- NATURAL 20
- Microsoft's New Future of Work Report 2023 | Gates "You All Should Pay Attention" | AI and Work
Microsoft's New Future of Work Report 2023 | Gates "You All Should Pay Attention" | AI and Work
PLUS: Humanoid Robot Funding, OpenAI explores AI chip collaboration and more...
Microsoft's New Future of Work Report 2023 | Gates "You All Should Pay Attention" | AI and Work
Microsoft's recent report shines a spotlight on AI, highlighting its potential to reshape how we work, especially remotely. Bill Gates and Microsoft CEO Satya Nadella are big believers. They've been in the AI game early, focusing on making work less of a drag and more efficient, from writing emails to tackling security tasks.
Microsoft's betting big, weaving AI into the fabric of work life, making sure we're all speaking AI fluently. And as we get more tangled up with our digital co-workers, figuring out how to work alongside them without going off the rails is the next big challenge.
Microsoft revenue boosted by AI excitement and cloud strength
Microsoft's revenue soared due to strong cloud computing demand and the hype around artificial intelligence. In the last quarter of 2023, the company's cloud division, including Azure, saw a 20% revenue jump to $25.9 billion, exceeding expectations. Overall, Microsoft's revenue hit a record $62 billion, with earnings per share at $2.93, surpassing forecasts.
Investments in AI, especially a partnership with OpenAI and a commitment of up to $13 billion, have put Microsoft at the forefront of the generative AI field. This move has driven Microsoft's shares up by over 60% in the past year, making it more valuable than Apple with a market cap above $3 trillion.
Microsoft CEO Satya Nadella highlighted the company's progress in applying AI on a large scale. The Azure cloud platform benefited significantly from this, with AI-related demand boosting its revenues. Microsoft is also expanding its AI capabilities by investing in data centers and servers, without hurting profit margins due to high cloud service demand and cost control.
Alphabet posted $307.4 billion in revenue for the 2023. It also spent $2.1 billion on layoffs.
Alphabet, the big tech giant, raked in a whopping $307.4 billion in 2023, a solid 13% jump from the previous year. But it wasn't all smooth sailing. They shelled out a hefty $2.1 billion for employee severance, linked to the job cuts announced early in 2023.
And that's not counting the latest round of layoffs this month. This info comes straight from their latest earnings report.
Humanoid Robot Startup Figure AI in Funding Talks With Microsoft, OpenAI
Figure AI Inc., a company developing robots that act like humans, is chatting with Microsoft and OpenAI about getting up to $500 million. They're still looking for more folks to chip in.
The plan might have Microsoft tossing in $95 million and OpenAI adding $5 million. All this info comes from someone who knows the deal but wants to stay under the radar.
OpenAI CEO Sam Altman explores AI chip collaboration with Samsung and SK Group
OpenAI's CEO, Sam Altman, is scoping out a deal with Samsung and SK Group in South Korea. He's eyeing a partnership to make AI chips, crucial for AI tech, especially since there's a chip shortage looming.
Altman checked out Samsung's Pyeongtaek factory and chatted with bigwigs from both companies. The goal? To possibly work together and invest in AI chip production, with OpenAI potentially buying High Bandwidth Memory from them. Altman's also meeting with local AI startups tied to Samsung and SK.
Meanwhile, OpenAI, already big with ChatGPT and working with Microsoft, is looking to be a heavy-hitter in AI chips too. They're talking with big investors like G42 and SoftBank for a global network of chip factories. This move could make OpenAI a key player in the AI game, from chips to consumer AI hardware.
Robot trained to read braille at twice the speed of humans
Cambridge University's team whipped up a robot that reads braille way faster than we can, hitting about 315 words a minute with 90% right answers. This robot, using AI smarts, glides over braille like a pro, not for helping folks directly, but to test out super-sensitive robot hands or fake limbs. Our fingertips are ace at feeling tiny details and knowing how much oomph to use, like not smashing an egg or dropping a bowling ball. Replicating this in a robot hand, without guzzling energy, is tough cookies.
The team's using a fingertip camera and smart algorithms to tackle reading braille, which needs a lot of sensitivity due to its tiny dot patterns. Their robot doesn't just poke at each letter like old-school robotic readers; it smoothly moves along the text. After training their system with blurred images to make it sharper at reading, their robot nailed it, getting the balance of speed and accuracy just like a human would. They're aiming to shrink this tech to fit on a humanoid hand or skin, getting a hand from the Samsung Global Research Outreach Program.
InternLM-XComposer2 is a top-notch model that's ace at mixing text and images to create unique content. It's smarter than typical models, handling various inputs like outlines or images. Using a special technique, PLoRA, it balances sharp image understanding with creative text crafting. Tests show it's a real champ, even beating big names like GPT-4V and Gemini Pro in some areas. It's especially good at understanding and creating lengthy, mixed text-image content. This model, with its 7 billion parameters, is up for grabs online.
MoE-LLaVA is a cutting-edge framework for large vision-language models. It uses a novel 'Mixture of Experts' approach, allowing massive models to run with constant computational costs and improved performance. MoE-LLaVA stands out in visual understanding and reduces output errors, performing comparably to larger models with fewer parameters. This breakthrough offers a new direction for efficient and effective multi-modal learning.
"Rephrasing the Web: A Recipe for Compute and Data-Efficient Language Modeling," introduces WRAP, a method to train large language models (LLMs) more efficiently. It involves paraphrasing web content in various styles to enhance training. This approach speeds up training, improves model performance, and ensures better question-answering accuracy. The paper highlights how rephrased data can be more beneficial than raw web data for training LLMs.
