Notes on The Turing Lectures: "What is generative AI?"

 The Turing Lectures: What is generative AI?

Generative - create new content (audio, code, images, text, video)

Artificial Intelligence - automatically using a computer program "a lot of myth about it, its just a tool", "we have to spell it out"

Generative AI is not a new concept. Google translate (17 yo) is an example of generative AI.

Another example is Apples Siri (11yo), another generative AI. Phone autocomplete or Google search autocomplete

Not that new so "what is the fuss"? Answer: chatGpt was launched. chatGpt is a lot more sophisticated than older AIs. You can have a conversation.

The technology is not extremely new. It uses language modeling. Given a context, the language model can predict

e.g. context: "I want to"

  • prediction: play, eat, shovel


AIs used to count to predict but latest use Neural Networks. Feed context (prompt text) into neural language model which predicts the answer.

Basic steps to build a language model:

  1. Step 1. Collect very large data (aka "corpus")  e.g. the web (wikipedia books, github, stack overflow)
  2. Step 2. Ask language model to predict the next word in a sentence
    • randomly truncate last part of input sentence, calculate probability of missing works, adjust and feedback to the model
    • e.g. "the trevi fountain is in ??"  Rome - good  Berlin - bad
  3. Step 3. Repeat over whole corpus. Keep going over months.
    • So it learns over time by practicing and correcting.


Neural Network Language Models

Parameters = # input units * other nodes


Transformers - "king of AI architectures"

 - uses mini neural networks


gpt: "generative pretrained transformer"

pre-training how is it done?

  • hire expert trainers who train the AI (correct answers)
  • in the wild, how people respond and provide multiple replies for users to choose


"Fine tuning" - specializing the network e.g. add medical data to generic pretrained model

gpt is fine tuned to be general purpose


How can a model become great? "Size matters, bigger is better"

Since 2018 model sizes have grown. gpt1 "ant brain" ...now gpt4 has 1 trillion parameters  "more than rat brain"

The more parameters, the more powerful the AI. When you get to half a billion parameters then "common sense reasoning" and "summarization" and "arithmetic" etc become possible and powerful

gpt4 has read almost all the human written text

gpt4 cost $100 million, openAI can do it because they have Microsoft backing them (not everyone can)


"HHH" framework i.e. Helpful, Honest and Harmless

  • How do you make the model HHH? You Fine tune it by letting it know if answers are correct or not.


LLMs generate hallucinations by fabricating nonexistent or false facts


Llama 2, a chatGpt like model from Meta

Comments

Post a Comment

Popular posts from this blog

deep dive into Material UI TextField built by mui

Image
The Material Design system is a design system used to create consistent digital experiences across platforms such as android, web and ios. Includes: a color system a typography system a shape system a grid system it addresses motion and lots more Also includes Material Theming which makes it easy to customize Material Design to match your brand. Material Design is not just a set of guidelines it is a whole ecosystem. Material itself stopped maintaining it's react implementation and recommends opensource versions.  A very popular open source implementation is  MUI ,  a react component library  built to implement Material Design . Bills itself as the "worlds most popular React UI framework" and at this time has 1.2M weekly downloads on npm and used by over 750k projects on github (did I mention it's very popular?).  Note: there are other react component libraries implementing bootstrap and ant design for example, so up to you and your design team on what you choose. 
Read more

angular js protractor e2e cheatsheet

Many from the protractor test suite category detailed task syntax example load page load a page browser.get('yoururl') browser btn actions browser back/forward browser.navigate().back() OR browser.navigate().forward() access elements on page See Protractor API list and examples from Protractors test suite access by angular ngmodel name (must match ngmodel in the html) this is the preferred approach element(by.model('person.name')); access by the css id on the page e.g, #user_name (but this can be brittle) element(by.id('user_name')) access angular binding to get generated value i.e. for {{}} element(by.binding('person.concatName')); access angular repeat list data (must match repeat stmt in the html) See Protractor test suit l
Read more

react-router v6.4+ loaders, actions, forms and more

react-router  react-router is used for in app single page app navigation, we configure routes for pages in the app to map to components react-router has over 1bn downloads, its an industry standard (but there are others e.g. Nextjs router) many apps use react-router, many on v5.x the new react-router 6.4.x the newest major version of react router v6.4 is a significant enhancement which adds a lot of new features. react-router v6.4 offers new features to load data (loaders) and handle actions (actions).  It's done gone all Remix-y which makes sense because the same guys wrote both. eliminates a lot of boilerplate code to fetch data in useEffect() and set various loading states put your data load and update logic in react-route components instead of custom code unlike Remix, react-router v6.4 is still just all on the client (not server like Remix) paradigm shifts Lets start with a couple of high level ideas; "back to the future" so to speak OR "what's old is new ag
Read more