AI Weekly: 04/10/23

Good morning and welcome to this week’s edition of AI Weekly! While this week was relatively subdued compared to other weeks, it feels like it’s simply the calm before another wave of the storm.

Since leading AI researchers publicly called for a halt on AI development in order for the larger risks to be evaluated, other prominent individuals in tech refuted the idea. Although these parties acknowledge the need for a better understanding of the challenges associated with such fast-paced AI development, their concerns primarily exist around what a temporary pause would mean for China’s ability to continue gaining steam from a competitive standpoint.

As that complicated situation continues to unfold, enjoy reading about the other important AI happenings of this past week!

- ZG

Here are the most important stories of the week:

TEXT

Microsoft confirms more ads coming to Bing's AI-powered chatbot. Link.

• Director of Communications at Microsoft Caitlin Roulston says ads will appear in chat and traditional search results.

• The company has reportedly been testing ads in Bing since February.

• Thus far, ads have appeared as sponsored links and inside shopping results.

• The volume of ads and how they will be displayed remains uncertain.

• Microsoft plans to share ad revenue with partners whose content contributes to chat responses.

IMAGE/VIDEO

Clearview AI's controversial facial recognition database contains 30 billion photos scraped from social media without user permission. Link.

• The company claims the technology helps identify criminals, save abused children, and exonerate wrongfully accused individuals.

• Critics argue that privacy violations and faulty identifications made by facial recognition can lead to wrongful arrests.

• Clearview AI CEO, Hoan Ton-That, claims the database is lawfully collected and only accessible to law enforcement for after-the-crime investigations.

• Major social media platforms have taken measures to combat unauthorized scraping and protect user data, though scraped photos are difficult to remove.

• Lack of transparency and regulation around law enforcement's use of facial recognition technology raises concerns about civil liberties and privacy.

SPEECH/AUDIO

Music app Moises, powered by artificial intelligence, has surpassed 30 million registered users and raised $10.25 million in funding from investors. Link.

• The Salt Lake City-based startup was founded in 2019 and has ranked No. 1 in Apple's app store and is available in 33 languages.

• Moises uses AI to offer artists tools to customize music or practice, such as separating or removing vocals and instruments, detecting a song's key, changing its pitch, and identifying chords.

• The app also introduced an AI-powered songwriting assistant, which the CEO describes as "like Grammarly for musicians."

• Moises aims to lead a new generation of AI technologies and become a service provider for the music industry.

• The company has attracted investors from multiple continents and claims to be one of few companies with the infrastructure to run large AI models that render results quickly.

Adthos has launched a platform that uses AI to generate scripts for audio ads and even add voiceovers, sound effects, and music, allowing advertisers to produce ads efficiently and instantly. Link.

• The platform also provides a browser-based, multi-track editor that enables customers to go in and edit the ads, creating a "fully produced" ad that can be played out via integrations with radio automation systems and ad servers.

• Adthos's platform is a mix of in-house tech and third-party APIs, with the company leveraging OpenAI's recently released GPT-4 text-producing model to generate ad scripts.

• Adthos also compiled a library of hundreds of synthetic voices, including "Emmy Award-winning" voice actors and talents, which raises ethical questions about whether all of those voice actors chose to be included in Adthos' library in the first place.

• Adthos CEO Raoul Wedel says that the voice actors are "fully aware" of the platform and paid either a flat fee or a licensing fee for the use of their voice, based on their preferences, and that the actors have signed consent and licensing agreements.

• Adthos claims to have several paying customers, including Dutch media conglomerate Talpa, and advertisers and agencies like GroupM, Dentsu, and Sportradar.

CODE/DEVTOOLS

LangChain recently raised $10 million in seed funding led by Benchmark to empower developers in building applications powered by language models. Link.

• LangChain, released six months ago, provides a framework for developing LLM applications, integrating with various components and data sources.

• Initially an open-source side project, it has gained significant community support with over 20K GitHub stars and 350+ contributors.

• The framework includes modular components, pre-built chains, and agents, enabling developers to create custom solutions or use out-of-the-box implementations.

• Future plans involve expanding TypeScript package capabilities, implementing OutputParsers and Retriever abstraction, and integrating with other solutions for enhanced observability and experimentation.

• LangChain remains open-source and developer-focused, actively hiring founding engineers and seeking community feedback for improvement.

Replit has announced the Chroma template, which allows developers to create AI applications with state and memory easily. Link.

• The Chroma template adds capabilities to large language models (LLMs) that allow for question answering bots, personal assistant agents, and applications that interact with other APIs.

• The template uses embeddings, numerical vector representations of all kinds of data, generated using an embedding model, to store and interact with data in an AI native way.

• Chroma is the easiest embeddings store for AI applications, handling embedding documents and queries and storing documents alongside their embeddings, with support for filtering and dynamic updating.

• The Chroma template includes a demo for question-answering using OpenAI's ChatGPT API, embedded with the textbook "Artificial Intelligence: A Modern Approach" by Stuart Russell and Peter Norvig.

• Replit and Chroma make it easy for developers to experiment with, build, and share powerful AI-driven applications.

