Microsoft Research presented VASA, a AI framework that – from a single static image and a speech audio clip – can create lifelike talking faces of virtual characters with appealing visual affective skills.

 

 

 

VASA-1, is capable of not only producing lip movements that are exquisitely synchronized with the audio, but also capturing a large spectrum of facial nuances and natural head motions that contribute to the perception of authenticity and liveliness.

This development paves the way for real-time engagements with lifelike avatars that emulate human conversational behaviors.

 

VASA-1 website

 

What is an AI search engine?
An AI search engine is a search platform that uses machine learning models and artificial intelligence to provide results. With the help of other AI-based algorithms and natural language processing, it can understand more complex search queries and provide improved, more relevant, and accurate results to users.

 

Moreover, an AI search engine automatically generates more accurate search results and detailed, unique answers to complex questions. Instead of relying on keywords, these tools examine your queries and check your preferences, past interactions, interests, and behavior to provide a personalized answer that follows your requirements.

 

How does an AI-based search engine work?

AI search engines work similarly to their traditional counterparts but employ additional AI technologies, algorithms, and deep learning to understand queries better and offer improved results.

 

Here’s what a typical AI search engine does:

 

An AI search engine begins its search by indexing content from the web.
Once a user types in a question, the tool uses natural language processing to understand the meaning and intent of the query.
The engine determines the relevance of its indexed items to the query and assigns a relevancy score to each item.
The AI-based search engine provides the most relevant results at the top and goes down the list based on relevancy scores.
Moreover, machine learning models enable AI search engines to improve based on historical data and offer more relevant content and more accurate results after every interaction.

 

AI-driven search engine vs traditional search engine:

 

AI search engines can understand natural language queries and content better than regular search engines can.
Where regular search engines use only words and phrases to determine their results, AI tools consider the context and actual meaning behind the question.

AI-based search tools can cover advanced features like image search or voice search.
Traditional search engines work from scratch with every query, while their AI counterparts can learn from past interactions and improve over time as a result of machine learning.

AI search engines can answer complex questions and longer texts instead of being limited to simple queries.

Unlike traditional search engines, most AI-based ones offer customized and personalized search experiences.

AI search engines can interact conversationally and handle follow-up questions, whereas conventional search engines can only answer queries.

 

Here’s a list of AI search engines for browsers, along with links to their relevant pages
The list is relevant for Feb. 2024, but new startups and developments happen quickly now.

 

Perplexity AI
– Perplexity.ai
Perplexity.ai is an advanced answer engine founded by experts connected to DeepMind and OpenAI. It uses avant-garde AI, specifically OpenAI’s GPT-3 model, to provide direct answers.
However, instead of the traditional list of links, it offers summaries and citations as search results. This means that when users ask questions, the search engine not only provides relevant information but also presents concise summaries and references to relevant sources.

Moreover, Perplexity.ai offers an uninterrupted and straightforward browsing experience for users. It is add-free and does not require users to sign up or log in. More importantly, this AI search engine does not collect any personal information from its users, ensuring a secure and confidential search experience.

Anthropic’s Claude
–  Anthropic’s Claude
– Description: Claude is a conversational AI assistant developed by Anthropic that can search the internet, answer questions, and assist with various tasks.

 

Brave Search
– Brave Search
– Description: Brave Search is a privacy-focused search engine developed by the Brave browser, which uses AI to filter out bias and deliver relevant results.

 

You.com
– You.com
– Description: You.com is a search engine that uses AI to provide direct answers to queries instead of just listing web pages.

 

Neeva
– Neeva
– Description: Neeva is an ad-free search engine powered by AI that aims to provide personalized and relevant search results without tracking user data.

 

Wolfram Alpha
– Wolfram Alpha
– Description: Wolfram Alpha is a computational knowledge engine that uses AI to answer queries by computing answers from curated data.

 

Presearch
– Presearch
– Description: Presearch is a decentralized, community-driven search engine that rewards users with cryptocurrency for searching and contributing to the platform.

 

Dejian
– dejian
– Description: Dejian is an AI-powered search engine that aims to provide more accurate and relevant search results by understanding natural language queries.

