AI Weekly: 08/21/23

Good morning and welcome to this week’s edition of AI Weekly! This past week was filled with some exciting updates in the ecosystem. A group of researchers from MIT who created a new neural network architecture that supposedly outperforms traditional transformer neural networks have started to gain significant attention.

Also, one of the co-authors on the famous “Attention Is All You Need” Google Brain paper, has launched his own startup in an attempt to overcome the inflexibility of current AI systems.

In recent funding news, Anthropic raised $100M from South Korean mobile carrier SK Telecom in order to co-develop a multilingual LLM customized for global telco firms.

One of the more unique and somewhat unprecedented stories of the week revolves around Index Ventures, who has partnered with Oracle to provide Nvidia GPUs to its portfolio companies to address the GPU shortage issue. It will be very interesting to see if this is the catalyst to other venture firms entering the current GPU shortage war.

Continue reading below for more noteworthy AI happenings from this past week!

- ZG

Here are the most important stories of the week:

TEXT

Anthropic has raised $100 million in funding from South Korean mobile carrier SK Telecom (SKT). Link.

• SKT's investment in Anthropic comes three months after Anthropic raised $450 million in its Series C funding round, in which SKTVC also participated.

• Anthropic and SKT plan to co-develop a multilingual LLM customized for global telco firms, with Anthropic's chief science officer, Jared Kaplan, leading the customization and product roadmap.

• SKT intends to use the customized LLM to provide various AI services to potential customers in the telco industry, including customer service, marketing, sales, and interactive consumer applications.

• Anthropic is building an AI system called Claude, similar to OpenAI's ChatGPT, which can perform tasks like searching, generating answers, automating workflows, coding, and processing text in natural conversations.

• The LLM developed by Anthropic and SKT will support multiple languages, including English, Korean, German, Japanese, Arabic, and Spanish. The investment is expected to close in the third quarter of this year.

Liquid neural networks, also known as liquid networks, are a type of neural network architecture that emphasizes adaptability and compactness. Link.

• Liquid networks were first introduced in research papers dating back to 2018, with a significant paper called "Liquid Time-constant Networks" published in 2020, gaining attention for their unique properties.

• These networks are designed to be adaptable even after training, enabling them to adjust to incoming inputs dynamically.

• The "liquid" aspect refers to the flexibility and adaptability of the networks, and they use fewer but richer nodes compared to traditional neural networks.

• Liquid networks have shown potential benefits in terms of running on less computing power, being more interpretable due to their smaller size, and addressing issues related to the black box nature of complex neural networks.

• They require time series data and are particularly suited for robotics applications and safety-critical systems where understanding and transparency are crucial.

McKinsey and Company, a global consulting firm, is launching its own gen AI tool called Lilli, developed by its ClienTech team. Link.

• Lilli is a chat application that offers information, insights, data, plans, and recommends internal experts for consulting projects based on over 100,000 documents and interview transcripts.

• The tool aims to serve as a repository of McKinsey's collective knowledge, answering questions and offering valuable insights.

• The AI tool has been in beta since June and will be rolled out across McKinsey this fall.

• Lilli operates with two tabs: "GenAI Chat" sources data from a generalized LLM backend, while "Client Capabilities" sources responses from McKinsey's internal documents.

• McKinsey intends for Lilli to be used throughout various consulting tasks, from initial research to drafting client plans, with a focus on quality information over rapid responses.

OpenAI enhances ChatGPT platform with multiple updates for improved user experience and engagement. Link.

• Example prompts at start of chats offer context and ease for new users.

• Suggested replies enable easy continuation of conversations and exploration of insights.

• Plus users now default to GPT-4 model for richer interactions.

• Code Interpreter beta allows Plus users to analyze data across multiple files.

• Extended login duration, improved interface, and keyboard shortcuts enhance overall usability.

IMAGE/VIDEO

Alias Technologies, led by former Machine Zone CEO Kristen Garcia Dumont, has launched BeFake, a social media app that allows users to express themselves through AI-generated visuals. Link.

