Welcome back to AI Weekly Digest #33, your go-to resource for the most important artificial intelligence news and events from Week 48 of 2023.
As an AI enthusiast, staying informed about the rapid advancements in the field is crucial for both personal and professional growth. Our mission is to bring you a concise and comprehensive roundup of the latest breakthroughs, innovations, and discussions shaping the world of AI.
Dec 02, 2023
Table Of Contents
- AI Image Generation Has Similar Carbon Footprint to Charging a Smartphone
- Google Pushes Back Launch of Powerful AI Model Gemini to 2024
- Google AI Introduces Translatotron 3 for Unsupervised Speech-to-Speech Translation
- Israel Uses AI to Double Airstrike Targets in Gaza
- Meta AI Announces A Suite Of AI Language Translation Models
- OpenAI Updates & Improved Custom GPTs
- Mastercard’s AI Shopping Assistant Aims to Take the Stress Out of Gift Giving
- Microsoft’s $3.2 Billion UK Investment: A Major Boost for AI Development
- AI wars: Amazon is coming for Microsoft and Google
- Perplexity.ai Unveils Online Large Language Models
- Stability AI Considers Sale Amid Investor Pressure Over Finances
- Sam Altman is back as CEO, Mira Murati as CTO and Greg Brockman as President
- A Google AI Unveils 2.2 Million Hidden Materials, Rewriting Scientific Frontiers
- Nvidia CEO Predicts AI Will Rival Humans in 5 Years
- Amazon Launches AI Model Benchmarking with Human Evaluators
- Amazon Unveils Titan Image Generator at AWS re:Invent 2023, Strengthening AI Image Generation Landscape
- Stability AI Introduces Real-Time Text-to-Image Model SDXL Turbo
- Amazon Announced Amazon Q: A Generative AI Assistant
- AWS and NVIDIA Announce Strategic Collaboration to Offer New Supercomputing Infrastructure, Software and Services for Generative AI
- Pika’s Breakthrough in AI Video Editing: A $55M Leap Forward
- Google, Symphony unveil AI JV to help banks manage voice calls compliance risk
- Meta AI’s Breakthrough in Non-Invasive Speech Decoding: A Ray of Hope for Speech Impairments
- Cradle’s Groundbreaking AI in Protein Design: A Leap Forward with $24M Funding
- ChatGPT Outperforms Students in Essay Writing, Study Reveals
- Generative AI Used By Three-Quarters Of Young Brits, Ofcom Says
- Salesforce and Amazon Web Services Expand Partnership to Enable AI Innovation
- UK Pupils Exploiting AI to Manufacture Explicit Images of Peers
- Neuralink Raises Additional $43 Million Amid Mounting Controversies
- Global Agreement Ensures Ethical AI Development: 193 Countries Commit to Secure-by-Design Approach
AI Image Generation Has Similar Carbon Footprint to Charging a Smartphone
Generating images with AI systems like Dall-E and Midjourney could have nearly the same carbon footprint as charging a smartphone, according to a new study by Hugging Face and Carnegie Mellon. The research found that creating AI-generated images uses substantially more energy than other AI tasks. Testing 88 models on 30 datasets, the study concluded that large, multipurpose models like ChatGPT are more energy intensive than specialized models.
Google Pushes Back Launch of Powerful AI Model Gemini to 2024
Google has quietly delayed the launch of its much-anticipated artificial intelligence system Gemini until 2024. Originally planned to debut next week, Gemini is Google’s next-generation multimodal AI that can understand and generate text, images, websites and more based on textual descriptions or sketches. Concerns over Gemini’s inability to reliably respond to some non-English prompts led Google to postpone its launch events in New York, Washington and California.
