Category: HAL149

The real AI thing will be some kind of AI agent that can handle all of your online digital interactions, including but not limited to:

– Managing the agenda,
– Writing and answering emails,
– Web browsing and online transactions,
– Summarising and commenting on news stories
– Taking part in meetings and phone calls, etc.

Imagine having an AI agent that efficiently takes care of all your online interactions, meeting with it for 20-30 minutes a day to give further instructions or get updates, and having the rest of the day to think or do whatever you want.…

Everyone is confused and stressed about what to build on AI because everyone is focused on a moving target: technology.

No one is talking about what companies really need. The real challenge is not the AI models, the blockchain, the apps and so on, but how to create value.

Technology is a moving target and will remain so. Its only part of the equation: the fundamental part is people.…

In this insightful talk, Sam Altman emphasises that only a few companies are likely to have the resources to build and maintain Large Language Models (LLMs) such as GPT-3. However, he foresees the emergence of many “layer two” companies with valuations in excess of a billion dollars over the next decade.

“Layer two” is referring to companies built on top of fine-tuned base models that unlock efficiency and progress in domain-specific industries.…

An interesting take on a very interesting topic: can we get creative, valuable ideas from generative AI?.

Some weeks ago I wrote about the core of this problem: finding “what we don´t know that we don´t know“.

But could AI generate the next billion-dollar business idea? Product ideation is the best place to start. If you only have vague initial ideas, generative AI can really help to crystallise them.…

Achieving real value depends on moving from theoretical models to production-level accuracy. This shift requires investment in data tagging and development.

Specialised, fine-tuned models consistently outperform generic counterparts such as ChatGPT, and outperform alternative approaches such as zero-shot learning and prompt-based methods, as shown in studies.…

The fact that OpenAI has just released enough updates to destroy a lot of AI startups and plugins should make everyone think.

Does it still make sense to build products on top of the platform? Why should you invest time and money in creating new products that could soon be cannibalised by these people?

Yet it does make sense why? because most users wont get into paid options, not to mention complex configuration menus.…

Artificial intelligence finds and builds patterns in data. But arranging alphabetical symbols in a way that makes sense to us does not imply intelligence or a sense of meaning.

A stable diffusion model helps you design cars or houses, but it has no idea of the real meaning of those shapes. It goes no further than knowing that they belong to a particular archetypal class or ‘latent space’.…

Goal is to convert lengthy documents into numerical representations (vector embeddings) and store them in a vector search engine.

When users engage in conversation with the documents, the system employs Approximate Nearest Neighbor search to find and return relevant text responses. This is achieved using:

– OpenAI’s cost-effective ChatGPT API (gpt-3.5-turbo)
– The vector database, Chroma, is suitable when used alongside LangChain for building applications with Large Language Models (LLMs).…

Most of the queries that people and businesses make to chatgpt are to retrieve known data that can be found online. And so we expect it to provide real answers to real, concrete questions.

But knowing what the network knows, or even what it ‘knows it doesn’t know’, doesn’t really make a competitive difference: there are lots of people and resources working on it.…

Every technological revolution creates opportunities because it reshuffles the deck. But the rules of the game are the same: it’s always about human needs.

ChatGPT is great compared to what we had before, in the same way that a car is better than a horse: its optimal for certain tasks under certain constraints.

But it is not a one-stop solution for all business problems or some kind of oracle that knows everything.…

Google’s current business model is primarily focused on links and clicks, rather than on AI technology. This is based on the assumption that users interact with the WWW, which is a network of websites and links.

However, as AI assistants become more advanced they will start working as augmented search engines, capable of mimicking human interaction, summarizing and presenting information in a fully customized way, with:

– No ads
– No SEO-optimized article crap
– No cookie banner
– No popups, no usability tricks, etc.…

Machine learning is fundamentally about reducing entropy: finding the basic parameters that define a set of observations so we can either predict or generate new observations.

Reducing entropy can be seen as building a model that explains data. By reducing entropy, we aim to capture the underlying patterns and structure in the data, allowing us to better understand and explain the observed phenomena.…

Here’s some code for chatting and asking questions about the first book of Don Quixote (or any other long pdf, for that matter).

It is about 500 pages in .txt format.

Need to check what´s wrong with langchain CharacterTextSplitter when using texts in SpanishI I created a provisional script to split the text with just length criteria.