AI Weekly: 04/17/23

Good morning and welcome to this week’s edition of AI Weekly! When I said last week that it currently felt like “the calm before another wave of the storm,” I’ll be honest, I didn’t think we’d see a wave like this the following week. Amazotn has officially arrived to the party with Bedrock and CodeWhisperer, placing the life expectancy of many emerging startups into question.

In other news, I genuinely think we are very close to being able to create Westworld in real life as a result of Stanford and Google’s recent research simulation. They created a virtual town filled with 25 agents prompted with similar information to simulate human behavior and the results were fascinating.

Also, multiple countries have come out and threatened to ban ChatGPT as questions around privacy and cybersecurity continue to circulate. Learn more about all of that and a ton of other exciting AI news below. Happy reading!

- ZG

Here are the most important stories of the week:

TEXT

Microsoft is integrating its Bing chatbot into SwiftKey on Android and iOS, allowing users to chat with the bot directly from their mobile keyboard and search for things without having to switch between apps. Link.

• Bing on SwiftKey has three major features: Chat, Tone, and Search.

• The Chat functionality lets users access the new Bing on the go for more detailed queries.

• The Tone feature uses AI to customize text to fit any situation, making it easier for users to communicate effectively.

• With the Search functionality, users can quickly search the web directly from their keyboard, removing the need for switching between apps.

• Microsoft also announced that Bing is now available via Microsoft Start, the company's personalized news reading experience, and that Bing is now available in group chats on Skype.

Poe, a chatbot platform, has introduced a new feature allowing users to create their own chatbots in just a few words. Link.

• The feature lets users write a short prompt and choose an existing LLM, either Claude Instant or ChatGPT, as a base for their new creation.

• The user-generated prompt and the base LLM are then used to generate a new chatbot that can be customized and shared.

• According to Poe, this method can produce a wide variety of chatbots for different purposes and personalities.

• Poe offers a user-friendly interface, a unique URL for each bot, a selection of the best bots, and a cost-free service.

• This update puts Poe in competition with Character.AI, another AI startup that can generate chatbots that mimic different personalities.

IMAGE/VIDEO

Meta has open-sourced Animated Drawings, an AI project that allows anyone to bring their doodles to life. Link.

• The project was initially released as a web-based tool in 2021 and uses object detection models, pose estimation models, and image processing-based segmentation methods to create an animated version of a drawing.

• Users can upload a drawing of a single human-like character or select a demo figure and then use a pen and eraser to tweak the drawing before adjusting where the joints should be.

• After that, users can choose from preset animations in four categories: dance, funny, jumping, and walking.

• Meta's goal is to encourage other developers to create new and richer experiences using Animated Drawings as an open-source project.

• The project has already received permission to use over 1.6 million images for training purposes, including logos, anime characters, fish, and stuffed animals, despite the tool stipulating that only human figures would work.

OpenAI has decided to open-source its Consistency Models technology for generative image models. Link.

• Consistency Models are designed to enable one-step and few-step image generation with high sample quality and minimal compute requirements.

• Open-sourcing the technology is a significant step for the company, which has been criticized for its closed-door policy.

• The Consistency Models paper highlights how the consistency distillation method outperformed the diffusion model approach on various benchmarks.

• The model can pick up an image at any level of obscuration and generate a perfect image in a single step, even when fed with missing information.

• While the image quality may not be as amazing as diffusion models, the speed and minimal compute required make it a valuable option for real-time applications.

Bria aims to minimize legal risks by training image-generating AI in an "ethical" way. Link.

• Text-to-image AI is facing legal challenges due to the tendency of AI to regurgitate the data it's trained on and the composition of its training data.

• Bria's platform enables businesses to create visuals for social media posts, ads, and e-commerce listings using its image-generating AI, and it was built from the start with ethical considerations in mind.

• The company’s AI is trained on "authorized" datasets containing content that Bria licenses from partners, including individual photographers and artists, as well as media companies and stock image repositories.

• Bria's revenue share model rewards data owners based on their contributions' impact, allowing artists to set prices on a per-AI-training-run basis.

• Bria is currently serving a range of clients, including marketing agencies, visual stock repositories, and tech and marketing firms.

CODE/DEVTOOLS

Amazon launched Bedrock for generative AI, a new cloud service for building and scaling generative AI chatbots and applications. Link.