With vastly more computing power than OpenAI’s GPT-4, Gemini has impressed Google executives by creating novel images that demonstrate an ability to “speak in imagery.” While OpenAI’s ChatGPT currently leads consumer awareness, analysts believe Gemini could shift sentiment when it eventually launches. For now, the world will wait a bit longer to experience Google’s most complex AI to date. Gemini promises to transform how humans interact with AI, but Google wants to ensure it works reliably first.
Dec 01, 2023
Google AI Introduces Translatotron 3 for Unsupervised Speech-to-Speech Translation
Google has introduced Translatotron 3, a novel unsupervised speech-to-speech translation model that can directly translate speech between languages without parallel bilingual speech data. Previous Translatotron models required supervised parallel speech data which is scarce.
Translatotron 3 incorporates pre-training, unsupervised embedding mapping, and a back-translation reconstruction loss to enable direct speech translation using just monolingual datasets. An encoder is shared between the source and target languages and trained to output multilingual embeddings understandable by both decoders. The model is trained using a combination of unsupervised losses.
Experiments between Spanish and English audio show Translatotron 3 outperforms a baseline cascaded system on translation quality, speaker similarity after translation, and speech naturalness. The model particularly excelled on conversational speech data.
Israel Uses AI to Double Airstrike Targets in Gaza
The Israeli Defense Force (IDF) is using an AI-powered targeting system called “The Gospels” to identify potential airstrike sites in Gaza. The system processes intelligence and generates target recommendations for human analysts to review. An IDF spokesperson claimed the system helps “cause great damage to the enemy and minimal harm to non-involved civilians.” However, The Guardian reports it has also recommended strikes on private homes of suspected Hamas or Islamic Jihad members.
Since fighting began in October, the AI-enhanced system has identified 15,000 potential targets in Gaza – potentially doubling the number from the 2014 war with Hamas. As AI becomes more common, militaries worldwide are looking to leverage it on the battlefield. The U.S. uses AI to monitor airspace in Washington D.C. and recently announced efforts to establish global standards for responsible military use of AI. However, some companies like Shield AI see bringing AI tools to war as a way to save lives by deterring adversaries.
Meta AI Announces A Suite Of AI Language Translation Models
Meta AI has unveiled ‘Seamless’, a suite of AI language translation models, heralding a new era in real-time, expressive cross-lingual communication. This groundbreaking system, which includes the SeamlessExpressive and SeamlessStreaming models, is built on the SeamlessM4T v2 foundation model. SeamlessExpressive is engineered to maintain elements crucial to human speech such as tone, emotion, and style, particularly in speech-to-speech translation between six major languages. SeamlessStreaming, on the other hand, offers near real-time translation while a speaker is still talking, supporting nearly 100 languages for both speech recognition and translation.
OpenAI Updates & Improved Custom GPTs
OpenAI has introduced significant updates to enhance user experience and functionality of their Generative Pre-trained Transformers (GPTs). Key improvements include a revamped configuration interface, the introduction of one-click testing for ease of use, the addition of debug messages in previews for better troubleshooting, and support for multiple domains to broaden application scope. Additionally, addressing concerns regarding file uploads, the feature of downloading uploaded files via Code Interpreter has been set to off by default, with added explanatory messaging for clarity.
In other news, the much-anticipated launch of the OpenAI GPT Store is now scheduled for early next year, slightly delayed from its initially planned release this month. This postponement is attributed to unforeseen engagements that have required additional focus. Meanwhile, users can look forward to more exciting updates to ChatGPT, as OpenAI continues to innovate and enhance its offerings.
Mastercard’s AI Shopping Assistant Aims to Take the Stress Out of Gift Giving
Finding the perfect gift can be extremely difficult, but Mastercard is aiming to make the process easier with its new AI-powered shopping assistant called Shopping Muse. Powered by personalization platform Dynamic Yield, which Mastercard acquired last year, Shopping Muse acts as a gift concierge chatbot that provides personalized and relevant product recommendations to shoppers. It translates shoppers’ requests into tailored suggestions by taking into account their browsing history, past purchases, and affinity profile.
