by setting them a number of problems in arithmetic. The machine would be unmasked because of its deadly accuracy.” — Turing, 1950, “Computing Machinery and Intelligence”. “Artificial Intelligence has achieved super-human performance in some tasks, such as arithmetic or games; in this article we argue that sometimes AI’s ability might need to be artificially constrained. Such deliberate limiting is called Artificial Stupidity.” * – Trazzi, M., & Yampolskiy, R. V. (2020). Artificial stupidity: Data we need to make machines our equals. Patterns, 1(2)* For a Philosophy of Stupidity / Bêtise: Avital Ronell, Bernard Stiegler. |
| Symbolic Reasoning |
Connectionist Reasoning |
| Symbolic Reasoning - Deduction All men are mortal Socrates is a man Socrates is mortal Modus Ponens All A are B (or “if A, then B”) C is A Therefore C is B ) Example: Semantic Web; “Reasoning” Language Models? |
Connectionist Reasoning - Induction If “men” and “mortal” appear in the same sentence, assume they are connected (semantically, grammatically etc) Freudian free association (some meaningful relationship due to contiguity: ”It is a rule of psycho-analytic technique that an internal connection which is still undisclosed will announce its presence by means of a contiguity—a temporal proximity—of associations; just as in writing, if ‘a’ and ‘b’ are put side by side, it means that the syllable ‘ab’ is to be formed out of them.” – Freud, S. (1997). Dora: An Analysis of a Case of Hysteria. Simon and Schuster. Markov Chains - relationship between letters, next letter prediction Markov, A. A. (2006[1913]). An example of statistical investigation of the text Eugene Onegin concerning the connection of samples in chains. Science in Context, 19(4), 591-600. Example: ChatGPT |
| Supervised Learning Statistics: Linear Regression, Correlations. Is this Hypothesis True? Machine Learning: classify this text / image Discriminative (this or* that – cat ****or**** dog)* |
Unsupervised Learning Statistics: Factor Analysis, Principal Component Analysis – What’s in this Data? Machine Learning: predict or generate next token Generative (what comes next? Next token prediction) |
| Trained Data from a specialized field or domain Train a model to understand the data Test model prediction reproduces the training data Do not expect generalization (to other data sets) |
Pre-trained Very large data set from many domains / fields / languages 2. Train a model to understand general patterns in the data 3. Test model prediction reproduces the training data 4. Do expect generalization (to other data sets) 5. Fine-tune for specialization – or other techniques |
| Trained Data from a specialized field or domain Train a model to understand the data Test model prediction reproduces the training data Do not expect generalization (to other data sets) |
Pre-trained Very large data set from many domains / fields / languages 2. Train a model to understand general patterns in the data 3. Test model prediction reproduces the training data 4. Do expect generalization (to other data sets) 5. Fine-tune for specialization – or other techniques … Generalized Pre-Training model = “GPT” for short |
| Base model (GPT-3 - 2020) “as-is” But it is biased, unaligned with human values, so… |
Instruct Model (GPT-3.5, ChatGPT - 2022) 1. Take the base model, and generate many statements 2. Have humans score those statements on criteria: Helpful, Truthful, Harmless 3. Re-train the model, providing “rewards” as its outputs measure closer to “good” statements Ouyang, et al (2022). Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35, 27730-27744. Reinforcement Learning - with Human Feedback (RLHF) - with AI Feedback (RLAIF) |
| In-context “training” Fine-tuning Or give examples in a prompt No examples = Zero-shot learning 1 example = One-shot learning Few examples = Few-shot learning GPT-3: Brown, T., et al. (2020). Language models are few-shot* learners*. Advances in neural information processing systems, 33, 1877-1901. |
Retrieval-Augmented Generation I want to know what I was thinking when I was happy… But GPT doesn’t know my style / facts about my life etc. Fortunately my life is recorded in Dropbox… Can I (a) search Dropbox for happy memories, and add these as context to GPT? Steps: Do Semantic Search Retrieve relevant excerpts Add these excerpts to my prompt GPT can answer due this new context Note: relevant for Literature Reviews, Qualitative Data Analysis!! |
| Closed “Open”AI? GPT-1, GPT-2: open source GPT-3 onwards: closed source Anthropic’s Claude Google Gemini |
Open GPT-NEO Llama (Meta) Mistral DeepSeek Qwen (Alibaba) Phi (Microsoft) |
| Remote Large models Data is online - may be used for training (fails IRB tests) Monthly subscription Additional features Examples: ChatGPT, Anthropic’s Claude, Google Gemini, DeepSeek, OpenRouter |
Local Small models only Data is offline - private, secure, Hardware costs (Macbook or NVidia GPUs) Relies upon open source software Examples: Llama, DeepSeek, Phi, etc |
| Chat Web Interface Non-technical Less customization Pay by month |
API Code Interface (Python / JavaScript) Some coding required Can be integrated into apps, websites, devices, drones etc Pay by token |
| Regular Model (OpenAI GPT-4o; DeepSeek-v3) Expensive Training Cheap Inference Next token prediction |
Thinking Model (OpenAI o1, o3; DeepSeek-R1) Expensive Training Expensive Inference (“Thinking”) Next token prediction… Then review…. Then predict again Kahneman, D. (2011). Thinking, fast and slow. Farrar, Straus and Giroux. |
| Free vs Paid Text vs Multimodality (Images, video) Geopolitics of AI: US / China / EU / ??? Need to stay on “cutting edge” for research / work? Consider access to ChatGPT Plus (perhaps Claude / Google Gemini) |
Tools Web Search Voice interface Code Execution Computer control (Agents) Task reminders Project organization Collaboration Deep Research Memorization / Personalization |
| Large Language Models OpenAI ($20 per month (Plus) / $200 per month (Pro)) Anthropic Claude Google Gemini DeepSeek QwenChat Meta.ai Grok (x.com) Microsoft Co-pilot Mistral |
Tools Search: Perplexity Summarization: NotebookLM Research Assistance: Google DeepResearch / OpenAI DeepResearch Multiple AI models: Poe / OpenRouter (CGScholar) Image Generators: Midjourney Voice: ElevenLabs Video: Kling, Luma Labs |
| Free vs Paid Text vs Multimodality (Images, video) Geopolitics of AI: US / China / EU / ??? Legality (can DeepSeek be used? downloaded?) Model augmentations (code execution, project management, collaboration, etc etc) & associated effects of vendor lock-in |
Need to stay on “cutting edge” for research / work? Consider access
to ChatGPT Plus (perhaps Claude / Google Gemini) |
| Prompting How do we talk to AI? “Explain it to me like I’m 5 years old” “Here’s an example” “Here are a series of documents. Compare and contrast”. “Develop a structure / conceptual map for my literature review…” “Review my draft and give me critical feedback…” “Here’s my draft. I’ve got writer’s block. Give me some inspiration.” “Here’s my first draft, some feedback, and follow-up draft. Have I addressed everything?” |
Role-playing in Research Who is the AI? “I am an expert / a student“ Your Co-pilot / Co-worker / Colleague? Your (Anonymous) Critic? Your Student? Your Advisor, Mentor or Boss? Your Research Assistant? Your Research Participant? Which models respond best to different roles? |
| Prompting What do you find works well? |
AI In General What do you think will happen next? |