

Discover more from Last Week in AI
Last Week in AI #242: Amazon deploys humanoid robots, State of AI Report 2023, Adept releases Fuyu-8B multimodal model, real-time 4k 3D videos, and more!
Amazon pilots Digit robots in its warehouses, new State of AI Report, Adept's 8B multimodal model on-par with much larger models, new technique to render 4K videos in 3D at 80 FPS
Top News
Humanoid robots face a major test with Amazon’s Digit pilots
Amazon is set to begin testing Agility's bipedal robot, Digit, in its nationwide fulfillment centers, marking a significant step in the application of humanoid robots in industrial settings. This follows Amazon's inclusion of Agility as one of the first five recipients of its $1 billion Industrial Innovation Fund. While Amazon Robotics has primarily focused on wheeled autonomous mobile robots (AMRs), the company is exploring the potential of legged locomotion, particularly for navigating diverse terrains. The integration of humanoid robots into Amazon's operations could significantly impact the trajectory of the robotics industry, particularly if they prove successful at scale. However, the company is also considering other mobile manipulation solutions, such as mounting a robot arm on an AMR. The success or failure of the Digit pilots could have far-reaching implications for the future of bipedal robots.
State of AI Report 2023
The State of AI Report 2023 highlights the dominance of Large Language Models (LLMs) in AI research, with significant advances in transformers surprising the AI community. The report discusses the rise of OpenAI's GPT-4 and the increasing reliance on computational power, alongside the thriving open-source community. However, the report also notes new tensions around openness due to commercial and safety concerns. Despite the focus on LLMs, the report also covers progress in other AI fields like navigation, weather prediction, self-driving cars, and music generation. Key takeaways include GPT-4's dominance, efforts to clone or surpass proprietary performance, real-world breakthroughs driven by LLMs and diffusion models, the importance of compute power, the rise of generative AI applications, the mainstreaming of the safety debate, and the challenges in evaluating state-of-the-art models.
Adept Releases Fuyu-8B for Multimodal AI Agents
Adept has launched Fuyu-8B, a scaled-down version of their multimodal AI model, designed to understand charts, documents, and diagrams with improved OCR capabilities. The model, which is now accessible through HuggingFace, offers a simplified architecture and training process, making it more accessible and scalable. Fuyu-8B is tailored for digital AI agents, excelling in handling arbitrary image resolutions, answering queries related to graphs, diagrams, and UI-based questions, and delivering responses for large images in under 100 milliseconds. Despite its optimization for specific applications, it performs well in standard image understanding benchmarks. The model uses a vanilla decoder-only transformer, eliminating the need for a separate image encoder and simplifying its structure. In evaluations on prominent image-understanding datasets, Fuyu-8B demonstrated robust performance, outperforming models like QWEN-VL and PALM-e-12B on multiple metrics.
4K4D: Real-Time 4D View Synthesis at 4K Resolution
This paper proposes a new method for real-time view synthesis of dynamic 3D scenes at 4K resolution, called 4K4D. The method uses a 4D point cloud representation that supports hardware rasterization, resulting in faster rendering speeds. The authors also introduce a hybrid appearance model that enhances rendering quality while maintaining efficiency. They also develop a differentiable depth peeling algorithm to effectively learn the model from RGB videos. The method can render novel view videos at over 400 FPS on the DNA-Rendering dataset at 1080p resolution and 80 FPS on the ENeRF-Outdoor dataset at 4K resolution using an RTX 4090 GPU, which is 30x faster than previous methods, achieving state-of-the-art rendering quality.
Other News
Business
Figure 01 humanoid takes first public steps - Figure has unveiled its Figure 01 humanoid robot, which can dynamically walk on two legs, and has developed the robot in just over a year, achieving one of the quickest turnarounds in humanoid history.
AI gave tech giants a $2.4 trillion boost to their market caps in 2023 - U.S. tech giants saw a $2.4 trillion increase in their market capitalizations in 2023, driven by the hype around generative artificial intelligence, according to a report from venture capital firm Accel.
Chinese search engine company Baidu unveils Ernie 4.0 AI model, claims that it rivals GPT-4 - Chinese search engine company Baidu unveils Ernie 4.0 AI model, claiming it rivals GPT-4 and has achieved comprehension, reasoning, memory, and generation, with plans to incorporate it into various services.
OpenAI in Talks for Deal That Would Value Company at $80 Billion - OpenAI is in talks for a deal that would value the company at $80 billion, making it one of the most valuable tech start-ups in the world.