Nov 30, 2023
Microsoft’s $3.2 Billion UK Investment: A Major Boost for AI Development
Microsoft’s recent announcement to invest $3.2 billion (£2.5 billion) in the UK over the next three years marks its largest investment in the country and a significant milestone in advancing artificial intelligence (AI) technology. This move is set against the backdrop of Britain’s push for private investment in AI and other growth industries, amidst a forecasted sluggish economy.
This funding, initially unveiled by Prime Minister Rishi Sunak, is expected to more than double Microsoft’s datacenter footprint in the UK. This expansion is pivotal for developing new AI models, which require substantial infrastructure. Sunak hailed this investment as a turning point for the UK’s AI infrastructure and development.
AI wars: Amazon is coming for Microsoft and Google
In a bold move to challenge tech giants Microsoft and Google in the AI arena, Amazon’s AWS re:Invent developer conference unveiled significant advancements, signaling a fierce competition ahead. Central to Amazon’s strategy is its focus on generative AI, aiming to offset Microsoft and Google’s early dominance in this field. The conference highlighted three major developments: the introduction of the Q chatbot, an advanced AI chip named Trainium2, and an enhanced partnership with Nvidia.
The Q chatbot, tailored for enterprise customers, is designed as a versatile assistant, capable of drafting emails, generating business ideas, summarizing reports, and planning workshops. Trained on 17 years of AWS data, Q represents a significant step for Amazon in the realm of AI, aligning with similar services offered by Microsoft and Google, like Microsoft’s Copilot and Azure AI studio, and Google’s Duet for Workspaces.
Furthermore, Amazon’s new AI chip, Trainium2, marks a significant leap in AI model training, promising four times faster training and double the energy efficiency compared to its predecessor. This advancement, coupled with a commitment of up to $4 billion in AI developer Anthropic, places Amazon at the forefront of AI technology development.
Perplexity.ai Unveils Online Large Language Models
Artificial intelligence startup Perplexity.ai has announced two new online large language models (LLMs) – pplx-7b-online and pplx-70b-online. These models aim to provide helpful, factual, and up-to-date responses by leveraging Perplexity’s in-house search technology. The search infrastructure allows the models to access the most relevant and timely information from the internet to inform their responses.
In evaluations, human testers preferred Perplexity’s models over OpenAI’s GPT-3.5 and Anthropic’s LLama2-70b for accuracy and freshness of information. For example, on a prompt asking for current data on US licensed drivers, Perplexity’s models provided 2023 projections while the other models could only offer outdated 2021 statistics.
The improved performance stems from fine-tuning techniques and pairing the models with Perplexity’s search index of high-quality, frequently updated websites. By enabling access to live data, the company’s “online LLMs” can respond to queries requiring niche or time-sensitive knowledge more effectively than offline models.
With the launch, Perplexity is making the pplx-api publicly available, letting users tap into pplx-7b-online and pplx-70b-online for the first time. The access aims to power more helpful, up-to-date AI applications.
Stability AI Considers Sale Amid Investor Pressure Over Finances
The artificial intelligence startup Stability AI is exploring a potential sale of the company due to mounting scrutiny from investors over its financial stability. Stability AI, which achieved unicorn status in 2022, has held preliminary talks with multiple interested buyers, positioning itself as an attractive acquisition target, according to a Bloomberg report. However, the British firm could ultimately decide not to sell.
Tensions have risen with lead investor Coatue Management, which called for the resignation of Stability AI’s CEO in a letter citing concerns about leadership, staff departures, and precarious finances. While the startup earned praise for its ability to generate striking images from text prompts, its high expenses have surpassed revenue. With spending of $8 million per month and August revenue of just $1.2 million, Stability AI faces doubts about its ability to achieve profitability. Despite recent financing rounds, the pressure from key backers like Coatue appears to have Stability AI looking for an exit.
