Among the obvious things with ChatGPT
- Zero-shot: no structure provided in the question. It solves it on its own
- One-shot: one step but can be difficult to follow
- Few-shot: roadmap or steps to solve the problem
- Style / Audience / Length / Tone
- Four corner stones of evaluation: LARF: Logical consistency, Accuracy, Relevance, Factual correctness
- Use “—” for distinct points
- Use “” quotes for specifc words
- Define a Chain of thoughts (CoT): breaks down the problem into smaller steps and makes it easier for chatGPT
- Can you provide sources? and really think about it_
- Avoid discontinuous tasks between prompts
- Person-problem framework: Instruction, Persona, output format, Context and examples
- Keep it mind how old the training set was for the version of the AI you are using
- 2017 transformers were introduced y Google with attention mechanism that weight the significance of each words, irrespective of its position. Used previously Recurrent neural Networks (RNN) with hidden states to retain memory, but ran into vanishing gradient issues
- Customise your output from the prompt by giving specific instructions
- The term Ai was coined in 1956 in the Dartmouth conference
- AI ) ML) NN) DL
- Ai encompasses all
- ML: Machine learning (Reinforcement learning, supervised learning, classification regression, unsupervised learning for anomaly detection clustering and association, clustering with Kmeans DBSLAN etc..
- DL: deep learning
- NLP: natural language processing with bag of words, word embeddings capturing meaning as numbers
- LLM: Large language models with tokenzation, stop word removal, lemmatization (group word with the same meaning)
- ANI: Artificial narrow intelligence
- AGI: Artificial general intelligence
- XAI: Explainable AI
- Transparency, fairness, accountability. Train models without race, gender, age, socioeconomic statistics, sexual orientation, religion etc…
- LIME (Local interpretable Model agnostic Explanation)
- Use feature importance (SHAP: Shaply additive explanation)
- Regulations
- EU regulations: GDPR, DSA, EU Ai act
- US: Executive order on AI
- Canada: Bill c27 and code of conduct
- UK: Bletchely declaration
- China: Multiple laws
- Voluntary guidelines
- OECD: Ai principles
- GPAI: experts and guidlines
- ISO standards (IEC42001)
- Main processes:
- Text processing
- Text representation
- Pre-training requiring a attention mechanisms (self attention or multi-head)
- transformers with relationships between words, processing multiple words at the same time
- pre-processing
- positional encoding
- encoders
- decoders
- Fine tuning
- Advanced fine tuning
- RLHF: Reinforced learning through human feedback (for accuracy, relevance and coherence)
- GAN: 2 model competing (2014) with feedback loop known as adversarial training
Maximising on AI
- Finding something unique, not code or architecture, the code is the vehicle as you are not safe from copycat
- Focus on the data as uniqueness, data is the differentiation
- You need data with context
- you need intentional data collection
- Data supports multiple implementation, never taps out
- Seek high value, non-obvious opportunities, products are generated not coded
- Focus on the data the business has easy access to
- Create a vision
- find the opportunity that will be ahead in 2,3, 5 years
- you need to look ahead
- the internet of things
- where will data come from ?
- Position yourself where the data will be (ex: streaming real time data from 5G from places that were not data centers before)
- Align on strategy first
- Why use data and AI for that
- Reliability?
- Maybe go for incremental improvement
- Estimate ROI and value of time saving
- Consistent delivery, break down in small increments, start simple, use a maturity model
- Watch out for costs