MULTIMODAL

Researchers have developed CHAMPAGNE, a multimodal dialog model that uses visual and verbal context from a scene to generate dialogue that feels intuitive. Link.

• CHAMPAGNE was built using a large-scale dataset called YTD-18M, constructed from 20M YouTube videos.

• The model's ability to understand contextual cues in conversation highlights the potential for AI systems to observe and understand organic forms like humans.

• In evaluations, CHAMPAGNE models performed better on a range of open-domain text conversations and understanding social interactions.

• The development of such models suggests the potential for AI systems to read faces as well as text, predict human behavior, and personalize interactions to satisfy users.

• Future AI systems that understand contextual cues can create more intuitive dialogue and personalized interactions by modeling human behavior.

Cornell University researcher Ruidong Zhang has developed sonar glasses that can interpret silent speech to control devices. Link.

• The glasses use tiny microphones and speakers to read the words people mouth as they silently command it to do things like pause or skip a music track or enter a passcode without touching their phone.

• The system builds off a similar project the team created using a wireless earbud, which relied on cameras.

• The glasses form factor removes the need to face a camera or put something in your ear.

• The system requires only a few minutes of training data to learn a user’s speech patterns, then uses a deep learning algorithm to analyze echo profiles in real time with about 95% accuracy.

• The glasses have potential for hands-free control in quiet environments like libraries and for people with speech disabilities to feed dialogue into a voice synthesizer to speak the words aloud.

HEALTHCARE

Researchers are employing AI and wireless tech to discreetly observe elderly individuals in their living spaces. Link.

• The system detects emerging health issues without necessitating wearable devices, providing early detection.

• Low-power millimeter-wave radio systems are compatible with homes, hospitals, and long-term care facilities.

• The technology sends alerts to healthcare professionals regarding falls and changes in health conditions.

• This innovation aids overburdened healthcare systems in managing the growing elderly population.

• Researchers have teamed up with Gold Sentintel to commercialize the technology, with installations in several long-term care homes.

POLICY/LAW/ETHICS

CAIDP files a complaint against OpenAI with the FTC, claiming consumer protection rule violations. Link.

• The complaint follows an open letter signed by AI researchers, Elon Musk, and CAIDP President Marc Rotenberg.

• CAIDP highlights risks from GPT-4, such as malicious code, propaganda, and biased training data.

• The complaint alleges OpenAI released GPT-4 for commercial use despite knowing the risks.

• CAIDP requests a halt on further GPT model deployment and independent assessments before future rollouts.

• An FTC investigation into OpenAI would mark a significant escalation in regulating AI tools.

Metaphysic CEO Tom Graham submitted his AI likeness for copyright registration with the U.S. Copyright Office, making him the first person to do so. Link.

• The company creates authorized “deep fakes” of people using AI and is championing individual ownership and control of their AI likenesses and biometric data.

• The registration of Graham’s AI likeness with the U.S. Copyright Office provides a framework for how other individuals and public figures can take steps to protect their identities, performances, and brands.

• The company’s technology creates content that looks and feels real, and it champions the idea that individuals should never lose control over their identity, privacy, or biometric data.

• Graham recorded a three-minute video of himself on a mobile phone to capture his likeness, voice, and biometric data, which the company utilized to create an AI avatar of present-day Graham.

• The company hopes that its actions advance the evolving discussion surrounding privacy and individual rights in the context of rapidly advancing generative AI technologies that are becoming increasingly realistic and indistinguishable from reality.

India will not regulate the growth of AI within the country, identifying the sector as significant and strategic for the nation. Link.

• The Ministry of Electronics and IT has assessed the ethical concerns and risks of bias and discrimination associated with AI and is implementing policies and infrastructure measures to cultivate a robust AI sector.

• The expansion of AI will have a "kinetic effect" on entrepreneurship and business development in India.

• Despite the burgeoning interest in AI in many developed markets, India has seen a limited number of startups entering the field.

• India's decision comes as tech entrepreneurs and academics increasingly call for government intervention in AI.

• Italy recently imposed a temporary ban on OpenAI's ChatGPT, citing concerns that it violated the European Union's General Data Protection Regulation.

OTHER

Covariant raises $75 million in a Series C extension led by existing investors Radical Ventures and Index Ventures, bringing the total funding to $222 million. Link.

• CEO Peter Chen aims to develop the Covariant Brain as a more capable foundation model and apply it to various sectors.

• Covariant's technology, the Covariant Brain, builds a massive database of potential package sizes, shapes, and materials for logistics.

• The company's founders have a direct connection to OpenAI, and Chen draws parallels between generative AI/ChatGPT and Covariant's approach.

• The funds will be used to expand the deployment of Covariant's system to retailers and logistics firms.

Fourthline, an Amsterdam-based startup, raises €50 million ($54 million) to expand its AI-based identity verification and anti-money laundering solutions. Link.

• Finch Capital leads the all-equity round, with other undisclosed investors participating.

• The company has grown 80% annually since its launch in 2018, with clients including N26, Western Union, and Wish.

• Fourthline's technology conducts around 200 checks, including ID document examination, biometric data parsing, and verifying names against sanction lists.

• The company claims its approach can identify 60% more fraud with a 99.98% accuracy rate.