Nov 29, 2023
Sam Altman is back as CEO, Mira Murati as CTO and Greg Brockman as President
A Google AI Unveils 2.2 Million Hidden Materials, Rewriting Scientific Frontiers
In a groundbreaking revelation, Google’s artificial intelligence (AI) has unearthed 2.2 million previously unknown materials that had eluded scientific scrutiny. The discovery, a result of DeepMind’s tireless efforts, promises to revolutionize material science and propel technological advancements.
Nvidia CEO Predicts AI Will Rival Humans in 5 Years
Nvidia CEO Jensen Huang made a bold prediction about artificial intelligence at the New York Times DealBook conference. He said that in the next five years, AI systems will likely be able to perform at a “fairly competitive” level to human intelligence on benchmark tests. This would represent a major milestone towards artificial general intelligence (AGI).
Amazon Launches AI Model Benchmarking with Human Evaluators
Amazon Web Services (AWS) has introduced a new AI model evaluation service called Model Evaluation on Bedrock. The goal is to help developers better test AI models before deploying them. The service includes both automated benchmarking based on metrics like accuracy and robustness as well as human evaluation of subtle qualities like empathy.
Teams of human evaluators from AWS will be available for customers who want human assessment. The benchmarks can be run on popular third-party AI models available in AWS’s model repository, Amazon Bedrock. By evaluating models ahead of time, developers can select the right model for their needs and budget. They can also check that models meet responsible AI criteria around toxicity and fairness.
While automated tools are useful, Amazon says human judgment is still important to fully understand model performance. The service aims to promote safer, more transparent AI development.
Amazon Unveils Titan Image Generator at AWS re:Invent 2023, Strengthening AI Image Generation Landscape
Amazon has made a significant leap in the AI image generation field with the unveiling of the Titan Image Generator at the AWS re:Invent 2023 conference. This new tool, available in preview on Amazon’s AI development platform Bedrock, stands out in its ability to create new images from text descriptions or modify existing ones. Swami Sivasubramanian, VP for data and machine learning services at AWS, emphasized its ease of use in changing image backgrounds while preserving the main subject.
To combat the spread of misinformation and abuse imagery, all images created with the Titan Image Generator will feature a tamper-resistant invisible watermark, aligning with Amazon’s commitment to responsible AI use, signed alongside other tech giants like Google and Microsoft.
Stability AI Introduces Real-Time Text-to-Image Model SDXL Turbo
Stability AI has released a new text-to-image model called SDXL Turbo that can generate images from text prompts in real time. The model uses a novel technique called Adversarial Diffusion Distillation (ADD) that allows it to synthesize images in a single step while maintaining high image quality.
In blind tests, SDXL Turbo was able to outperform state-of-the-art diffusion models using 50x fewer steps without sacrificing quality. It can generate a 512×512 image in just 207ms on an A100 GPU.
SDXL Turbo represents major advances in speed and performance compared to previous diffusion models. The research paper detailing the technical innovations behind the model is available online.
Nov 28, 2023
Amazon Announced Amazon Q: A Generative AI Assistant
Amazon has unveiled Amazon Q, a generative AI-powered assistant designed to revolutionize how businesses interact with data and perform various tasks. Amazon Q enables fast, relevant responses to questions and problems, generating content and actions utilizing a company’s data and expertise. It’s tailored for business use, connecting to company data, systems, and over 40 built-in connectors, ensuring personalized, secure interactions based on user permissions.
AWS and NVIDIA Announce Strategic Collaboration to Offer New Supercomputing Infrastructure, Software and Services for Generative AI
Amazon Web Services (AWS) and NVIDIA are taking a monumental step in advancing generative AI technologies. They have announced an expansion of their strategic collaboration, aimed at delivering cutting-edge infrastructure, software, and services for generative AI applications. This collaboration will merge the strengths of both companies: NVIDIA’s next-gen GPUs, CPUs, and AI software, along with AWS’s advanced virtualization, security, and scalability solutions, creating an ideal environment for training foundational models and building generative AI applications.