• Built around the concept "Why be real when you can be fake?", BeFake aims to offer a unique, creative platform as an alternative to the conventional reality depicted on existing social media.

• By using text prompts, users can transform themselves into engaging, AI-augmented visuals, going beyond physical limitations to express their identity and creativity.

• Kristen Garcia Dumont believes BeFake is more than just an app; it is a revolutionary way of interacting on social media that allows for new forms of human connection. It aims to democratize social media by reducing the pressure of posting "real, candid" photos.

• The company is fully remote and has raised seed funding from Khosla Ventures, Next Coast Ventures, and Joe Lonsdale. BeFake will have a freemium model where power users can opt for a subscription to generate more images.

• Garcia Dumont's background in gaming, particularly with Machine Zone which was acquired for $600 million, informed her perspective on the role of AI and machine learning in building communities, stimulating user engagement, and monetization strategies.

An AI camera system installed by Cornwall police on a major road in England captured nearly 300 law-breaking drivers within 72 hours and 1,000 after 5 days. Link.

• Acusensus, an AI tech company, developed the system to identify offenses such as seatbelt violations and mobile phone usage while driving.

• The camera system, placed on the A30 road, uses high shutter speeds, infrared flash, and AI for reviewing images of passing vehicles.

• Detected violations are sent for human review, and drivers receive warning notices or intended prosecution based on offense severity.

• Similar AI surveillance systems have been used worldwide, including catching littering drivers in the UK and fare evaders in Barcelona and NYC.

• Adrian Leisk of Devon and Cornwall Police expressed disappointment in drivers not wearing seatbelts or using phones while driving and emphasized the technology's role in enforcement.

SPEECH/AUDIO

Voiceflow, a platform for creating conversational AI experiences, has raised $15 million in a funding round led by VC firm OpenView. Link.

• The funding brings Voiceflow's total raised to $35 million and values the startup at $105 million post-money.

• Voiceflow provides a collaborative design platform for building AI agents, allowing teams to deploy agents powered by LLMs, including OpenAI's GPT-4.

• The platform enables companies to design, test, and deploy AI agents without being locked into a specific model or technology, offering flexibility to mix and match different models for conversational AI experiences.

• Voiceflow has over 450 customers, including Amazon, JP Morgan, The Home Depot, State Farm, Vodafone, and government agencies.

• The funding will be used to enhance product innovation, including the addition of an AI builder powered by LLMs for customer support and more.

CODE/DATA/DEVTOOLS

The Allen Institute for AI (AI2) has released a new large text dataset called Dolma, which is open and free for the AI research community to use and inspect. Link.

• Dolma is intended to serve as the foundation for AI2's planned open language model (OLMo), and the dataset's openness aligns with the AI model's goal of being freely accessible for modification and research.

• Unlike some companies that guard their language model training data as proprietary, AI2 aims to make Dolma transparent and well-documented, revealing sources and processes used in its creation.

• Dolma is among the largest open datasets of its kind, containing 3 billion tokens, and is designed to be user-friendly with straightforward usage permissions.

• Users of Dolma are required to provide contact information, disclose any derivative creations, distribute derivatives under the same license, and agree not to use the dataset for prohibited purposes like surveillance or disinformation.

• Access to Dolma is available through Hugging Face, and AI2 encourages users to make removal requests for personal data if necessary.

Arthur, a machine learning monitoring startup, has developed Arthur Bench, an open-source tool aimed at helping users find the most suitable LLM for their specific dataset. Link.

• The company's CEO and co-founder, Adam Wenchel, notes that many companies lack an organized way to assess the effectiveness of different LLMs, leading to the creation of Arthur Bench.

• Arthur Bench provides a suite of tools to systematically test LLM performance, with a focus on evaluating how user prompts for a given application fare with different LLMs.

• The tool enables testing various prompts against different LLMs at scale, aiding users in making informed decisions about which model suits their use case best.

• Arthur Bench is being released as an open-source tool, with a SaaS version available for customers seeking simplified management or more extensive testing requirements.