• Bedrock uses internal organizational data to fine-tune on a variety of leading pretrained LLMs from Anthropic, AI21, and Stability AI, as well as two new LLMs in Amazon's Titan model family.

• Amazon CEO Andy Jassy explained that most companies want to use large language models, but they don't want to spend billions of dollars and years to train them, so they prefer to work off of a foundational model and customize it for their own purposes, which Bedrock enables.

• Bedrock provides a meta layer of usability for foundation models on AWS, and Amazon is emphasizing its ability to provide a secure environment for organizations to use this type of AI.

• Bedrock's offering of multiple models, including Stable Diffusion, plays to Amazon's history of focusing on choice.

• Low-code platform Pega was noted as one of Bedrock's early adopters, and they intend to use it for a range of use cases in their platform, which they will make available to their customers.

Amazon has made its AI-powered coding assistant, CodeWhisperer, free for individual developers, undercutting the pricing of its Microsoft-made rival. Link.

• Originally only available to Amazon Web Services customers, the newly announced free tier should make CodeWhisperer much more accessible to developers who don’t use AWS.

• CodeWhisperer filters out potentially biased or unfair code suggestions and flags any code similar to open-source training data.

• It also comes with security scanning features that can identify vulnerabilities within a developer's code and suggest ways to close any security gaps.

Chroma, an AI native open-source embeddings database, has raised $18 million in a seed round led by Astasia Myers from Quiet Capital, with participation from notable angels including Naval Ravikant and Jordan Tigani. Link.

• The company also announced a pre-seed round from May 2022, led by Anthony Goldbloom, James Cham, and Nat Friedman.

• Chroma's goal is to make intelligence too cheap to measure and permeate every product and facet of our lives.

• Chroma is an AI-enabled application that provides pluggable knowledge about data, facts, and tools to LLMs (large language models).

• The company is committed to building open source software to democratize robust, safe, and aligned AI systems.

• Chroma is working on new features like query relevancy and an open source distributed system to replace the current database for client/server Chroma.

POLICY/LAW/ETHICS

Italy's data protection watchdog, Garante, has ordered OpenAI to comply with GDPR regulations regarding ChatGPT's processing of Italian users' data. Link.

• OpenAI has been given until April 30 to comply with the measures set out by Garante.

• The compliance measures include transparency in data processing, age gating to prevent minors from accessing the service, clarifying the legal basis for processing people's data, and providing users with ways to exercise rights over their personal data.

• OpenAI must also conduct a local awareness campaign to inform Italians about its processing of their information to train its AIs.

• The company has until May 31 to submit a plan for implementing age verification tech to filter out underage users, with the deadline for having that system in place set at September 30.

• The Garante's formal inquiry into OpenAI's GDPR breaches is ongoing, and additional or different measures may be taken depending on the results of the fact-finding exercise.

Spain's data protection authority has launched a preliminary investigation into OpenAI over suspected GDPR breaches related to its ChatGPT service. Link.

• Italy's DPA had previously ordered OpenAI to stop processing data of local users over a range of GDPR concerns.

• ChatGPT is currently still accessible via a Spanish IP address, and the regulator has not yet issued a suspension order.

• The AEPD has asked the European Data Protection Board to include ChatGPT in a plenary discussion.

• The EDPB task force on ChatGPT will work in parallel to individual authority probes to potentially coordinate GDPR enforcement on the technology across the bloc.

• The AEPD emphasized that development of innovative technologies like AI must fully comply with the EU's data protection framework and the rights and freedoms the GDPR affords individuals.

The German data protection commissioner has warned that Chat-GPT could face a potential ban in Germany due to data security concerns. Link.

• Ulrich Kelber clarified that such a decision would fall under state jurisdiction, but did not mention any current plans to take such action.

• Kelber revealed that Germany has asked Italy for more details regarding its temporary ban on ChatGPT.

• Italy recently banned ChatGPT, making it the first known instance when a government has blocked an artificial intelligence tool.

• Italy's data protection authority accused OpenAI of stealing user data and not having an age-verification system in place.

• OpenAI's Sam Altman said Italy is one of his favorite countries and added, “We of course defer to the Italian government and have ceased offering ChatGPT in Italy (though we think we are following all privacy laws)."