 

Askwise
– Askwise
– Description: Askwise is an AI-powered search engine that uses natural language processing to provide direct answers to questions, rather than just listing web pages.

 

Yieko
– Yieko
– Description: Yieko is an AI-powered search engine that uses natural language processing and machine learning to provide relevant and personalized search results.

 

Some experts warn that artificial intelligence (AI) could pose catastrophic risks to national security and even human extinction.
A March 2024 report commissioned by the US state department and produced by Glassstone AI, a US-based company that promotes responsible AI development, states that advanced AI systems could be stolen, weaponized, or escape controls.
The report also says that AI labs are concerned about losing control of the systems they’re developing, which could have devastating consequences for global security.
The report’s title is “An Action Plan to Increase the Safety and Security of Advanced AI“.

An Action Plan to Increase the Safety and Security of Advanced AI

 

Other dangers of AI might be:
– Automation of jobs
– Spread of fake news
– Arms race of AI-powered weaponry
– Lack of transparency
– Bias
– Lack of protection for online data privacy
– Opportunity for people to cheat
– Overall human obsolescence

 

Some significant publications and open letters that have warned about the potential risks and implications of advanced AI development:

 

“Potential Risks from Advanced Artificial Intelligence” (2017)
– An open letter signed by over 8,000 individuals, including prominent figures like Elon Musk and Stephen Hawking, expressing concerns about the potential risks of advanced AI and calling for increased research into AI safety.

Potential Risks from Advanced Artificial Intelligence

 

 

“Autonomous Weapons: An Open Letter from AI & Robotics Researchers” (2015)
– This open letter, signed by over 3,000 AI researchers, called for a ban on offensive autonomous weapons and expressed concerns about the potential dangers of artificial intelligence in warfare.

Autonomous Weapons: An Open Letter from AI & Robotics Researchers

 

“Research Priorities for Robust and Beneficial Artificial Intelligence” (2016)
– This publication by researchers from the Future of Humanity Institute and the Machine Intelligence Research Institute outlined key research priorities for ensuring that advanced AI systems remain robust and beneficial.

Research Priorities for Robust and Beneficial Artificial Intelligence

 

“Asilomar AI Principles” (2017)
– A set of principles developed at the Asilomar AI conference, which aimed to guide the development of beneficial AI systems and address issues such as safety, ethics, and social impact.

Asilomar AI Principles

 

“The Malicious Use of Artificial Intelligence” (2018)
– A report by researchers from various institutions, including the University of Cambridge and the Future of Humanity Institute, highlighting the potential risks of AI systems being used for malicious purposes.

The Malicious Use of Artificial Intelligence

 

“Windfall: The Path to Prosperity After Artificial Intelligence” (2019)
– A book by Stanford University professor Erik Brynjolfsson and researcher Andrew McAfee, which discussed the potential economic disruptions and societal implications of advanced AI.

https://erikbryn.com/windfall/

 

“The One Hundred Year Study on Artificial Intelligence” (2016-2021)
– A multiyear study by Stanford University aimed at analyzing the potential impacts of AI on various domains, including ethics, economics, and policy.

The One Hundred Year Study on Artificial Intelligence

 

“The Ethics of Artificial Intelligence” (2020)
– A report by the IEEE Global Initiative on Ethics of Autonomous and Intelligent Systems, which outlined ethical considerations and recommendations for the development and deployment of AI systems.

The Ethics of Artificial Intelligence

 

“Superintelligence: Paths, Dangers, Strategies” (2014)
– A book by philosopher Nick Bostrom, which explored the potential risks and existential threats posed by advanced artificial superintelligence.

Superintelligence: Paths, Dangers, Strategies

 

“AI Ethics in Research and Development” (2022)
– A report by the AI Ethics Board of the European Union, which provided guidance and recommendations for the ethical development and deployment of AI systems within the EU.

AI Ethics in Research and Development

 

These and other publications, open letters, and reports have raised awareness about the potential risks, ethical considerations, and societal implications of advanced AI development. They have sparked discussions and calls for responsible AI governance, safety research, and proactive measures to mitigate potential negative consequences.

