AI Weekly: 03/06/23

Good morning and welcome to another edition of AI Weekly! Recently, we’ve been seeing platforms that are making it super straightforward to build and deploy your own AI models.

SambaNova Systems has built a platform to help enterprises and government organizations create their own proprietary generative AI models. Replicate allows you to package your model in a production-ready container so that people can run your model in a few lines of code. Agent deployment platforms like Fixie and Langchain allow you to plug into your personal databases to create personally tailored ChatGPTs for your organization. At a certain point, and we may already be there, the model layer will be commoditized. Once that happens, where does the value accrue?

One place to look is towards platforms that help make the building, testing, and evaluation processes of generative AI models as streamlined as possible. While there are numerous companies currently claiming that they can help you structure your unstructured data (i.e. Galileo, ScaleAI, Cleanlab, etc.), I am yet to see a company that supports the end-to-end deployment process.

We may end up in a world where every individual has their own personal model. Now, of course, most people will never have the technical ability to build their own model from scratch, but they may be able to leverage a user-friendly platform that ingests their data (documents, emails, texts, etc.), automatically fine-tunes it, and tells them the strengths and weaknesses of their model given its training data and amount of compute being used.

Some factors that the platform may evaluate are:

• Accuracy: Automatically runs cross-validation, hold-out testing, or A/B testing

• Speed: Measures how quickly the platform can collect, clean, preprocess, and analyze data. Should also recommend ways to increase speed

• Robustness: Automatically recognizes and creates adverse conditions such as noisy data, missing data, or adversarial attacks and measures how well a machine learning model performs under those adverse conditions

• Flexibility: Measures how well a platform can adapt to changing data and modeling requirements. A flexible platform can accommodate different types of data, modeling techniques, and business requirements.

• Cost: Predicts the cost of the model currently and at scale by measuring the cost of training, support services, etc.

Given their specific needs, users of this platform may be able to upgrade their model in different areas so that it is optimized for their specific needs and use cases.

While this is all still theoretical, it’s exciting to think about the next evolution of the AI tech stack. Until we get there, enjoy reading about the latest developments in AI!

- ZG

Here are the most important stories of the week:

TEXT

Meta has announced a new AI-powered large language model (LLM) called LLaMA-13B that can reportedly outperform OpenAI's GPT-3 model despite being "10x smaller." Link.

• The LLaMA family of language models range from 7 billion to 65 billion parameters in size, with the top model (LLaMA-65B) comparable to similar offerings from DeepMind, Google, and OpenAI.

• Meta trained its LLaMA models using publicly available datasets, potentially allowing for open sourcing and reproducibility.

• LLaMA is intended to form the basis of future, more-refined AI models, and could be useful in natural language research and power applications such as question answering and natural language understanding.

• LLaMA-13B can reportedly outperform GPT-3 while running on a single GPU when measured across eight standard "common sense reasoning" benchmarks.

• Smaller-sized AI models like LLaMA could potentially enable running language assistants locally on devices such as PCs and smartphones.

Cohere Summarize is a new endpoint for text summarization, powered by the company's foundational language model. Link.

• The Summarize beta empowers users to efficiently extract crucial information from lengthy documents and articles, greatly increasing productivity and saving valuable time.

• The model powering this endpoint is customized specifically for summarization, ensuring factual accuracy.

• Users can summarize lengthy documents up to 50,000 characters, which is equivalent to 18-20 pages in a single-spaced format.

• Cohere Summarize offers customizable settings that allow users to control certain aspects of their summaries, such as the length of the summary and output formats.

• Accessing the Cohere Summarize API is easy and user-friendly, with a Summarize playground and detailed documentation available.

Anthropic, an AI startup founded by ex-OpenAI employees, is offering access to its AI text-generating models to commercial partners. Link.

• Robin AI, a legal tech startup, is the first commercial venture to integrate Anthropic models into its software.

• Robin AI has fine-tuned an Anthropic model on a dataset of legal text to draft and negotiate contracts.

• Anthropic recently launched a closed beta for an AI system called Claude, similar to OpenAI's ChatGPT, which uses a technique called "constitutional AI" to align AI systems with human intentions.