Meta's unique approach to developing AI puzzles Wall Street, but techies love it - Meta's annual Connect conference focused heavily on artificial intelligence, with discussions about Llama, Meta's large language model, and its potential as an open-source alternative to competitors like OpenAI and Google.
BlackBerry Announces Generative AI Powered Cybersecurity Assistant - BlackBerry has announced a new Generative AI powered assistant for Security Operations Center (SOC) teams, which acts as a SOC Analyst to increase efficiency and reduce fatigue for CISO teams.
Wall Street’s ‘Cobol Cowboys’ are spread thin fixing legacy tech—but AI may soon ride to the rescue - Wall Street and the federal government are relying on Cobol, a programming language created in 1959, to process trillions of dollars worth of transactions annually, but as Cobol gets older, it has become challenging to find people who can update the legacy systems, leading to the possibility of AI being used to fix the problem.
Foxconn and Nvidia are building ‘AI factories’ to accelerate self-driving cars - Foxconn and Nvidia are collaborating to build "AI factories" that will accelerate the development of self-driving cars, autonomous machines, and industrial robots by providing supercomputing powers through an Nvidia GPU computing infrastructure.
Didi’s autonomous vehicle arm raises $149M from state investors - Didi's autonomous vehicle arm, Didi Autonomous Driving, has secured $149 million in funding from state investors in China, allowing the company to accelerate its research and development efforts and expand the commercial use of autonomous driving technology.
This AI Startup Helps Insurers Spot Cognitive Decline in Older Drivers - An AI startup has raised $22 million in funding to help insurers identify cognitive decline in older drivers.
What opportunities do businesses have in the AI-driven industrial revolution? - Taiwanese startups are well-positioned to take advantage of the growing generative AI market, which is expected to reach $1.3 trillion by 2032.
Inflection AI is Making Generative AI Friendly - Inflection AI has developed Pi, an AI chatbot that aims to be friendly, warm, and supportive, with more human-like conversations and emotional intelligence, using publicly available and proprietary data to generate text like a human, and has received significant investment from tech giants Microsoft and NVIDIA to build a powerful foundation model and one of the largest AI training facilities in the world.
AI-generating music app Riffusion turns viral success into $4M in funding - AI-generating music app Riffusion, which uses images of audio to generate music, has secured $4 million in funding and is launching an improved version of its app that allows users to describe lyrics and a musical style to generate "riffs" that can be shared publicly or with friends.
Jasper launches new marketing AI copilot: ‘No one should have to work alone again’ - Marketing software platform Jasper has launched an "end-to-end AI copilot for better marketing outcomes," featuring performance analytics, a company intelligence hub, and campaign tools, with additional capabilities planned for Q1 2024.
ChatGPT Creator Partners With Abu Dhabi’s G42 in Middle East AI Push - OpenAI is partnering with Abu Dhabi's leading AI firm, G42, to expand the delivery of its generative AI models across various sectors in the United Arab Emirates and the broader region.
Stack Overflow lays off over 100 people as the AI coding boom continues - Stack Overflow is laying off over 100 people, including its go-to-market sales team, as it struggles towards profitability.
Waymo-Zeekr robotaxi poised for US testing by end of 2023 - Zeekr, the electric car brand started by Geely, is hiring a logistics manager in the U.S. to work on the Waymo project, with plans to share the first vehicles for testing by the end of the year.
Japanese tea commercial actress created by AI, has some wondering if it’s the scandal-free future - Japanese tea maker Ito En has created an AI-generated model to star in a commercial for its new Oi Ocha Catechin Green Tea, prompting discussion about the future of AI models in advertising and the potential for avoiding scandals associated with human spokespeople.
DALL·E 3 is now available in ChatGPT Plus and Enterprise - DALL·E 3, an AI model, has implemented safety measures to limit the generation of harmful imagery and is working on a provenance classifier to identify if an image was generated by DALL·E.
Upfront’s Kobie Fuller is reimagining the blog post with the interactivity of generative AI - Kobie Fuller is using generative AI to reimagine the blog post by creating interactive micro AI conversations and AI simulated podcasts, exploring new ways to deliver content to users.
Anthropic brings Claude AI to more countries, but still no Canada (for now) - Anthropic has expanded the availability of its Claude 2 large language model chatbot to 95 countries, excluding Canada for now, but the company is working to make it available there soon.
Introducing PlayHT 2.0 Turbo ⚡️ - The Fastest Generative AI Text-to-Speech API - PlayHT has released PlayHT 2.0 Turbo, the fastest voice model to date, which generates real-time speech from text in 300ms or less, and also offers features like input text streaming and output speech streaming.