Pika’s Breakthrough in AI Video Editing: A $55M Leap Forward
The generative AI sector is witnessing another remarkable advancement with the startup Pika’s recent funding success. Pika, known for developing AI tools to generate and edit videos, has secured a whopping $55 million in a funding round led by Lightspeed Venture Partners, featuring contributions from a string of notable investors. This milestone comes just half a year after Pika debuted, coinciding with the early access release of “Pika 1.0,” a suite of videography tools using a generative AI model to edit videos in diverse styles such as 3D animation, anime, and cinematic.
Google, Symphony unveil AI JV to help banks manage voice calls compliance risk
Symphony, a key player in market infrastructure and technology, has joined forces with Google to enhance voice analytics services for banks and investment firms. This collaboration responds to stringent regulatory demands for communication compliance, particularly highlighted by the U.S. Securities and Exchange Commission’s hefty fines totaling over $2 billion for non-compliance issues.
Symphony will integrate Google Cloud’s generative AI platform, Vertex AI, into its Cloud9 voice product, boosting its speech-to-text and natural language processing capabilities. This upgrade aims to streamline routine tasks and detect potential misconduct by analyzing anomalies in data. According to Zac Maufe, the head of Regulated Industries at Google Cloud, this generative AI can significantly transform the trading landscape.
Meta AI’s Breakthrough in Non-Invasive Speech Decoding: A Ray of Hope for Speech Impairments
Meta AI’s latest innovation presents a significant leap in assistive technology for individuals with speech impairments. Their non-invasive method, detailed in Nature Machine Intelligence, harnesses magneto-encephalography combined with machine learning to decode speech from brain activity. This approach offers a safer alternative to neural prostheses, which typically require risky brain surgery and the use of electrodes.
The research team conducted a study with 175 participants, achieving up to 41% average accuracy in decoding speech segments from brain activity, with some instances reaching 80% accuracy. This method stands out for its non-invasive nature, avoiding surgical risks and potentially simplifying implementation in real-world settings.
Cradle’s Groundbreaking AI in Protein Design: A Leap Forward with $24M Funding
Cradle, an innovative biotech and AI startup, is revolutionizing protein design with a generative approach, recently bolstering its capabilities with a substantial $24 million investment. The company, emerging from stealth mode just over a year ago amid rising interest in large language models, utilizes AI to interpret complex amino acid sequences akin to an alien programming language, a task beyond human capability.
Nov 27, 2023
ChatGPT Outperforms Students in Essay Writing, Study Reveals
In a pivotal study conducted by the University of Passau, published in Scientific Reports, it was found that the AI-based chatbot ChatGPT writes better English essays than secondary school students. This research demonstrates the remarkable progress in AI language models, particularly highlighting the advancements from ChatGPT version 3.5 to version 4.
The study, titled “A large-scale comparison of human-written versus ChatGPT-generated essays,” evaluated both AI-generated texts and human-written essays based on criteria set by the Ministry of Education of Lower Saxony. The outcome was unequivocal: ChatGPT’s performance surpassed that of the students, especially in terms of language mastery. Professor Steffen Herbold, Chair of AI Engineering at the University of Passau and initiator of the study, expressed surprise at the clear results, noting that both versions of the OpenAI chatbot scored higher than the student essays, with GPT-4 achieving the best score.
Generative AI Used By Three-Quarters Of Young Brits, Ofcom Says
In a groundbreaking revelation, Ofcom, the UK’s communications authority, has reported that a staggering three-quarters of young Britons are now utilizing generative AI technologies. This significant adoption rate highlights a remarkable shift in the way technology is integrated into daily life, particularly among the younger generation. Generative AI, known for its ability to create new content based on existing data patterns, has seen a surge in applications ranging from writing assistance to artistic creation.