• Anthropic has not disclosed how many partners it's working with or how many models it plans to open up to commercial usage.

• Anthropic has received hundreds of millions of dollars in investments, including a recent $300 million pledge from Google for a 10% stake in the startup.

Brave Search has launched a new "Summarizer" feature powered by its own LLMs. Link.

• The Summarizer provides a synopsis of a search query using different sources, and is available to all Brave Search users on desktop and mobile.

• The feature is reliant on a mix of three models: question-answering, classifier, and sentence rewriter.

• Brave Search also highlights relevant sentences in listed results as news articles.

• The Summarizer generates plain-written summaries at the top of the search results page, aggregating the latest sources on the Web and providing source attribution for transparency and accountability.

• Brave Search warns users that they should not believe everything produced by AI-powered search results and that the Summarizer feature won't generate a summary for all queries.

Hundreds of books written by ChatGPT have appeared on Amazon in recent weeks, ranging from fantasy fiction to self-help and non-fiction. Link.

• Some authors have used ChatGPT alongside other generative AI software to create illustrated children’s books entirely produced by artificial intelligence through simple text prompts.

• Due to the nature of ChatGPT and many authors’ failure to disclose they have used it, it is nearly impossible to get a full accounting of how many e-books may be written by AI.

• ChatGPT learns how to write by scanning millions of pages of existing text, which raises concerns over authenticity and plagiarism.

• There are already concerns being raised over authenticity and low-quality books, which could turn book writing from a craft into a commodity.

• Amazon did not address whether it had plans to change or review its Kindle store policies around authors’ use of AI or other automated writing tools.

Robin AI, an AI-driven platform that helps people negotiate and edit contracts, has raised $10.5m in a funding round led by Plural, with participation from Episode 1 and angel investors. Link.

• The company's machine learning model is trained on proprietary data from 4.5 million legal documents, allowing users to draft and negotiate contracts 60-80% faster, saving up to 75% on legal fees.

• Robin AI's SaaS + services business model combines a 'lawyer-in-the-loop' with machine learning technology and integrates the latest models from Anthropic.

• The company has grown revenue 20x since its last fundraising and is already cash flow positive, reviewing over 12,000 contracts per year.

• This latest round will drive further development of its tech team and enable the partial acquisition of competitor LawGeex.

• Robin AI's existing clients include law firm Clifford Chance and impact fund Blue Earth Capital, and it is adding a cohort of customers from LawGeex, including UBS, PWC, Ebay, and Pepsi.

Shortwave, an email app launched by Google executives, has introduced an AI-powered summary feature to allow users to quickly grasp the content of long emails or threads. Link.

• The summary feature, powered by OpenAI's GPT-3, is free for all platforms and works well in translating emails from other languages. It has been tested on different use cases, from newsletters to large numbers of emails.

• Users can include a summary of a previous message when forwarding an email, allowing others to quickly understand the context of the conversation even if they don't use Shortwave.

• Shortwave plans to introduce more AI-powered features in the future, such as additional summarization methods and smart composing, as well as a semantic search function.

• Shortwave offers both free and paid subscription plans for users, with basic features free and additional AI-powered functions available for a fee.

IMAGE/VIDEO

YouTube is incorporating new generative AI features into its video-sharing platform to keep up with AI rivals such as OpenAI and Microsoft. Link.

• YouTube creators will soon be able to swap outfits or create fantastical film settings through AI's generative capabilities in the coming months.

• Google is under pressure to release generative AI products that can generate detailed text answers or create brand-new images based on information they’ve ingested from the internet and other digital sources.

• Some critics have said Google has been too slow to roll out its own tools and services that are ready for public consumption, while rivals have released popular products like OpenAI's ChatGPT and Microsoft's Bing chatbot.

• Google announced its own ChatGPT competitor, called Bard, in February.

• YouTube's head, Neal Mohan, signaled that Google will develop its generative AI features with thoughtful guardrails and protections to embrace the technology responsibly.

Snapchat is introducing a chatbot named "My AI," powered by OpenAI's latest version of ChatGPT. Link.

