Last Week in AI #234: Nvidia's moat in AI chips, State of AI Q2'23, Meta's new coding AI, DeepMind's algorithm to improve LLMs by themselves, and more!

Nvidia has 70% of AI chip sales, AI funding drops but deal size grows, Meta's code-specialized Llama 2 approaches GPT-4, more evidence of LLMs self-improvement

Top News

How Nvidia Built a Competitive Moat Around A.I. Chips

Nvidia has established a dominant position in the AI chip market, accounting for over 70% of sales and holding a significant role in training generative AI models. This dominance has been achieved through early recognition of AI trends, tailoring chips for complex AI tasks, and developing key software for AI development. Nvidia's strategy also includes offering customers access to specialized computers, computing services, and other tools, effectively making it a one-stop shop for AI development. Despite competition from tech giants like Google, Amazon, Meta, and IBM, Nvidia's market position remains strong, with a recent projection of a 64% leap in quarterly revenue and a market capitalization surpassing $1 trillion, making it the world's most valuable chip maker.

State of AI Q2'23 Report

The State of AI Q2'23 Report reveals a 38% quarter-over-quarter (QoQ) drop in global AI funding, settling at $9.4B, but this remains on par with Q4'22 figures. Despite the decline, the average AI deal size has increased by 48% in 2023 YTD, driven by mega-rounds. The quarter also saw the emergence of seven new AI unicorns, five of which are generative AI companies. M&A deals in the AI sector increased by 13% QoQ, while public exits remained low. Notably, funding for Canadian AI startups surged by 13x QoQ, largely driven by mega-rounds to generative AI model developer Cohere and AI traffic management startup Miovision. The report also highlights that over 40% of the 590 deals in Q2'23 went to US-based startups, and generative AI continues to attract significant investor attention.

Introducing Code Llama, an AI Tool for Coding

Code Llama, a large language model (LLM) designed to generate and discuss code, has been released for both research and commercial use. This AI tool is a code-specialized version of Llama 2, trained on code-specific datasets to enhance its coding capabilities. It can generate code and natural language about code, support code completion, and debugging in popular programming languages like Python, C++, Java, PHP, Typescript, C#, Bash, and more. Three versions of Code Llama, with 7B, 13B, and 34B parameters respectively, have been released to cater to different serving and latency requirements. Two additional variations, Code Llama – Python and Code Llama – Instruct, have been fine-tuned for Python code and instruction alignment respectively. The tool aims to make developer workflows more efficient and lower the barrier to entry for coding learners.

Reinforced Self-Training (ReST) for Language Modeling

Researchers from Google DeepMind recently published the Reinforced Self-Training (ReST) method for language modeling. ReST is a novel approach that combines reinforcement learning and self-training to improve the performance of language models. The method involves training a model on a large dataset, then using the model's own predictions to generate new training data. This data is then used to further train the model, a process that is repeated iteratively. The reinforcement learning aspect comes into play by rewarding the model for making accurate predictions and penalizing it for incorrect ones. ReST can bring significant improvements to LLMs without obtaining costly human labeled data, and it's one among many recent works that hint at the promise of LLMs' self-improvement capabilities.s

Other News

Research

A Stroke Stole Her Ability to Speak at 30. A.I. Is Helping to Restore It Years Later. - Scientists have made a remarkable advance in neuroscience and artificial intelligence by using implanted electrodes to decode brain signals and enable a stroke patient to speak again through an avatar on a computer screen.

Parkinson's disease could be detected early with AI scans, scientists say - AI scans of eye scans could potentially detect Parkinson's disease before symptoms appear, according to a study analyzing retinal markers in a dataset of over 150,000 patients.

AI helps robots manipulate objects with their whole bodies - MIT researchers have developed an AI technique called smoothing that simplifies contact-rich manipulation planning for robots, enabling them to manipulate objects with their whole bodies, potentially leading to the use of smaller, more mobile robots in factories and exploration missions.

Graph of Thoughts: Solving Elaborate Problems with Large Language Models - A framework called Graph of Thoughts (GoT) is introduced, which advances prompting capabilities in large language models (LLMs) by modeling LLM thoughts as an arbitrary graph, enabling the combination and enhancement of thoughts, and offering advantages over existing paradigms.

Open-sourcing SQLCoder: a state-of-the-art LLM for SQL generation - Defog has open-sourced SQLCoder, a state-of-the-art LLM for converting natural language questions to SQL queries, which outperforms major open-source models and slightly outperforms gpt-3.