 

The AI Index report tracks, distills, and visualizes data related to artificial intelligence (AI). Its mission is to provide unbiased, rigorously vetted, broadly sourced data in order for policymakers, researchers, executives, journalists, and the general public to develop a more thorough and nuanced understanding of the complex field of AI.

Here a summary from 2023:

1. Al beats humans on some tasks, but not on all.
Al has surpassed human performance on several benchmarks, including some in image classification, visual reasoning, and English understanding.
Yet it trails behind on more complex tasks like competition-level mathematics, visual commonsense reasoning and planning.

 

2. Tech Industry continues to dominate frontier Al research.
In 2023, industry produced 51 notable machine learning models, while academia contributed only 15. 

 

3. Frontier AI models get way more expensive.
According to Al Index estimates, the training costs of state-of-the-art Al models have reached unprecedented levels.
For example, OpenAl’s GPT-4 used an estimated $78 million worth of compute to train, while Google’s Gemini Ultra cost $191 million for compute.

 

4. The United States leads China, the EU, and the U.K. as the leading source of top Al
models.
In 2023, 61 notable Al models originated from U.S.-based institutions, far outpacing the European Union’s 21 and China’s 15.

 

5. Robust and standardized evaluations for LLM responsibility are seriously lacking.
New research from the Al Index reveals a significant lack of standardization in responsible Al reporting.
Leading developers, including OpenAl, Google, and Anthropic, primarily test their models against different internal Al benchmarks. This practice complicates efforts to systematically compare the risks and limitations of top Al models.

 

6. Generative Al investment skyrockets.
Despite a decline in overall Al private investment last year, funding for generative Al surged, nearly octupling from 2022 to reach $25.2 billion.
Major players in the generative Al space, including OpenAl, Anthropic, Hugging Face, and Inflection, reported substantial fundraising rounds.

 

7. Al makes workers more productive and leads to higher quality work.
In 2023, several studies assessed Al’s impact on labor, suggesting that Al enables workers to complete tasks more quickly and to improve the quality of their output. These studies also demonstrated Al’s potential to bridge the skill gap between low- and high-skilled workers.
Still, other studies caution that using Al without proper oversight can lead to diminished performance.

 

8. Scientific progress accelerates even further, thanks to Al.
In 2022, Al began to advance scientific discovery. 2023, however, saw the launch of even more significant science-related Al applications – from AlphaDev, which makes algorithmic sorting more efficient, to GNoME, which facilitates the process of materials discovery.

 

9. The number of Al regulations in the United States sharply increases.
The number of Al-related regulations in the U.S. has risen significantly in the past year and over the last five years. In 2023, there were 25 Al-related regulations, up from just one in 2016. Last year alone, the total number of Al-related regulations grew by 56.3%.

 

10. People across the globe are more aware of Al’s potential impact -and more nervous.
A survey from Ipsos shows that, over the last year, the proportion of those who think Al will dramatically affect their lives in the next three to five years has increased from 60% to 66%. Moreover, 52% express nervousness toward Al products and services, marking a 13 percentage point rise from 2022. I

n America, Pew data suggests that 52% of Americans report feeling more concerned than excited about Al, rising from 37% in 2022.

DALL-E and Midjourney are both AI image generation tools that can create images from text prompts.
They use diffusion training, which involves training the AI with image-word associations and breaking down images until there is only random noise.
Both tools are already used by artists and graphic designers and “hobby creators”.

 

DALL-E website

Midjourney showcase website

 

Generative AI are artificial intelligence systems that are capable of generating new content, such as text, images, audio, or video, based on the patterns and relationships learned from training data. These systems use deep learning techniques, particularly generative models like variational autoencoders (VAEs), generative adversarial networks (GANs), and transformer-based language models, to create novel and realistic outputs.

 

 

Timeline of some significant milestones in the development of Generative AI, along with the companies and research groups involved:

 

2018: GPT (Generative Pre-trained Transformer) (OpenAI)
– OpenAI introduced GPT, a transformer-based language model that could generate human-like text by learning from a large corpus of online data.
– GPT demonstrated the potential of transformer models for natural language generation tasks.