• The goal is for My AI to eventually be available to all of Snapchat's 750 million monthly users.

• My AI adheres to Snapchat's trust and safety guidelines and has limited functionality compared to ChatGPT.

• Snap treats generative AI as a persona, with My AI's profile designed to look like any other Snapchat user's profile.

• My AI could boost Snapchat's paid subscriber numbers in the short term and open up new revenue streams in the future.

• Snap plans to use data from the chatbot to inform its broader AI efforts and may incorporate LLMs from other vendors besides OpenAI over time.

D-ID, an Israeli company using AI for unique experiences, has launched a new chat API to enable face-to-face conversations with an AI digital human. Link.

• The API provides a "human" interface for conversational AI and can integrate large language models like GPT-3 and LaMDA to deploy interactive digital humans.

• Developers can create photorealistic digital assistants that can interact with consumers in a more human and engaging way.

• The new offering can be used for branding and customer experience purposes, and for brands to engage with customers through a more personal and interactive experience.

• D-ID is part of a framework for the ethical and responsible development, creation and sharing of synthetic media, along with other companies like Adobe and OpenAI.

• D-ID aims to make interactions with AI feel more natural and help brands leverage the trending technology, given the increasing popularity of AI and chatbots for consumer interactions.

SPEECH/AUDIO

OpenAI launched the Whisper API, a hosted version of its open-source Whisper speech-to-text model. Link.

• Whisper is an automatic speech recognition system that offers robust transcription and translation from multiple languages into English.

• Whisper was trained on 680,000 hours of multilingual and "multitask" data collected from the web, leading to improved recognition of unique accents, background noise, and technical jargon.

• The Whisper API is priced at $0.006 per minute and can take files in various formats.

• Despite its limitations in "next-word" prediction and language bias, OpenAI sees Whisper's transcription capabilities being used to improve existing apps, services, products, and tools.

• The speech-to-text market could be worth $5.4 billion by 2026, and OpenAI aims to be a "universal intelligence" that can handle any kind of data or task.

Spotify is adding to its suite of personalization tools with “DJ”, a personalized AI guide that creats your own custom radio station full of new music. Link.

• DJ will deliver a curated lineup of music with commentary around the tracks and artists, constantly refreshing the lineup based on your feedback.

• DJ will scan the latest releases and take you back to nostalgic playlists, making listening feel completely personal to each user.

• The powerful combination of Spotify's personalization technology, generative AI, and the expertise of music editors will make the AI DJ possible.

• The generative AI tooling allows editors to scale their innate knowledge and expertise in genres and music culture.

• The AI DJ will use a dynamic AI voice platform from the Sonantic acquisition, with the first voice model being the Head of Cultural Partnerships, Xavier "X" Jernigan.

CODE/DEVTOOLS

Microsoft's Responsible AI Toolbox is a collection of tools designed to help practitioners maximize the benefits of AI systems while mitigating harms. Link.

• The targeted model improvement approach identifies solutions tailored to the causes of specific failures, supporting practitioners in better addressing failures without introducing new ones or eroding other aspects of model performance.

• The Responsible AI Mitigations Library allows practitioners to experiment with different techniques for addressing failures, while the Responsible AI Tracker uses visualizations to show the effectiveness of different techniques for more informed decision-making.

• The tools in the Responsible AI Toolbox have been designed with each stage of the model improvement life cycle in mind, informing targeted model improvement through error analysis, fairness assessment, data exploration, and interpretability.

• Building tools for responsible AI that are intuitive, effective, and valuable can help practitioners consider potential harms and their mitigation from the beginning when developing a new model.

• The Responsible AI Toolbox is available in open source and through the Azure Machine Learning platform offered by Microsoft.

Prog.ai is a startup that uses AI to infer skills from public code on GitHub to help recruiters find the most suitable candidates for technical roles. Link.

• The platform uses GitHub's "git clone" command to create a copy of millions of public repositories and branches and analyzes each git commit to figure out the roles and engagement levels of project contributors.