The US Commerce Department is seeking public comment on how to create accountability measures for AI and develop audits and assessments of AI tools created by private industry. Link.

• The National Telecommunications and Information Administration (NTIA) is seeking feedback from researchers, industry groups, and privacy and digital rights organizations on how to establish guardrails for AI systems that are safe, effective, non-discriminatory, and respect individuals’ privacy.

• The Biden administration has introduced a voluntary “bill of rights” and the National Institute of Standards and Technology has published an AI risk management framework for companies to limit the risk of harm to the public.

• US lawmakers introduced more than 100 AI-related bills in 2021, and federal agencies are looking at ways to apply current rules to AI.

• European regulators have proposed a legal framework that would categorize AI systems by risk: unacceptable risk, high risk, limited risk, and minimal risk.

• Microsoft has argued that chatbots can't be easily categorized as they have more than one purpose and are used for low-risk activities, but the government seeks input from the public to develop a responsible AI regulatory framework.

Chinese regulators have proposed restrictive rules for AI models like ChatGPT, requiring user identification and security reviews, and prohibiting certain types of content that "subvert state power" or "undermine national unity." Link.

• The rules come after Chinese tech companies, such as SenseTime, Baidu, and Alibaba, have rolled out their versions of general-purpose language models.

• The Cyberspace Administration of China's proposed restrictions could hinder relevant innovations and the Chinese AI industry's ambitions.

• The draft rules require providers to assume liability and responsibility for training data of models, users of services must be verified as real people, and generated content must be labeled as such.

• Some requirements may be difficult, perhaps impossible to implement by today's companies and R&D efforts, leading to a setback for the industry.

• The draft rules are open for comment for the next month, after which they may or may not be revised and will enter into effect later this year.

Japan is considering the adoption of AI technology such as OpenAI's ChatGPT if privacy and cybersecurity issues are resolved, according to Chief Cabinet Secretary Hirokazu Matsuno. Link.

• OpenAI's CEO, Sam Altman, recently met with Japanese Prime Minister Fumio Kishida to discuss potential business opportunities in Japan and to improve the company's AI models for the Japanese language and culture.

• Japan is aware of the ChatGPT ban in Italy and will evaluate introducing AI to reduce government workers' workload while considering data breach concerns.

• OpenAI presented measures to remedy privacy concerns to the Italian regulator in response to Italy's restriction of ChatGPT.

• The company is working to develop "nuanced policies against behavior that represents a genuine risk to people."

• Altman discussed the potential benefits of AI and how to mitigate the downsides with Japan's Kishida during the Tokyo meeting.

OTHER

Researchers at Stanford and Google used ChatGPT to create a virtual town filled with 25 agents prompted with similar information to simulate human behavior. Link.

• Each agent is essentially a text-based chatbot, and the simulation is more like an improv role-playing game than an advanced AI system.

• The agents are prompted with specific information about their circumstances and are asked to come up with their next actions, which are organized by a complex hidden text layer.

• The experimental setup informs the agents of other agents' actions to create realistic interactions between them.

• Users can write in events and circumstances, and the agents respond appropriately, as any text for them is reality.

• While the experiment is not scalable or practical for games or virtual environments, it has huge implications for simulations of human interactions.

OpenAI will pay up to $20,000 to people who help find bugs in its AI systems, including ChatGPT. Link.

• The bug bounty program was rolled out in partnership with Bugcrowd, a bug bounty platform.

• Rewards for uncovering bugs will range from $200 for low-severity findings to $20,000 for exceptional discoveries.

• OpenAI believes that transparency and collaboration are key to finding vulnerabilities in its technology.

• Certain safety issues related to the models, such as jailbreak prompts, malicious code, and bad language, are not eligible for rewards.

• Greg Brockman, OpenAI's president, recently mentioned on Twitter that the company had been considering starting a bounty program or a network of "red-teamers" to detect weak spots.

Betaworks is launching a program that will award $500k in funding to 10 companies working on AI from mid-June until mid-September. Link.

• The program will provide startups with access to business-building curriculum, accelerated compute, mentorship, events and activities, and collaboration opportunities with other Betaworks cohort ventures.

• Betaworks is not looking at writing a small check into companies across a range of categories; instead, it is looking at the "evolution" of technology and betting on a cohort that's creating or defining a new category.