Salesforce and Amazon Web Services Expand Partnership to Enable AI Innovation
Salesforce and Amazon Web Services (AWS) announced an expansion of their partnership at the AWS re:Invent conference on November 27. The companies aim to make it easier for customers to build trusted AI applications by enhancing data management and improving customer experiences.
Key expansions include deeper integration between AWS and Salesforce products to streamline purchasing and unify customer data. Salesforce will also expand its use of AWS technologies, while AWS adopts Salesforce’s customer data platform. Additionally, Salesforce will offer products like Data Cloud and Sales Cloud on the AWS Marketplace for simplified procurement.
UK Pupils Exploiting AI to Manufacture Explicit Images of Peers
A group of UK experts on technology and child safety have warned that students are using AI image generators to create realistic fake photos and videos depicting the sexual abuse of other children. Multiple schools have reported incidents of students making deepfakes that legally constitute child sexual abuse material. Experts warn that easy access to AI is enabling harmful behaviors and overwhelming safety systems. They call for urgent action from schools to implement stronger blocking systems and educate students on responsible technology use before the problem grows. The highly realistic AI creations are extremely difficult to distinguish from real abuse photos and remain illegal in the UK.
Neuralink, the brain implant startup founded by Elon Musk, has quietly raised an additional $43 million in funding, according to an SEC filing. This brings the total raised to $323 million. However, the company faces increasing scrutiny over its workplace culture and animal testing practices. Former employees have described a “culture of blame and fear” fostered by Musk, while outside groups allege Neuralink rushed tests on monkeys, leading to complications.
The company was investigated by the USDA for potential animal welfare violations, though the agency ultimately found no evidence. Most recently, lawmakers have asked the SEC to investigate Neuralink’s omission of details surrounding the deaths of over a dozen animals fitted with its implants. Despite its cutting-edge technology, Neuralink must now contend with growing controversies surrounding its internal operations and research methods.
Nov 26, 2023
Global Agreement Ensures Ethical AI Development: 193 Countries Commit to Secure-by-Design Approach
In a groundbreaking move, 193 countries have come together to adopt the first-ever global agreement on the Ethics of Artificial Intelligence (AI). This historic accord, facilitated by the United Nations Educational, Scientific and Cultural Organization (UNESCO), aims to define common values and principles necessary for the responsible and healthy development of AI.
- Ethical Framework: The adopted agreement provides a guide for constructing the legal infrastructure required to ensure the ethical advancement of AI. It addresses critical issues such as gender and ethnic bias, privacy threats, and the use of unreliable AI technologies in law enforcement.
- Positive Impact: While acknowledging the unprecedented challenges posed by AI, the agreement emphasizes its positive contributions. AI is already integral to everyday life, from booking flights to cancer screening. It also supports decision-making by governments and the private sector, addressing global problems like climate change and hunger.
- Data Protection: The recommendation underscores the need to protect data, going beyond existing efforts by tech firms and governments. It advocates for transparency, agency, and control over personal data, explicitly banning the use of AI systems for social scoring and mass surveillance.
- Climate and Environment: AI actors are encouraged to adopt energy-efficient methods, positioning AI as a powerful tool in the fight against climate change and environmental issues.
UNESCO Chief Audrey Azoulay aptly summarizes the significance of this milestone: “The world needs rules for artificial intelligence to benefit humanity. The Recommendation on the ethics of AI is a major answer. It sets the first global normative framework while giving states the responsibility to apply it at their level.”
Every week, we’ll meticulously curate a selection of stories from top AI media outlets and distill them into a digestible format, ensuring you stay up-to-date with the latest developments without having to spend hours browsing the web. From groundbreaking research to real-world applications, ethical debates to policy implications, AI Weekly Digest will be your essential guide to navigating the ever-evolving landscape of artificial intelligence. Join us on this exciting journey as we explore the future of AI together, one week at a time.