Job Nijenhuis’ Post

It is so strange to talk about "AI" in those days. It could mean anything to nothing. from Superhuman Agency to the simplest automation. To me a ChatBot is no "AI". Its a tool. It's a compressed knowledge retrieval system, it can process natural language. That's really it. There is no understanding or usability, without sophisticated systems in place that integrate that ChatBot. To me the notion of AI is about automation. Controlled, reliable scaling and the build of systems that solve problems by themselves. A ChatBot is not able to do that, because it just does not have any "Intelligence", it's a tool. If it would have any understanding, it would fail very complex tasks much more likely than the simplest tasks. In fact, it would not fail ANY of the simple tasks. And the fact that those systems can fail so dumbly makes scale very unefficient. And given that we understand very well WHY those systems fail so dumbly, tells us quite good, that this cannot ever really change. For example: LLM supervision with other LLM's must also break at any point due to the lack of understanding. So what system is supervising the supervisor?

Interesting read. My knowledge of deep learning is not very deep, but my experience with simple ANNs and basic knowledge about experience with the current LLMs let me think that "reasoning" is not yet what they do. However, they can follow the "reasoning" they learned from their training data, like following the steps for solving a Solow model, as in one of my previous examples.

How Quickly Do Large Language Models Learn Unexpected Skills?

Large language models (LLMs) are machine-learning models designed to understand and generate human language.

6 Best Large Language Models (LLMs) in 2024 https://lnkd.in/eHKTBUQH via drumup.io

"Long-Form Factuality in Large Language Models" introduces a new approach to evaluating and benchmarking the factuality of long-form responses generated by large language models (LLMs). ▪️ LongFact: 2,280 fact-seeking prompts across 38 topics, generated by GPT-4 to assess factuality in long-form responses. First dataset for evaluating LLM factuality over multiple paragraphs. ▪️ SAFE Methodology: It uses LLMs to evaluate response factuality. By breaking down responses into facts, checking their accuracy using Google Search, and aggregating evaluations, it provides a factuality score. SAFE outperforms human annotators in cost and reliability, with 72% agreement on ~16k facts. In 100 disagreements, SAFE was preferred 76% of the time, showing its efficiency and effectiveness. The authors provide open access to LongFact, SAFE, and all experimental code: https://lnkd.in/gq47hmue

How Quickly Do Large Language Models Learn Unexpected Skills?

6 Best Large Language Models (LLMs) in 2024 https://drumup.io/s/sY2U3H

https://lnkd.in/e2evrdTU Chain-of-Symbol Prompting For Large Language Models #llms #genai #generatieveai #cos Cobus Greyling

In the first post in this series, we introduced retrieval augmented generation (RAG) , explaining that it became necessary to expand the capabilities of conventional large language models (LLMs) . #teachthemachine