 

2019: BigGAN (DeepMind, Google Brain)
– DeepMind and Google Brain researchers introduced BigGAN, a large-scale GAN model capable of generating high-resolution, diverse, and coherent images.
– BigGAN showed the scalability of GANs and their ability to generate high-quality images across various domains.

 

2020: DALL-E (OpenAI)
– OpenAI introduced DALL-E, a transformer-based model that could generate realistic images from text descriptions.
– DALL-E demonstrated the ability of generative models to understand and translate natural language into visual representations.

 

2021: Stable Diffusion (Runway)
– Runway released Stable Diffusion, an open-source text-to-image generative model based on latent diffusion.
– Stable Diffusion made high-quality image generation accessible to a wider audience and fueled the growth of AI art.

 

2022: DALL-E 2 (OpenAI), Imagen (Google Brain), and Parti (Google Brain)
– OpenAI introduced DALL-E 2, an advanced version of DALL-E with improved image generation capabilities.
– Google Brain released Imagen, a text-to-image model that could generate high-resolution images with unprecedented quality and detail.
– Google Brain also introduced Parti, a generative model that could create 3D objects from natural language descriptions.

 

2023: GPT-4 (OpenAI), Claude (Anthropic)
– OpenAI released GPT-4, a powerful language model with multimodal capabilities, including image and text generation.
– Anthropic introduced Claude, a conversational AI system capable of generating human-like responses and handling a wide range of tasks.

 

This timeline highlights the rapid progress in generative AI, driven by advancements in deep learning architectures, increasing computational power, and the availability of large datasets. 

 

Here are some applications of generative AI across different fields:

 

1. Art and Design:
– Generating realistic images, artwork, and designs from text descriptions or sketches
– Creating unique textures, patterns, and visual effects
– Exploring new artistic styles and creative expressions

 

2. Media and Entertainment:
– Generating realistic characters, environments, and animations for movies and video games
– Creating synthetic voice-overs, music, and sound effects
– Generating personalized stories, scripts, and dialogue

 

3. Marketing and Advertising:
– Generating personalized marketing content, such as product descriptions, advertisements, and social media posts
– Creating synthetic product images and visualizations
– Generating realistic virtual models and influencers for marketing campaigns

 

4. Scientific Research:
– Generating synthetic data for training and testing machine learning models
– Simulating complex physical or biological systems for experimentation
– Generating novel molecular structures for drug discovery and material design

 

5. Natural Language Processing:
– Generating human-like text for content creation, summarization, and translation
– Creating synthetic training data for language models and conversational AI
– Generating personalized responses and recommendations in virtual assistants

 

6. Computer Vision:
– Generating synthetic images for training and data augmentation
– Creating realistic virtual environments and scenarios for simulations
– Generating image-to-image translations, such as style transfer and image enhancement

 

7. Finance and Business:
– Generating synthetic financial data for risk modeling and scenario analysis
– Creating personalized reports, summaries, and business analyses
– Generating realistic simulations for training and decision-making

 

Rather than aiming to understand words individually, BERT algorithms helps Google understand words in context. This led to a huge quality improvement across Search, and made it easier for people to ask questions as they naturally would, rather than by stringing keywords together.

 

Google’s research on Transformers led to the introduction of Bidirectional Encoder Representations from Transformers, or BERT for short.

It helped Searchunderstand users’ queries better than ever before.

 

Understanding searches better than ever before

The Transformer has revolutionized what it means for machines to perform translation, text summarization, question answering and even image generation and robotics.

 

The Google Research paper “Attention Is All You Need” introduced the Transformer, a new neural network architecture that helped with language understanding.

 

Before the Transformer, machines were not very good at understanding the meaning of long sentences — they couldn’t see the relationships between words that were far apart.

 

The Transformer hugely improved this and has become the bedrock of today’s most impressive language understanding and generative AI systems.

 

Transformer: A Novel Neural Network Architecture for Language Understanding

Artificial Neural Networks or Neural Nets are computational models – computer algorithms – inspired by the structure and function of the human brain. They consist of interconnected nodes, called neurons, organized in layers. Neural networks are capable of learning complex patterns in data and are widely used in various machine learning tasks.

