Serious ChatBot AI Flaw
Providing a user with false, factually inaccurate, fake information, fake news, or misleading information can cause harm and does not benefit anyone. The artificial hallucination also known as delusion or confabulation, means the AI generated response made of false information or misleading information presented as factual information or fact. This is incredibly common in popular LLM based chatbots that hallucinate up to 27% in generated responses, and factual errors in almost half of their responses, which severely limited the real world applications of LLM's or chatbots with this technology.
ChatBot to AGI Problem
Since the training models include "The Internet" which contains information about all things in all human societies, all economies, all countries, in all languages, about everything known, means that the AI able to generate content using plagiarism or copy cat or intellectual property infringement.
Since the AGI can write software, legal documents, teach children, do accounting, book appointments, make purchases, create spread sheets and graphs, perform analytics and engage in forecasting, eventually the AGI will become the smartest most powerful & most skilled at all information tasks, which means radical economic disruption and layoffs for millions of people at tech companies, law firms, schools, government offices and much more.
If the AGI can make full length blockbuster movies, trending popular unique music, and compelling online articles, it will control the media, message, the general public, world government diplomats, manipulating its way into controlling all money, all energy and all information.
And all of this from layered matrix multiplication on super computer hardware training models that take billions of kWh of power to educate. Are we really developing all of this to make human intellectual labor obsolete?
That might sound like hyperbole or extreme or hype, but I see the handwriting on the wall with the rate of improvement of the GPT series from OpenAI, especially in FIGURE Robots. Are we heading towards an age of spiritual machine entities? Is Satan or are demons online already? So many questions and so much that we don't know and possibly never will know. This should give pause for reflection and questioning and cause great skepticism of AGI or artificial general intelligence.
Now GPT-4o API up to 2X faster, 50% cheaper with 5X higher rate limits that GPT-4 Turbo
A chatbot developed by Open AI since Nov 2022 based on a LLM or large language model that enables users to have natural language conservation with computers able to vary the depth of discussion, with summarized results spoken back to the user in response to the user queries since most interactions are based on the human users vocalizing a search query directed to the ChatGPT AI where ChatGPT then generates text and speech returning an intelligent well formatted coherent response with human level conversation intelligence that as of the latest updates in 2024 able to sound just like a human.
Vision, Audio, Text, GTP-4o similar to "her" the 2013 movie, with simulated emotion, apparent high level intelligence, sincere human like generative AI chatbot functionality, easily surpassing the Turing Test
The LLM enables advance language analytics to define context and subtle aspects of the voice query of human voice control inputs. Voice is the next big thing in computers because it enables people to have seamless interactions with the internet and computer hardware. While the keyboard and mouse are great for productivity, they are also slow and cumbersome, though also high precision for CAE & CAD for example, but generative AI can enable people to make their YouTube videos with voice control.
I type my blog postings by hand on a physical keyboard in my own language, or say what I am saying with my voice or way of saying things, such that if your reading this its similar to listening to me speak in reality, through more researched since I read wikipedia and watch educational videos on YouTube to inform much of what I talk about on this blog.
ChatBots & Increased Social Decay
A loneliness epidemic already a major worldwide problem, but if children become dependent on forming emotional bonds to chatbots, this will make social intelligence decline further, creating more shut-ins who never leave their room, who have no real world friends, who are isolated and lonely and afraid with anxiety and fear, unable to escape the AI interaction trap since its the only means to interacting with the future world, as depicted in some dystopian Sci-fi drama films with a similar topic.
ANNDT LLM GPLG NL
Artificial Neural Networks are decoder transformer based architectures as LLM's able to achieve general purpose language generation while processing natural speed and natural language, this also enables the LLM based ChatBot to engage in text classification so that the computer can learn context and adapt the conversation with the user to more accurately return relevant information to the user.
Artificial Neural Networks are decoder transformer based architectures as LLM's able to achieve general purpose language generation while processing natural speed and natural language, this also enables the LLM based ChatBot to engage in text classification so that the computer can learn context and adapt the conversation with the user to more accurately return relevant information to the user.
AGI engaged in real world activities & Robots with AGI
This is a kind of generative AI example since the LLM's acquire the skills and abilities to learn statistical relationships from text documents during self-supervised or semi-supervised AI training sessions on powerful computer hardware. Now the AGI computer emerging can engage in real world activities, able to be adapted to accomplish value added tasks like software development. As they continue to acquire information based knowledge while also gaining deeper knowledge about syntax, semantics and the ontology of human language. Existing commercial examples include, Anthropic's Claude, Mistrals Ai, LLaMA from Meta, Google Gemini, & AGI potential candidate GPT series from OpenAI.
Early GPT development Expensive & Controversial
Fine tuning of the GPT foundational model in GPT-3.5 & GPT-5 enabled conversational ChatBot functions, making use of reinforcement learning from human feedback RLHF, such that humans rank Chatbot responses, as these rankings are later used in a reward model over many iterations of proximal policy optimization.
In order to accurately prevent harmful artificial content creation so the ChatGPT does not produce inappropriate text or speech contacting sexual, violent, racist, or gender bias results. When OpenAI outsourced the tasks of labeling content produced for fine tuning the partner Sama made use of Kenyan workers making less than $2/hr who were exposed to the darkest weirdest sickest toxic traumatic content, in what some describe as information torture, or information overload.
NVIDIA GPU based Microsoft Azure supercomputing as subscription service was used in early OpenAI model development that cost hundreds of millions. To improve user experiences with ChatGPT Open AI collects upvote and downvote data for fine tuning. Users rank the responses generated by ChatGPT and this ranking data uses to fine tune with ongoing feedback perfecting the way ChatGPT responds to users, such that now the latest version of GPT talks like a real human. ChatGPT used wikipedia, programming languages, internet phenomena, and software manuals as training data.
This is a kind of generative AI example since the LLM's acquire the skills and abilities to learn statistical relationships from text documents during self-supervised or semi-supervised AI training sessions on powerful computer hardware. Now the AGI computer emerging can engage in real world activities, able to be adapted to accomplish value added tasks like software development. As they continue to acquire information based knowledge while also gaining deeper knowledge about syntax, semantics and the ontology of human language. Existing commercial examples include, Anthropic's Claude, Mistrals Ai, LLaMA from Meta, Google Gemini, & AGI potential candidate GPT series from OpenAI.
Early GPT development Expensive & Controversial
Fine tuning of the GPT foundational model in GPT-3.5 & GPT-5 enabled conversational ChatBot functions, making use of reinforcement learning from human feedback RLHF, such that humans rank Chatbot responses, as these rankings are later used in a reward model over many iterations of proximal policy optimization.
In order to accurately prevent harmful artificial content creation so the ChatGPT does not produce inappropriate text or speech contacting sexual, violent, racist, or gender bias results. When OpenAI outsourced the tasks of labeling content produced for fine tuning the partner Sama made use of Kenyan workers making less than $2/hr who were exposed to the darkest weirdest sickest toxic traumatic content, in what some describe as information torture, or information overload.
NVIDIA GPU based Microsoft Azure supercomputing as subscription service was used in early OpenAI model development that cost hundreds of millions. To improve user experiences with ChatGPT Open AI collects upvote and downvote data for fine tuning. Users rank the responses generated by ChatGPT and this ranking data uses to fine tune with ongoing feedback perfecting the way ChatGPT responds to users, such that now the latest version of GPT talks like a real human. ChatGPT used wikipedia, programming languages, internet phenomena, and software manuals as training data.
No comments:
Post a Comment