• This launch follows the release of Arthur Shield in May, which serves as an LLM firewall to detect model hallucinations, guard against harmful information, and prevent private data leaks.

Swedish company Apica, which specializes in testing applications using synthetic data, has acquired Logiq.ai to expand into post-production monitoring. Link.

• The acquisition enables Apica to improve its synthetic datasets based on the application's behavior in production, enhancing the quality of the synthetic data over time.

• Logiq.ai's technology helps collect, organize, and enrich data from various sources, providing a consolidated view of data for analysis and visualization.

• The acquisition brings together Apica's testing capabilities with Logiq.ai's data consolidation and enrichment tools.

• Apica aims to provide customers with a comprehensive solution that combines testing, monitoring, and data analysis.

• The acquisition was accompanied by a $10 million investment and is expected to result in new functionality being integrated into Apica's offerings in the third quarter of the year.

HARDWARE/ROBOTICS

Cruise, GM's self-driving car subsidiary, has been asked by the California DMV to cut its robotaxi fleet in San Francisco by 50% following a crash with a fire truck. Link.

• The DMV, which regulates autonomous vehicle testing and deployment in the state, is investigating concerning incidents involving Cruise vehicles and has ordered the reduction in operations.

• Cruise is to limit its fleet to a maximum of 50 driverless vehicles during the day and 150 at night until the investigation is concluded.

• Cruise is complying with the request and has issued a blog post explaining its perspective on the crash, emphasizing its focus on road safety.

• This comes after Cruise has faced other issues, including reports of its driverless cars stalling and causing traffic disruptions.

• Despite these challenges, Cruise recently received approval from the California Public Utilities Commission (CPUC) to expand its commercial operations and charge for rides in San Francisco.

GPUs

Index Ventures, a global investor, has partnered with Oracle to provide GPUs (Nvidia’s H100 and A100 chips) to its portfolio companies to address the GPU shortage issue. Link.

• Founders of AI startups are struggling to secure the necessary GPUs for their projects due to high demand from larger companies and nations.

• The scarcity of GPUs is not just about cost but the allocation of over 95% of GPU capacity to larger players due to pre-commitments with cloud vendors.

• Index Ventures aims to remove the barrier of GPU access for early-stage portfolio companies, allowing them to focus on their workloads and accelerate development.

• The partnership involves Index making a pre-commitment and paying the cloud bill, while Oracle manages the cluster and the startups receive access to GPUs for free.

• Index's partnership with Oracle aims to level the playing field for AI startups and enable them to graduate to their own GPU clusters eventually.

POLICY/LAW/ETHICS

OpenAI might need to reset ChatGPT due to a potential lawsuit and could face fines up to $150,000 per copyrighted content piece used in training. Link.

• ChatGPT has gained attention for its various uses, but its success might have legal consequences.

• Language models like GPT-3.5 and GPT-4 that power ChatGPT are trained using third-party data, including copyrighted material without permission.

• The New York Times is considering suing OpenAI after changing its terms to forbid using its content for training language models.

• The lawsuit could result in substantial financial penalties and even force OpenAI to restart the training of its language model, damaging prior efforts.

• The situation highlights concerns about the reliance of language models on others' work and could lead to more legal actions from other companies.

United States District Court Judge Beryl A. Howell ruled that AI-generated artwork cannot be copyrighted, based on a lawsuit against the US Copyright Office by Stephen Thaler, the creator of an AI-generated image made with the Creativity Machine algorithm. Link.

• Thaler attempted to secure copyright multiple times for the AI-generated image but was denied, leading him to sue the Office.

• Judge Howell's decision emphasized that copyright requires human authorship and cannot be granted to works created "absent any guiding human hand."

• Past cases, including the monkey selfie case, were cited by the judge to support her ruling.

• The judge acknowledged the evolving nature of copyright in relation to AI and questioned how much human input is necessary for AI-created art to be copyrighted, given that AI models are often trained on existing works.

• Stephen Thaler plans to appeal the ruling; his attorney disagrees with the court's interpretation, and the US Copyright Office supports the decision. Ongoing lawsuits involving AI and copyright law continue to accumulate.