• The company is specifically looking to recruit companies creating AI tools that "augment the way humans behave, create, play, work and think."

• Betaworks CEO John Borthwick expects copyright, intellectual property, and attribution model ownership issues to play out over the next few years, but he doesn't think it will have an adverse impact on the startups in the space or their business models.

• The firm is funding Stability AI, which is currently embroiled in a legal battle over whether the company infringed on the rights of millions of artists by developing its text-to-image tools using web-scraped, copyrighted images.

Reef.ai, a platform that automates the collection of data around net revenue retention (NRR), has raised $5.2 million in a recent funding round led by Struck Capital with participation from SCV-SBI, Builders VC, Dig Ventures, and several industry angels. Link.

• The system can connect to each system in the organization that touches the customer, such as Salesforce and HubSpot for customer data, to collect revenue and engagement data and build a data warehouse to store the information.

• Reef.ai aims to create a platform for plug and play net retention excellence for early-stage to mature startups to more predictably drive revenue from their existing customer base.

• The founder, Brenton Grimes, was formerly the global head of customer success at MuleSoft and set up an internal system to track revenue data from existing customers, which led him to launch Reef.ai in 2021.

• Reef.ai already has over 30 customers, some with six-figure contracts, and expects to surpass $1 million in ARR this year.

Apptopia's report indicates that 158 AI Chatbot apps hit the app stores in the first quarter of 2023, a 1480% increase year-over-year, mostly thanks to OpenAI's public release of ChatGPT API in March. Link.

• Many AI chatbot apps have broken into top charts in app stores, such as Nova AI, Genie AI, and Chat with Ask AI.

• AI-powered app downloads have increased by 1506% YoY, reaching nearly 20 million in March, with in-app spending almost $3 million in March, a 4184% increase YoY.

• The top 10 GPT-3 powered AI apps reached $6 million in user spending in February 2023.

• Despite the rise in popularity, analysts are cautious about whether any of these pure-play AI-assistant apps have staying power.

• The onslaught of AI apps also brings dubious tools to the app stores.

OpenAI co-founder and CEO Sam Altman believes that we are approaching the limits of LLM size for size’s sake, and that we are at the end of the era where bigger is better. Link.

• Altman sees size as a false measurement of model quality and compares it to the chip speed races of the past, where there was too much focus on a big number.

• He believes that it’s important to keep the focus on rapidly increasing capability, rather than obsessing over parameter count.

• Altman defends OpenAI’s approach to developing large language models, stating that they spend a long time studying the safety model and mitigating as much risk as possible before releasing them.

• He believes that the recent letter requesting OpenAI to pause for six months missed the mark in some ways, and that moving with caution and increasing rigor for safety issues is important.

• Altman says he’s being open about the safety issues and limitations of current models because it’s the right thing to do and to engage in dialogue about this technology to create the future we all want.

GPT Tip of the Week:

Prompt: What are the top 5 books about the LLM era of AI?

Response: Here are five books that provide insight into the era of Large Language Models in AI:

1. "The Alignment Problem: Machine Learning and Human Values" by Brian Christian - This book explores the challenges of ensuring that AI models are aligned with human values, and discusses the ethical implications of LLMs and other machine learning technologies.

2. "Artificial Unintelligence: How Computers Misunderstand the World" by Meredith Broussard - This book critiques the limitations of AI and exposes its weaknesses, including the shortcomings of LLMs and the danger of relying too much on automation.

3. "Language and the Pursuit of Leadership Excellence: How Extraordinary Leaders Build Relationships, Shape Culture and Drive Breakthrough Results" by Chalmers Brothers - While not strictly about LLMs, this book delves into the importance of language and communication in leadership, and can offer insights into the role of language models in shaping our understanding of the world.

4. "Language Models are Few-Shot Learners" edited by Sam Bowman, et al. - This collection of essays explores the capabilities and limitations of LLMs, and discusses the implications of these models for language understanding and natural language processing.

5. "The Rise of Big Data Policing: Surveillance, Race, and the Future of Law Enforcement" by Andrew Guthrie Ferguson - This book examines how law enforcement agencies are increasingly using AI technologies, including LLMs, in their work, and discusses the implications for civil liberties and social justice.