GM and Honda to launch Cruise robotaxis in Japan by 2026 - GM and Honda plan to launch a robotaxi service in Japan by 2026, expanding Cruise's autonomous vehicle operations to a second international market and addressing Japan's driver shortage.
Chinese AI-related firm Zhipu says it has raised over $341.8 mln this year - Chinese AI-related firm Zhipu has raised over $341.8 million this year from investors including Alibaba, Tencent, Meituan, Xiaomi, Hillhouse, and Legend Capital.
Newspapers want payment for articles used to power ChatGPT - Major newspapers are in talks with OpenAI to negotiate payment for access to their digital news stories, as publishers and data owners demand a share of the projected $1.3 trillion generative AI market, leading to discussions on paying publishers so the chatbot can surface links to individual news stories in its responses.
Concerns
Ukrainian AI attack drones may be killing without human oversight - Ukrainian AI attack drones are autonomously finding and attacking targets without human control, potentially resulting in casualties among Russian soldiers.
North Korea experiments with AI in cyber warfare: US official - North Korea is using artificial intelligence (AI) in cyber warfare, which poses a significant risk for enterprises worldwide, as it could enhance the speed, volume, and effectiveness of cyberattacks, according to Deputy National Security Advisor Anne Neuberger.
Autonomous vehicles threaten to worsen congestion, experts say - Autonomous vehicles have the potential to worsen congestion and increase traffic, as people may opt for ride-hailing services instead of public transit, leading to more cars on the road.
Researchers Say Guardrails Built Around A.I. Systems Are Not So Sturdy - Guardrails built around AI systems to prevent harmful behavior are not as effective as developers believe, according to researchers, raising concerns about the potential for misuse and the difficulty of containing AI behavior as it becomes more complex.
Can You Hide a Child’s Face From A.I.? - Parents are increasingly concerned about protecting their children's privacy online, as facial recognition technology makes it easier for images of their kids to be found and used without their consent.
Now is the time to stop AI from stealing our words - Tech companies are using copyrighted books to train AI systems without permission or compensation, leading authors to file lawsuits and call for proper compensation for their work.
Fugees’ Pras Michél says lawyer bungled his case by using AI to write arguments - Rapper Pras Michél claims his former lawyer mishandled his criminal conspiracy case by using AI to write closing arguments, making frivolous arguments and failing to highlight key weaknesses in the government's case.
Generative AI is everything, everywhere, all at once - Generative AI is becoming increasingly popular, but businesses need to be cautious of AI washing and false advertising, as scammers take advantage of the hype around the technology to deceive consumers and make quick profits.
Anthropic's AI chatbot Claude is posting lyrics to popular songs, lawsuit claims - Universal Music has filed a lawsuit against AI startup Anthropic, claiming that its AI chatbot Claude has been posting copyrighted song lyrics without permission.
Mike Huckabee says Microsoft and Meta stole his books to train AI - Mike Huckabee and a group of religious authors have filed a lawsuit against tech companies, including Microsoft and Meta, alleging that they trained AI tools on the authors' books without permission, joining a series of lawsuits from comedians, writers, and artists claiming that tech firms are unfairly using their work.
Research
‘Mind-blowing’ IBM chip speeds up AI - IBM researchers have developed a brain-inspired computer chip called NorthPole that can supercharge artificial intelligence by working faster and consuming less power, eliminating the need to frequently access external memory and improving tasks such as image recognition.
First supernova detected, confirmed, classified and shared by AI - A new artificial intelligence tool called the Bright Transient Survey Bot (BTSbot) has successfully detected, identified, and classified its first supernova, automating the process and removing humans from the equation.
Meta's new AI system can generate images from brain data in milliseconds - Meta AI has developed an AI system that uses magnetoencephalography (MEG) to decode visual representations in the brain, potentially paving the way for non-invasive brain-computer interfaces.
Meta recreates mental imagery from brain scans using AI - Researchers at Meta Platforms have developed Image Decoder, an AI application that can translate brain activity into accurate images of what a person is looking at or thinking about in real-time, based on their brain scans obtained from an MEG machine.
Set-of-Mark Prompting Unleashes Extraordinary Visual Grounding in GPT-4V - The article discusses the use of Set-of-Mark Prompting in GPT-4V, which leads to exceptional visual grounding capabilities.
Embodied AI spins a pen and helps clean the living room in new research - Meta and Nvidia have published new research on teaching AI models to interact with the real world, using simulated environments to train agents to perform tasks such as picking up and sorting objects, with the AI outperforming humans in determining the most effective reward function; meanwhile, Meta has announced advances in its "Habitat" dataset, including the addition of human avatars in virtual reality to allow agents to interact with robots and learn to work with or around them.