OpenAI claims to have developed a content moderation technique using its GPT-4 model to reduce the burden on human moderation teams. Link.

• The approach involves prompting GPT-4 with a policy for moderation and creating a test set of content examples that violate or adhere to the policy.

• Policy experts label the examples and assess GPT-4's judgments, refining the policy based on discrepancies and reasoning provided by the model.

• OpenAI asserts that its process can accelerate the rollout of new content moderation policies, but skepticism remains.

• Existing AI-powered moderation tools, like Google's Perspective and various startup solutions, have faced challenges with biases and imperfections.

• OpenAI acknowledges that GPT-4's judgments are susceptible to biases and emphasizes the need for careful monitoring and validation by humans.

OTHER

The 2023 Gartner Hype Cycle places generative AI on the 'Peak of Inflated Expectations' for the first time. Link.

• Generative AI encompasses various technologies, from foundation and diffusion models to prompt engineering tools, leading to inflated expectations in the market.

• Many vendors claim to offer generative AI, causing confusion among enterprise executives about customization, security, and privacy practices.

• The market sees an overwhelming number of products with generative AI claims, and there is confusion regarding the actual benefits of the technology.

• Challenges in de-risking and deploying generative AI models contribute to the placement of generative AI on the 'Peak of Inflated Expectations.'

• The next stage in the Gartner Hype Cycle is the 'Trough of Disillusionment,' where competitive dynamics will play a role in shaping the technology's future trajectory.

Llion Jones, co-author of the transformative 2017 research paper on transformer architecture, “Attention Is All You Need,” left Google to co-found Sakana AI, based in Tokyo. Link.

• Sakana AI's approach is based on biomimicry, drawing inspiration from collective intelligence in nature (e.g., fish schools) to design flexible, reactive, and efficient AI models.

• The company develops numerous smaller AI models that collaborate like a swarm to deliver complex results, contrasting with the trend of creating extensive AI systems.

• Sakana AI's vision aims to overcome the inflexibility of current AI systems by embracing evolutionary computing principles inspired by nature.

• By utilizing swarm concepts and evolutionary computing, Sakana AI seeks to provide cost-effective and adaptive AI solutions while contributing to Tokyo's role as an AI innovation hub.

OpenAI has acquired Global Illumination, Inc., a NYC-based startup founded by former Facebook and Instagram employees, known for their work on creative tools, infrastructure, and digital experiences. Link.

• The Global Illumination team has a strong background, having made contributions at notable companies like YouTube, Google, Pixar, and Riot Games.

• The acquisition is seen as an excellent "acqui-hire," meaning OpenAI primarily acquired the startup to bring its talented team on board.

• OpenAI plans to have the newly acquired team work on core products, including ChatGPT, signaling possible advancements or new features for the language model.

• Given Global Illumination’s focus on visual-facing products and services, the acquisition suggests OpenAI may be looking to expand into multimedia features for ChatGPT, Dall-E 2, or even possibly launch a video-generation product.

• Financial terms of the deal were not disclosed, but the acquisition could significantly enhance OpenAI's capabilities in creative tools and possibly in realistic visual effects given Global Illumination's past work in those areas.

Many large companies are facing challenges in deploying generative AI technology, leading to cost and data management hurdles that are causing generative AI projects to remain in the pilot phase. Link.

• Deloitte and NVIDIA are establishing an "Ambassador AI program" to assist companies struggling with full-scale deployment of AI projects.

• More than half of AI decision-makers in top companies are encountering cost barriers to deploying advanced AI tools, according to S&P Global's 2023 Global Trends in AI report.

• Around 70% of surveyed companies have at least one AI project in production, but half of them are still in the pilot or proof-of-concept stage, with data management, security, and computing resources cited as top challenges.

• Outdated data infrastructure is negatively impacting the ability to achieve enterprise-scale AI deployment and sustainable AI use.

• Concerns about AI's potential to amplify biases and the inability of AI models to explain decisions remain, even as companies aim to leverage AI for productivity and insights.