Prompt2Model: Generating Deployable Models from Natural Language Instructions - A new method called Prompt2Model is proposed, which uses natural language task descriptions to train special-purpose models that outperform large language models (LLMs) while being significantly smaller, allowing for easier deployment and reliable performance estimates.

Language to rewards for robotic skill synthesis - A new approach to teaching robots novel actions through natural language input is proposed, using reward functions as an interface to bridge the gap between language and low-level robot actions.

Teach LLMs to Personalize -- An Approach inspired by Writing Education - A general approach for personalized text generation using large language models (LLMs) is proposed, inspired by the practice of writing education, which involves multiple stages and a multitask setting to improve generation ability.

Introducing IDEFICS : An Open Reproduction of State-of-the-art Visual Langage Model - IDEFICS is an open-access visual language model based on Flamingo, developed by DeepMind, that accepts image and text inputs and produces text outputs, aiming to bring transparency to AI systems and serve as a foundation for open research in multimodal AI systems.

How Midjourney Reads a Mugshot - Donald Trump's mugshot is analyzed using generative AI, specifically Midjourney's "/describe" function, which reveals the different prompts that would generate the mugshot and explores the concept of "Sumatraism" as a genre of style markers in AI image generation systems.

Meta releases an AI model that can transcribe and translate close to 100 languages - Meta has released an AI model called SeamlessM4T that can transcribe and translate close to 100 languages across text and speech, representing a "significant breakthrough" in the field of AI-powered speech-to-speech and speech-to-text.

Introducing Arthur Bench: The Most Robust Way to Evaluate LLMs - Arthur Bench is an open-source evaluation tool that helps businesses compare and select the best large language models (LLMs) for their applications, considering factors such as model performance, budget optimization, and translation of academic benchmarks to real-world scenarios.

Applications

How Artificial Intelligence Gave a Paralyzed Woman Her Voice Back - A breakthrough brain implant and digital avatar allow a stroke survivor to speak with facial expressions for the first time in 18 years.

Chinese firm launches WonderJourney satellite with AI-powered ‘brain’ - Chinese firm launches WonderJourney satellite with AI-powered ‘brain’ that has an onboard intelligent processing unit, making it the first satellite with AI at its core.

California deploys AI-powered wildfire detection systems - California's main firefighting agency, Cal Fire, is using AI-powered cameras to detect visual signs of wildfires and quickly warn firefighters of emerging threats, with a 40 percent success rate in detecting fires before 911 calls were made.

Meet Apollo, your new humanoid coworker - Apollo, a new humanoid robot, has been unveiled as a versatile and human-like robot designed to work alongside humans in various industries, with the potential to take over dangerous work environments; the robot is being positioned as a solution to labor shortages and retention issues in critical industries, but critics worry about job losses.

Snapchat is expanding further into generative AI with ‘Dreams’ - Snapchat is expanding its generative AI features with a new feature called "Dreams," which will allow users to place AI-generated selfies of themselves and their friends in imaginative backgrounds.

Developers are now using AI for text-to-music apps - Developers have created text-to-music apps using AI, such as Songburst and CassetteAI, which generate music clips based on prompts and offer features like editing and mixing, aiming to make music production accessible to everyone.

IBM taps AI to translate COBOL code to Java - IBM has developed Code Assistant for IBM Z, an AI-powered tool that translates COBOL code into Java, aiming to assist businesses in modernizing their mainframe applications while preserving performance and security.

Midjourney Inpainting Has Arrived. Let’s Play! - Midjourney has launched a new inpainting feature called "Vary (Region)" that allows users to modify selected regions of an image while keeping the rest intact, offering a range of practical use cases such as making subtle tweaks, removing unwanted items, adding new elements, and fixing AI artifacts.

Business

40% of workers will have to reskill in the next three years due to AI, says IBM study - 40% of workers will need to reskill in the next three years due to AI implementation, but a new IBM study shows that rather than fearing the technology, people should leverage it for their own gain, as executives expect generative AI to augment roles rather than replace them.

University of Michigan to provide custom AI tools to campus community - The University of Michigan is providing custom generative artificial intelligence platforms to its entire community, offering three levels of AI tools that emphasize equity, accessibility, and privacy.

New York's largest health system is pushing into generative AI with help from a former Amazon exec and a $100 million venture partnership - New York's largest health system, Northwell Health, is partnering with Aegis Ventures to develop generative AI models that can reduce the time doctors spend on paperwork and administrative tasks.

The AI-powered Adobe Express is now generally available - Adobe has incorporated AI features into its Adobe Express application, powered by Firefly, a proprietary generative AI model.