• Fourthline is focused exclusively on KYC, AML, and ID verification services for the financial sector in Europe.

Narrato, an AI content creation and collaboration platform, raises a $1 million pre-seed round led by AirTree Ventures. Link.

• OfBusiness and Shreesha Ramdas also participated in the round.

• Founded in January 2022 by Sophia Solanki, who previously created content marketing platform DrumUp.

• Narrato's main feature is an AI content assistant for planning, creation, optimization, collaboration, and publishing.

• The platform offers templates for various content types and plans to expand its generative AI-assisted content templates.

• Narrato aims to differentiate itself by embedding generative AI into the entire marketing and content creation workflow.

London-based Quantexa raises $129 million in Series E funding led by Singapore’s sovereign wealth fund GIC, valuing the startup at $1.8 billion. Link.

• Quantexa provides AI tools for risk and compliance, identity verification, and financial investigations to major banks and organizations.

• The company has doubled its business in the last 18 months and serves hundreds of customers in 70 countries.

• Quantexa plans to use the funding to develop new technology and make acquisitions.

• CEO Vishal Marria projects $100 million in revenues next year and profitability by 2025.

Glean announces new AI-based features to improve accuracy, security, and collaboration in enterprise data and content. Link.

• Features include AI answers, expert detection, and in-context recommendations.

• Glean's governance engine enforces real-time enterprise data permissions and governance rules.

• The company was founded in 2020 and has a $1 billion valuation after a $100 million Series C funding round.

• Glean serves over 70 customers across various industries, including technology, media, education, and healthcare.

• McKinsey report shows AI can help reduce knowledge workers' time spent searching for information by 35% and increase revenue by 6%.

Machine learning algorithms can be trained to detect cyberbullying patterns in real-time, paving the way for prompt intervention and prevention measures. Link.

• Social media and online communication have led to an increase in cyberbullying.

• NLP algorithms can analyze online communication language to determine the tone and identify specific terms or phrases associated with bullying behavior.

• Machine learning offers scalability advantages compared to conventional ways of preventing cyberbullying, which can be inefficient and time-consuming.

• Significant quantities of high-quality data are required to train the algorithms, which is a challenging aspect.

• Dr Manjeevan Singh and his team collected and labeled tweets in Malay to train deep learning models, achieving an 80% accuracy rate for cyberbullying classification, which can be further improved with additional training and the incorporation of hate speech data.

SetSail aims to compensate salespeople based on their position in the sales pipeline using data analysis to determine fair payments. Link.

• The company built a data infrastructure to provide a 360-degree view of the sales process, which now allows for querying the system in natural language through ChatGPT.

• ChatGPT can write code, including well-structured SQL queries, to answer specific questions and replace dashboards with interactive discussion on top of the data.

• The data infrastructure has value beyond its original purpose of building sales incentives and can answer questions such as which deals are close to closing or which salespeople are having trouble closing deals.

• SetSail recently announced new capabilities and an $11 million investment, including from Zoom Ventures, which recently announced its own conversational tool aligned with SetSail's capabilities.

• Founded in 2018, SetSail has raised almost $50 million to date.

KPMG partner Jonathan Dambrot and two other co-founders launch Cranium, a software startup addressing the challenges around AI security and compliance. Link.

• Cranium attempts to map AI pipelines, monitor for outside threats and validate their security, and it can address threats such as poisoning and text-based attacks.

• Customers can establish an AI security framework with Cranium, providing their security and data science teams with a foundation for building a security program, and capture both in-development and deployed AI pipelines.

• Cranium wants to be a source of data to identify risks proactively across its client base and make the security and trust of AI a priority.

• Cranium has raised $7m in venture capital from KPMG and SYN Ventures.

• The company currently has around 30 full-time employees, but it expects to end the year with around 40 to 50 assuming business picks up.

GPT Tip of the Week

Prompt: What are some common pitfalls to avoid when fine-tuning a large language model?

Response: When fine-tuning a large language model, there are several common pitfalls to avoid:

1. Overfitting: Fine-tuning a model on a small dataset can lead to overfitting, where the model performs well on the training data but poorly on new data. To avoid this, use techniques such as data augmentation, regularization, and early stopping.

2. Catastrophic forgetting: When fine-tuning a model, it's possible to forget what it has learned previously. This is called catastrophic forgetting and can be avoided by using techniques such as distillation, multitask learning, and continual learning.

3. Biased training data: If the training data is biased, the fine-tuned model will also be biased. To avoid this, ensure that the training data is diverse and representative of the real-world scenarios where the model will be used.

4. Insufficient training data: Fine-tuning a large language model requires a large amount of data. Insufficient training data can result in poor performance. Ensure that the dataset used for fine-tuning is large enough to produce accurate results.

5. Poor hyperparameter tuning: Choosing the right hyperparameters for fine-tuning can significantly impact model performance. Experiment with different values and use techniques such as grid search and random search to find the optimal values.

6. Lack of interpretability: Fine-tuned models can be difficult to interpret, making it hard to understand how they arrived at their decisions. To improve interpretability, use techniques such as attention mechanisms and explainable AI.