• Prog.ai leans on OpenAI's GPT to derive scores on code quality, and users can build lists of top experts in specific disciplines and generate a leaderboard of top performers in any given field.

• Prog.ai has a built-in email outreach engine that allows users to create personalized email sequences to connect with software developers.

• The platform also gleans publicly available data from LinkedIn and aligns it with the corresponding individual from GitHub to build a more detailed picture of potential candidates.

• Prog.ai's future plans include expanding beyond hiring to help developer tool companies pitch their wares and foster developer relations.

3D/ROBOTICS

Ford has launched a subsidiary, Latitude AI, to focus on commercializing hands-free, eyes-off driver-assistance systems. Link.

• The new company will be staffed by 550 employees, mainly former machine learning and robotics engineers of Argo AI, an autonomous vehicle startup that folded last year after Ford and VW pulled their backing.

• Latitude will improve Ford's existing advanced driver-assistance system, BlueCruise, and develop new automated driving technology for its next-generation vehicles.

• Latitude will be headquartered in Pittsburgh, with additional engineering hubs in Michigan and California and operate a highway-speed test track in South Carolina.

• Sammy Omari will serve as the CEO of Latitude, Peter Carr as CTO, and David Gollob as the president.

• Ford aims to redefine the relationship between people and their vehicles, making travel safer, less stressful, more enjoyable, and giving customers some of their time back.

MULTIMODAL

Typeface, a startup founded by former Adobe CTO Abhay Parasnis, has emerged from stealth with $65 million from Lightspeed Venture Partners, GV, M12, and Menlo Ventures. Link.

• Typeface offers an AI-powered dashboard for drafting marketing copy and images that combines generative AI with a brand's tone, audiences, and workflows to enhance content workflows and corporate content development.

• Using Typeface, customers can type in a command like "Write a fun blog post about apple juice" to have the platform execute it, writing a several-paragraph draft piece complete with images.

• There is a growing industry of generative AI startups focused on marketing and ad-specific applications, with startups like Movio, Copysmith, Copy.ai, Sellscale, Jasper, Omneky, and Regie.ai using generative AI to create better marketing copy, imagery, and even video for ads, websites, and emails.

• With increasing competition, it is not clear which startups will come to stand above the rest in terms of market traction. However, Parasnis asserts that Typeface has a fighting chance owing chiefly to its platform's safety and governance capabilities as well as its ability to incorporate "brand-specific" visual assets.

• Typeface already has customers in industries including marketing, advertising, sales, HR, and customer support, and its unique enterprise-minded vision for micro-personalized, secure content for teams is appealing to mid-large enterprises.

HEALTHCARE

Researchers have achieved the ability to reconstruct high-resolution and accurate images from brain activity using the Stable Diffusion image generation model. Link.

• Unlike previous studies, the researchers did not need to train or fine-tune the AI models to create the images.

• The researchers predicted a latent representation from fMRI signals and added noise to it through the diffusion process to produce a final image.

• Prior to this study, no other researchers had tried using diffusion models for visual reconstruction.

• The study provides a quantitative interpretation of the model from a biological perspective and shows a correlation between stimuli and noise levels in the brain.

• The advancement of generative AI has led to more researchers testing the ways AI models can work with the human brain.

British researchers are developing a new technology that uses artificial intelligence (AI) to increase the number of organs available for transplant. Link.

• The project, called OrQA (Organ Quality Assessment), has secured more than £1 million in funding from the National Institute for Health and Care Research (NIHR).

• The AI technology uses tens of thousands of images of donor organs to identify those that offer the best chance of transplant success, which could result in up to 200 more kidney and 100 more liver transplants every year in the UK.

• The technology is being developed by a team including experts from the University of Oxford and is backed by ministers, NHS Blood and Transplant (NHSBT) and the NIHR Blood and Transplant Research Unit.

• The AI algorithm will be trained using thousands of images of human organs to assess images of donor organs more effectively than what the human eye can see.

• The technology will revolutionize the way we care for people, improve organ transplant services, and help close the gap between those patients waiting for and those receiving lifesaving organs.