Helm.ai: $55 Million Secured For Developing AI Software For Autonomous Driving And Robotics Automation - Helm.ai, a provider of AI software for autonomous driving and robotics automation, has raised $55 million in its Series C financing round, bringing its total funding to $102 million, which will be used for R&D and commercial engagements in the automotive and robotics sectors.

OpenAI bought the makers of a Minecraft clone - OpenAI acquires Global Illumination, a New York-based digital product company, to work on its core products, including ChatGPT, without disclosing the terms of the acquisition or specifying the team's role.

Meet Lilli, our generative AI tool that’s a researcher, a time saver, and an inspiration - McKinsey has launched "Lilli," a generative AI solution that aggregates the firm's knowledge and capabilities to provide streamlined and efficient insights to clients, saving time and improving productivity.

ElevenLabs Comes Out of Beta and Releases Eleven Multilingual v2 - a Foundational AI Speech Model for Nearly 30 Languages - Voice AI platform ElevenLabs has released a new multilingual voice generation model, Eleven Multilingual v2, which supports speech in nearly 30 languages, allowing media companies, game developers, publishers, and independent creators to improve the accessibility of their content.

ChatGPT Killed the Old AI. Now Everyone Is Rushing to Build a New One - AI companies are facing a global GPU shortage and algorithmic constraints, forcing them to slow down their plans and search for new breakthroughs in order to train next-generation AI models.

South Korea's Naver launches generative AI services - South Korean internet giant Naver has launched its own generative AI tool, HyperCLOVA X, which includes a chatbot application and a generative AI function, aiming to provide generative AI-driven searches and customised services for enterprise clients.

Microsoft may bring AI capabilities to apps like Paint and Photos on Windows 11 - Microsoft is experimenting with bringing AI capabilities to apps like Paint and Photos on Windows 11, including features like object recognition, OCR, and generative AI.

‘Big Brother,’ ‘Survivor’ Producer Banijay Launches Fund to Develop TV Shows Using AI - Banijay, the superindie behind shows like Big Brother and MasterChef, has launched a creative fund to develop television formats focused on artificial intelligence, aiming to create fresh and forward-thinking formats in the technology space.

Google plans to bring AI-fueled security enhancements to Google Workspace - Google plans to bring AI-fueled security enhancements to Google Workspace, including automating certain tasks, enhancing the zero trust model with data loss prevention, adding context-aware controls in Drive, using AI to detect data breaches and suspicious actions, and offering client-side encryption on mobile versions of Gmail, Calendar, and Meet.

Twilio expands CustomerAI capabilities with generative and predictive AI - Twilio is expanding its customer artificial intelligence (AI) tools, called CustomerAI, by adding voice intelligence, predictive analytics, generative journeys, and AI Nutrition Facts Labels, as well as improving profile organization and sharing through a partnership with Databricks.

Wand.app raises $4.2M to scale its AI-powered creative tool for artists - Wand.app, an AI-powered creative tool for artists, has raised $4.2 million in seed funding to scale its platform, which allows artists to manually customize and visualize their ideas without compromising their specific visions.

Lex raises $2.75M for its AI writing tool that helps writers get past blocks - Lex, an AI-powered writing tool, has raised $2.75 million in funding to develop its platform that uses AI to help writers overcome blocks and improve their workflow, offering features such as formatting tools, markdown-based shortcuts, and AI-generated suggestions for when writers get stuck.

Nvidia’s Q2 earnings prove it’s the big winner in the generative AI boom - Nvidia's Q2 earnings demonstrate the company's success in selling AI chips for generative AI applications, with revenue doubling from last year and the data center unit overshadowing the gaming unit.

Hugging Face raises $235M from investors including Salesforce and Nvidia - AI startup Hugging Face has raised $235 million in a Series D funding round, bringing its valuation to $4.5 billion, and plans to expand its platform for creating, testing, and deploying machine learning models.

Modular secures $100M to build tools to optimize and create AI models - Modular, a startup creating a platform for developing and optimizing AI systems, has raised $100 million in a funding round to support product expansion, hardware support, and the expansion of its programming language, Mojo.

OpenAI partners with Scale AI to allow companies to fine-tune GPT-3.5 - OpenAI has partnered with Scale AI to allow companies to fine-tune its GPT-3.5 text-generating model using custom data, providing additional services and expertise beyond what OpenAI offers.

a16z-backed AI video generator Irreverent Labs raises funding from Samsung Next  - Bellevue-based startup Irreverent Labs, backed by a16z and Samsung Next, has raised funding for its AI technology that allows users to create videos using various inputs, from images to text to audio.

Ikigai lands $25M investment to bring generative AI to tabular data - Ikigai Labs, a no-code platform for AI, has raised $25 million in Series A funding to bring generative AI to tabular data and empower enterprises with AI.