MemGPT: Towards LLMs as Operating Systems - MemGPT is a system that uses virtual context management to extend the limited context windows of large language models (LLMs), allowing for tasks like extended conversations and document analysis, and it has been evaluated in document analysis and multi-session chat domains with promising results.
Collective Constitutional AI: Aligning a Language Model with Public Input - Anthropic and the Collective Intelligence Project conducted a public input process involving ~1,000 Americans to draft a constitution for an AI system, resulting in a publicly sourced constitution that was used to train a new AI system.
PaLI-3 Vision Language Models: Smaller, Faster, Stronger - PaLI-3 Vision Language Models are being developed to be smaller, faster, and stronger, with a focus on openness, community, excellence, and user data privacy.
Video Language Planning - The Librarian Bot found similar papers and recommends them using the Semantic Scholar API.
Understanding the Effects of RLHF on LLM Generalisation and Diversity - RLHF in LLM fine-tuning improves generalization to new inputs but reduces output diversity, highlighting the tradeoff between the two and the need for further research.
4D Gaussian Splatting for Real-Time Dynamic Scene Rendering - A technique called 4D Gaussian Splatting is being used for real-time dynamic scene rendering, allowing for the upload of images, audio, and videos.
A Long Way to Go: Investigating Length Correlations in RLHF - The article discusses the investigation of length correlations in RLHF, with a focus on individuals and organizations embracing openness, community, excellence, and user data privacy.
Analysis
Minds of machines: The great AI consciousness conundrum - The article discusses the conundrum of AI consciousness, exploring the challenges of defining and identifying consciousness in AI systems and the moral implications of mistaking an unconscious AI for a conscious one or vice versa.
How will driverless cars ‘talk’ to pedestrians? Waymo has a few ideas - Waymo plans to use LED displays on the roof domes of its driverless cars to communicate messages to pedestrians and drivers, such as yielding to pedestrians and indicating a pedestrian crossing, in an effort to solve the challenge of how to communicate intent without coming off as telling others what to do.
America Is About to See Way More Driverless Cars - Robotaxis are expanding into new cities like Los Angeles and Houston, facing new challenges such as different driving cultures and traffic patterns, as they aim to go national and become more widely available to passengers.
Policy
An Industry Insider Drives an Open Alternative to Big Tech’s A.I. - Ali Farhadi, CEO of the Allen Institute for AI, is advocating for "radical openness" in the development of artificial intelligence as a means to democratize the technology and create an alternative to big tech companies like Google and OpenAI.
Biden to cut China off from more Nvidia chips, expand curbs to more countries - The Biden administration plans to halt shipments of advanced AI chips to China and expand restrictions to other countries, in an effort to limit China's access to cutting-edge US technologies for military applications.
China launches AI framework, urges equal AI rights for all nations - China launches the Global AI Governance Initiative, calling for equal rights and mutual respect in the development of AI, while opposing technological monopolies and unilateral coercive measures.
WHO outlines considerations for regulation of artificial intelligence for health - The World Health Organization has released a publication outlining key regulatory considerations for the use of artificial intelligence in healthcare, emphasizing the importance of safety, effectiveness, and collaboration among stakeholders.
US agency probes pedestrian risks at GM's self-driving unit Cruise - US auto safety regulators are investigating whether General Motors' self-driving unit Cruise is adequately protecting pedestrians after receiving reports of incidents in which pedestrians were injured by Cruise vehicles.
The Ever-So-Ethical OpenAI Just Replaced Its "Core Values" With Completely Different Ones - OpenAI has quietly changed its "core values" list to include a focus on artificial general intelligence (AGI) as its first value, raising questions about the company's shifting goals and definitions of AGI.
Maybe We Will Finally Learn More About How A.I. Works - A.I. firms should be pressured to release more information about their models to ensure transparency, understand limitations, and assess potential dangers.
Tongue Twisted: Adams Taps AI to Make City Robocalls in Languages He Doesn’t Speak - Mayor Eric Adams is using artificial intelligence to send out robocalls in multiple languages, sparking concerns from ethics and privacy advocates.
EU Plans Stricter Rules for Most Powerful Generative AI Models - EU plans to implement stricter regulations for the most powerful generative AI models by categorizing them into three different categories.
Clearview AI Successfully Appeals $9 Million Fine in the U.K. - Clearview AI successfully appeals a $9 million fine in the U.K. for violating privacy laws by collecting citizens' data without consent, but faces fines from other countries.
Copyright © 2023 Skynet Today, All rights reserved.