 

Neural networks are trained using a process called deep learning, where they learn to perform tasks by adjusting the strengths of connections (weights) between neurons based on example inputs and corresponding outputs. Through this process, neural networks can learn complex patterns in data and make predictions or decisions based on new input.

 

if you really want to know how this works, watch this simple ;  ) video.

 

 

Research into Neural Networks has been going on since the 1980s. But because of the limited computing power and available data little progress was possible.

 

This changed dramatically later one especially from the 2010s onwards.
Here are some – more technical – details about the growing importance and use of Deep Neural Networks in different domains:

 

Neural Network Deep Learning Breakthroughs:
– The 2010s saw significant breakthroughs in deep learning research, with neural networks achieving state-of-the-art performance in various domains. This period witnessed the development of deep learning architectures such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and transformer models.

 

Image Recognition and Computer Vision
– CNNs, particularly deep CNNs, became the standard architecture for image recognition and computer vision tasks and achieved remarkable performance on image classification, object detection, and image segmentation tasks.

 

Natural Language Processing (NLP)
– Transformer models, such as BERT (Bidirectional Encoder Representations from Transformers) and GPT (Generative Pre-trained Transformer), revolutionized natural language processing tasks. These models surpassed previous benchmarks in tasks like language understanding, question answering, text generation, and machine translation.

 

Transfer Learning and Pre-training
– Transfer learning and pre-training became dominant paradigms in deep learning, where models are first trained on large-scale datasets and then fine-tuned for specific tasks. Pre-trained models, such as those released by OpenAI (GPT, GPT-2, GPT-3) and Google (BERT), demonstrated superior performance and generalization across diverse NLP tasks.

 

Reinforcement Learning Advances
– Reinforcement learning (RL) algorithms, particularly deep reinforcement learning (DRL), made significant strides in solving complex sequential decision-making tasks. Deep Q-Networks (DQN), Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), and AlphaGo demonstrated impressive results in games, robotics, and autonomous systems.

 

Autonomous Vehicles and Robotics
– Neural networks played a crucial role in the development of autonomous vehicles and robotics. Deep learning models enabled advancements in perception, localization, mapping, and control, leading to autonomous driving systems deployed by companies like Tesla, Waymo, and Uber, as well as robotic applications in manufacturing, healthcare, and logistics.

 

Healthcare and Biomedical Applications
– Neural networks were increasingly applied in healthcare and biomedical research for tasks such as medical image analysis, disease diagnosis, drug discovery, and personalized medicine. Deep learning models demonstrated promising results in detecting diseases from medical images, predicting patient outcomes, and analyzing genomic data.

 

Ethical and Societal Implications
– The rapid advancement of neural networks raised concerns about their ethical and societal implications. Debates around issues like bias in AI systems, privacy concerns, algorithmic fairness, and responsible AI research gained prominence, leading to efforts to address these challenges through interdisciplinary collaborations and policy interventions.

 

In 2001, Google introduced machine learning to enhance its search engine capabilities, particularly in correcting users’ spelling errors.

 

This innovation marked a significant advancement in improving the accuracy and relevance of search results. By leveraging machine learning algorithms, Google was able to analyze patterns in user queries and suggest corrections for misspelled words, ultimately leading to a more seamless and effective search experience for users.

 

This integration of machine learning into Google Search exemplifies the continuous evolution of technology to better serve user needs and enhance overall user satisfaction.

Long before Siri and Alexa an ChatGPT & Co, there was ELIZA, the grandmother of all chatpots :-D. 

 

In 1966, MIT researcher Joseph Weizenbaum created Eliza, the first chatbot, which marked the beginning of research into Natural Language Processing (NLP).  Eliza used pattern recognition and responded to programmed trigger-words to simulate human conversation. (Hint: Eliza was not super intelligent 😉

 

ELIZA simulated a conversational interaction somehow similar to what might take place in the office of a psychotherapist in an initial psychiatric interview.

Here is a recreation of the original program

 

Enjoy your free therapy session ;  )

 

 

 

John McCarthy, a American computer scientist and cognitive scientist held a workshop on “artificial intelligence” which is the first use of the word, and how it came into popular usage.

He is considered to be one of the “founding fathers” of Artificial Intelligence.