DP Technology Raises 100 million US dollars, Achieves AI Milestones in Science - DP Technology has raised $100 million in a recent round of financing, with investors expressing positive expectations for the company's AI for Science research paradigm and its potential impact on various industries.

CEOs extol benefits of AI on earnings calls but not in official filings - CEOs are praising the benefits of AI on earnings calls, but not including it in official filings.

Concerns

The human costs of the AI boom - Cloudwork platforms, which are increasingly used by AI companies to outsource manual work, are failing to meet basic workers' rights and protections, with none of the platforms evaluated meeting more than half of the standards of decent work, according to a study by the University of Oxford.

Amazon AI scammers blew millions on Lake Como wedding and cars, FTC alleges - The FTC alleges that two brothers scammed consumers out of millions of dollars by posing as e-commerce experts, promising success on platforms like Amazon and Walmart, and using the money for luxury purchases and a lavish wedding.

Beware the Emergence of Shadow AI - The emergence of Shadow AI, AI systems used or developed within organizations without explicit approval or oversight, poses cybersecurity risks, infrastructure integration challenges, and ethical concerns, highlighting the need for organizations to implement concrete measures to address this issue.

‘Very wonderful, very toxic’: how AI became the culture war’s new frontier - The rightwing backlash against "woke AI" and its alleged political bias is shaping the conversation around AI development and regulation, despite experts arguing that AI models are simply remixing and regurgitating existing content.

AI unlikely to destroy most jobs, but clerical workers at risk, ILO says - Generative AI is unlikely to completely replace most jobs, but it will automate certain tasks, particularly in clerical work, potentially impacting female employment more significantly.

Robbing Delivery Robots Is Now a Thing - Instances of theft and vandalism of delivery robots used by restaurants and other businesses have been noted in the past couple of years, despite security measures such as cameras and sirens.

Tricks for making AI chatbots break rules are freely available online - Prompts that encourage AI chatbots to ignore pre-coded rules have been shared online for over 100 days, potentially enabling criminal activity.

Analysis

Early days of AI - The article discusses the early days of AI, the advancements in machine learning models, and the potential impact of AI on various industries in the future.

How teachers and students feel about A.I. - Teachers and students share their experiences using AI chatbots for teaching, learning, and studying, highlighting the benefits of personalized tutoring, simplified understanding of difficult concepts, and concerns about potential laziness and the joy of independent learning.

A Creator (Me) Made a Masterpiece With A.I. - The article discusses the controversies surrounding intellectual property and AI-generated content, suggesting that computers may never be able to produce truly groundbreaking creative work and that AI could ultimately be more of a boon than a competitor to artists and creative industries.

Visualising AI - Diversifying perspectives of AI through more diverse and accessible representations is crucial in expanding the public's understanding and vision of what AI can look like and achieve.

Policy

Bletchley Park to host AI safety talks in November - World leaders and AI experts will gather at Bletchley Park in November to discuss the safe use of artificial intelligence, with some arguing for a UN-style regulator to establish global rules.

UK to spend £100m in global race to produce AI chips - The UK government plans to spend £100m to establish a national AI resource by ordering key components from major chipmakers, but concerns are rising that the investment is too low compared to other countries, leaving the UK exposed in the global race to produce AI chips.

Meta confirms AI ‘off-switch’ incoming to Facebook, Instagram in Europe - Meta has confirmed that it will introduce non-personalized content feeds on Facebook and Instagram in the European Union to comply with the bloc's Digital Services Act, which requires larger platforms to provide users with the ability to switch off AI-driven personalization.

Beijing to restrict use of generative AI in online healthcare activities - Beijing's new rules aim to restrict the use of generative AI in online healthcare activities, reflecting the disruptive impact of the technology on traditional industries in China.

As Fight Over A.I. Artwork Unfolds, Judge Rejects Copyright Claim - A federal judge rejected an attempt to copyright an artwork generated by artificial intelligence, highlighting the ongoing legal battle over authorship and intellectual property in the age of AI.

Explainers

How to help high schoolers prepare for the rise of artificial intelligence - The MIT Abdul Latif Jameel Clinic for Machine Learning in Health (Jameel Clinic) organized a summer program to educate high school students about the use of AI in healthcare, aiming to reduce financial barriers and reach students from diverse backgrounds.

Fun

A Fun Film About a Topic That Scares Me a Lot - A filmmaker's work embraces spontaneity, randomness, and error, but the author's fear of attempting a backflip prevented them from learning the trick.

---

Copyright © 2023 Skynet Today, All rights reserved.