SERL is a software suite designed for sample-efficient robotic reinforcement learning. It's crafted to address the complexity of implementing robotic RL methods by offering a library with efficient off-policy deep RL methods, tools for computing rewards, resetting environments, and a top-notch controller for a popular robot. SERL demonstrates its prowess with quick and robust learning in tasks like PCB assembly, cable routing, and object relocation, showing better results than existing methods. The implementation is open-source, aiming to advance robotic RL development.
"Motion-I2V" introduces a new framework for creating consistent and controllable image-to-video generation. It uniquely separates the process into two stages: first, predicting motion trajectories from images, and then enhancing these with motion-augmented attention. This method generates more reliable videos even with significant motion or viewpoint changes. Additionally, it allows users to control motion paths and areas, offering greater control than text-based instructions. The approach also supports zero-shot video-to-video translation.
"Scaling Face Interaction Graph Networks to Real World Scenes" addresses the challenge of simulating real-world object dynamics for applications like robotics, engineering, graphics, and design. It introduces a method to reduce memory usage in graph-based learned simulators, making them more efficient. Additionally, the paper presents a perceptual interface that converts real-world scenes into a structured representation, allowing graph network simulators to process them. This approach enables the application of learned simulators to real-world scenes captured from multiple camera angles, expanding their use in scenarios where only perceptual information is available.
Recraft - Generative AI design tool that lets users create and edit digital illustrations, vector art, icons, and 3D graphics in a uniform brand style.
Norrm Ai - offers AI agents for regulatory compliance, automating tasks like risk identification and report generation, enhancing efficiency and understanding of complex regulations.
Potion - video prospecting tool using AI to create personalized sales videos, increasing engagement and sales, user-friendly and integrates with many tools. Ideal for various professionals.
Assembly - workplace engagement platform that enhances team collaboration and recognition, integrating with various tools to create a more engaging and rewarding work environment.
Murf AI - AI voice generator that allows users to create natural-sounding voiceovers quickly. It offers lifelike AI voices for various applications, such as podcasts, videos, and presentations.
Replika’s new AI therapy app tries to bring you to a zen island
Replika's teamed up with Blush to drop Tomo, an AI wellness app. It's like a mini-vacay with an AI guide, offering talk therapy, yoga, and more. The trial's free for three days, then it's $7.99 a week or $49.99 a year. While it's got the usual guided meditations, chatting with Tomo feels no different than your average chatbot. It's iPhone-only for now, with an Android version coming. Replika's been in hot water before, but they say Tomo's chats stay private. THE VERGE
AI server maker Super Micro boosts sales guidance — another sign to stick with Nvidia
Super Micro, a key Nvidia partner, has significantly upped its fiscal 2024 revenue forecast. This is another nudge to stick with Nvidia, a top dog in AI chips, whose stock is on a roll, hitting record highs. The move underscores Nvidia's solid market position and hints at the booming demand for AI servers. This update is a big deal for investors eyeing the tech and AI space. CNBC
Yelp will use AI to tell you if that burger’s any good
Yelp's rolling out a new AI feature on iOS that summarizes user reviews. Using large language models, it'll spotlight key details like atmosphere, popular dishes, and prices. This AI summary pops up for food and nightlife spots with enough recent reviews. Yelp's also jazzing up its iOS app with a new home feed and photo-centric search results. Android users will get these updates later this year, but for now, they can use the "surprise me" feature introduced on iOS last year. THE VERGE
Microsoft AI engineer says company thwarted attempt to expose DALL-E 3 safety problems
Shane Jones, a Microsoft AI engineering leader, claims to have found serious flaws in OpenAI's DALL-E 3 image generator. According to Jones, these vulnerabilities could be exploited to create violent and explicit images. Despite his concerns, he alleges that Microsoft hindered his efforts to publicly address these issues. Jones' worries intensified following the recent spread of explicit deepfake images of Taylor Swift. He believes that products like Microsoft Designer, which uses DALL-E 3 technology, contribute to the ease of generating harmful AI images. He reportedly discovered this vulnerability in December and notified Microsoft and OpenAI. GEEKWIRE
AI can better retain what it learns by mimicking human sleep
Researchers at the University of Catania, Italy, have developed a new method of training artificial intelligence (AI) called wake-sleep consolidated learning (WSCL), inspired by the way humans consolidate memories during sleep. WSCL aims to address the issue of "catastrophic forgetting," where AI models lose the ability to perform previously learned tasks when trained for new ones. In WSCL, AI models have periods of "sleep" where they review previous lessons and merge new and old information, preventing the loss of previously acquired skills. This approach led to a significant increase in accuracy and the ability to remember old tasks in AI models, according to the researchers. NEWSCIENTIST
New AI model designs proteins to deliver gene therapy
Researchers at the University of Toronto have used artificial intelligence (AI) to redesign a crucial protein involved in the delivery of gene therapy. Their work, published in Nature Machine Intelligence, focused on optimizing proteins to mitigate immune responses, thereby improving the efficacy of gene therapy and reducing side effects. The study targeted hexons, a fundamental protein in adenovirus vectors used in gene therapy. These vectors often trigger immune responses, reducing their effectiveness. To address this issue, the researchers used AI to custom-design hexon variants that are distinct from natural sequences, making them unrecognizable by the immune system. PHYS ORG
What'd you think of today's edition?
What are MOST interested in learning about AI?
What stories or resources will be most interesting for you to hear about?