POLICY/LAW

The US military is increasing its commitment to autonomous weapons with a new Department of Defense directive, following NATO's implementation plan aimed at preserving technological edge in warfare. Link.

• Recent conflicts in Ukraine and Nagorno-Karabakh have shown that weaponized AI is the future of warfare.

• The pressure to achieve decisive battlefield advantages with fully autonomous weapons is increasing, as casualties mount in Ukraine.

• Proponents argue that fully autonomous weapons keep soldiers out of harm's way and allow for superhuman speed in military decisions, but critics argue that autonomous weapons lack human judgment and erode meaningful human control.

• The updated directive promises ethical use of autonomous weapons systems and oversight for developing and employing the technology.

• International law currently does not provide an adequate framework for understanding or regulating autonomous weapons systems.

OTHER

SambaNova Systems has launched the SambaNova Suite, aimed at helping enterprises build and deploy customized generative AI models. Link.

• The suite expands on the dataflow service with a collection of capabilities that enables organizations to customize both open-source and proprietary generative AI models to meet their specific requirements.

• SambaNova has its own custom silicon that is optimized for both machine learning-training and inference, enabling continuous training and inference capability to keep data that feeds the generative AI models up to date.

• The platform provides transparency to organizations for how a given AI model actually works and helps limit bias and provide responsible and explainable AI.

• Content creation is a particularly vibrant field of enterprise generative AI customization.

• SambaNova's key differentiator from others in the generative AI space is its own hardware, which is extremely power efficient and performative, and enables a continuous training and inference capability.

Elon Musk has approached AI researchers to form a new research lab to develop an alternative to OpenAI's high-profile chatbot, ChatGPT. Link.

• Musk has criticized OpenAI for installing safeguards that prevent ChatGPT from producing text that might offend users.

• Musk is recruiting Igor Babuschkin, a researcher who specializes in machine-learning models that power chatbots like ChatGPT, to spearhead the effort.

• Musk's goal is to improve the reasoning abilities and the factualness of language models, making their responses more trustworthy and reliable.

• The project is still in the early stages, with no concrete plan to develop specific products, and could end up within his social media app Twitter or become a standalone AI lab.

• Babuschkin left Alphabet's DeepMind last week but has not officially signed onto the Musk initiative.

EleutherAI, a community research group, is forming a nonprofit foundation to develop AI systems and avoid the pitfalls faced by other open-source projects. Link.

• The EleutherAI Institute will be funded by donations and grants from backers including AI startups Hugging Face and Stability AI, former GitHub CEO Nat FriedmanG, Lambda Labs and Canva.

• The organization started several years ago as a grassroots collection of developers working to open source AI research and has since grown to over 20 full-time staff members.

• EleutherAI members have co-authored 28 academic papers, trained dozens of models, and released ten codebases.

• The Foundation will focus on developing large language models, as well as devoting more resources to ethics, interpretability, and alignment work.

• The Foundation will strive to remain independent and avoid running dry, as OpenAI did when it transitioned to a "capped-profit" structure to fund its ongoing research.

GPT Tip of the Week

Prompt: "What are the key factors in evaluating the efficacy and defensibility of AI smart agents?"

Response:

• Task Performance: One key factor in evaluating the efficacy of an AI smart agent is its ability to accurately and efficiently complete its intended tasks. This can be measured through metrics such as accuracy, precision, recall, and F1 score.

• Adaptability and Generalization: Another important factor is the agent's ability to adapt to changing conditions and learn from experience. This can be evaluated through metrics such as the agent's learning curve and its ability to generalize to new situations.

• Data Privacy, Security, and Transparency: When evaluating the defensibility of an AI smart agent, it's important to consider factors such as data privacy, security, and transparency. The agent should be designed with strong security measures to protect sensitive data and prevent unauthorized access. Additionally, the agent's decision-making process should be transparent, so that it can be audited and explained if necessary.

• Bias Detection and Mitigation: Another important factor in evaluating the defensibility of an AI smart agent is its ability to be audited and tested for potential biases. This requires careful monitoring of the agent's decision-making process, and the ability to identify and address potential sources of bias in the data and algorithms used by the agent.