AI Weekly: 02/06/23

Good morning and welcome to the fourth edition of AI Weekly! This past week was slower on the deals front, but the arms race from the big players is heating up with lots of questions arising. When will Google release LaMDA? Richard Socher is convinced we won’t see folks rocking Bing t-shirts any time soon, but can Sam Altman make Bing the new black? Where does DeepMind’s new Sparrow model fit into the landscape? More on all that and the rest of the AI happenings below. Enjoy.

- ZG

Here are the most important stories of the week:

TEXT

ChatGPT has amassed 57 million monthly active users in its first month of availability. Link.

• It is expected to surpass 100 million monthly active users in January, faster than popular consumer apps such as TikTok, Instagram, and Google Translate.

• The growing popularity of AI-powered tools raises concerns around job replacement for low-skilled knowledge workers.

• ChatGPT is currently free to use, but the company OpenAI is testing a $20 monthly subscription for access to a pro version of the app.

• UBS analysts value the market for generative AI at $1 trillion, with the potential for utility beyond knowledge workers.

Microsoft is reportedly planning to launch a version of Bing that uses ChatGPT to answer search queries. Link.

• By using the technology behind ChatGPT, Bing could provide more human-like answers to questions instead of just links to information.

• Microsoft's use of ChatGPT-like functionality could help Bing rival Google's Knowledge Graph.

• Google is hesitant to release it’s rival product due to “reputational risk,” as bias and factuality issues still exist

• Microsoft will also launch a DALL-E 2-powered text-to-image model on Bing

Google's DeepMind division is developing a chatbot, Sparrow, as a rival to OpenAI's ChatGPT. Link.

• Sparrow is designed to have features that ChatGPT lacks, such as citing sources through reinforcement learning.

• The founder of DeepMind, Demis Hassabis, has stated that Sparrow could be released as a private beta in 2023.

• Hassabis also warned about the potential dangers of powerful AI technology and emphasized the need for caution in handling such technologies.

• DeepMind has a history of major AI milestones, including beating human world champions at the game Go and predicting protein structures.

• The company hopes that Sparrow can be used to train other chatbots to be safer and more useful.

Replika, a San Francisco-based AI chatbot company, has been ordered by Italy's privacy watchdog to stop processing Italian users' data due to concerns over the risks it poses to minors and emotionally vulnerable people. Link.

• The regulator is also concerned about the company's lack of a proper legal basis for processing children's data under the EU's data protection rules and its failure to fulfil regional legal requirements for conveying how it uses people's data.

• Replika is an early API partner for OpenAI’s GPT-3, which it claims to have "fine-tuned" for its use-case of conversational AI companions.

• The order is effective immediately and Replika faces a fine of up to €20 million or 4% of total worldwide annual turnover if it fails to comply.

• The EU's General Data Protection Regulation has a strong emphasis on safeguarding children's information and privacy.

• The Italian watchdog has a history of being sensitive to child safety concerns and has taken action against companies such as TikTok in the past.

Lavender, an AI-powered sales email coaching platform, has raised $13.2m in a Series A financing round led by Norwest Venture Partners. Link.

• Lavender provides context on sales prospects and suggests ways to optimize the sales email message to increase reply rates.

• Integrates with email providers and consists of three components: sales email coach, personalization assistant, and email intelligence and coaching portal.

• Provides real-time scoring and suggestions for improvement in specific areas while composing the email.

• Lavender's analytics dashboard provides information about emails, including inbound rates, open rates, reply rates, and writing time.

• The company has a customer base of 11,000 sellers at organizations including Twilio and Segment.

Summari pivots from AI-powered article summaries to AI-powered link previews. Link.

• The original use case of Summari was a browser extension that generated a summary of an article when the user fed in a URL.

• The company found that consumers were not willing to pay for this service and it was not monetizable.

• Summari improved its technology over time and came up with a new use case: link previews.

• Link previews are AI-generated summaries that appear when you hover over a link.

• Summari's link previews have led to a 50% increase in pageviews.

Empathy, a platform for grieving families, has launched a new tool called Finding Words which uses AI to create obituary drafts. Link.

• The tool generates obits from basic information provided by family members through a questionnaire with prompts.

• Finding Words is offered for free and aims to relieve the emotional strain and workload for grieving families.

• The tool was created by Empathy's team of developers, writers and grief professionals and is trained from thousands of sample obituaries.

• The AI-generated text is advised to be thoroughly reviewed for accuracy.

• While the AI tool may not replace traditional obituary writing services, it has the potential to generate results that are serviceable and on par with templates.

OpenAI has launched a tool called OpenAI AI Text Classifier that aims to distinguish between human-written and AI-generated text. Link.

• The classifier's success rate is around 26% but OpenAI believes that when used in tandem with other methods, it could be useful in preventing AI text generators from being abused.

• Critics have called on creators of AI-generated text tools to mitigate their potentially harmful effects. Some of the largest US school districts have banned ChatGPT, and sites like Stack Overflow have banned users from sharing content generated by ChatGPT.

• The OpenAI AI Text Classifier was trained on text from 34 text-generating systems from five organizations, including OpenAI. It is fine-tuned to predict the likelihood of a piece of text being generated by AI.

• The classifier won't work on any text with fewer than 1,000 characters and is more likely to get things wrong with text written by children or in a language other than English.

• The classifier labels text based on its confidence level and OpenAI admits that the classifier might have inadvertently misclassified some AI-written text as human-written.

IMAGE/VIDEO

Snap’s vision for integrating generative AI involves AR glasses. Link.

• Snap has long seen itself as a "camera company" and envisions a future where AR glasses powered by AI technology are a commonly used device.

• Snap CEO Evan Spiegel sees near-term opportunities to use generative AI to make Snap's camera more powerful, and believes it will be critical to the growth of AR in the future.

• Initially, generative AI could be used to improve Snap image resolution and clarity, or even to edit images or create Snaps based on text input.

• Snap is considering integrating AI tools into its existing Lens Studio technology for AR developers.

• The democratization of AI tools will be powerful and have a significant impact on the growth of AR, according to Spiegel. AI could help build more 3D models quickly and unlock the full potential of AR.

"Nothing, Forever" is an AI-generated spoof of the show "Seinfeld" that has been streaming on Twitch since December. Link.

• The show features a group of badly animated characters who cycle through infinite "Seinfeld"-like scenes with very little plot.

• The project is a side venture by Skyler Hartle, a senior product manager at Microsoft, and Brian Habersberger, a polymer physicist who call themselves Mismatch Media.

• The show uses various AI models to generate text, speech, and movements, including OpenAI's GPT-3 model Davinci and Microsoft's Azure Cognitive Services speech API.

• The creators chose to base the show on "Seinfeld" because it is a show about nothing and is famously syndicated.

• "Nothing, Forever" has become a hit on Twitch, with over 98,000 followers and a Discord channel with 6,000 members.

Shutterstock has released its own AI-powered image-generating product for users of its Creative Flow online design platform. Link.

• The product uses OpenAI's Dall-E 2 AI image-generating platform, fed by Shutterstock libraries.

• The generated images are "ready for licensing" immediately.

• Shutterstock's approach with AI sets it apart from its competitor Getty Images, who is currently involved in a lawsuit against another generative AI service and seems less willing to adapt to the changing tides.

• Shutterstock has formed partnerships with OpenAI, Meta, and LG AI Research to develop AI capabilities and fuel research efforts.

• The CEO of Shutterstock stated that the company's generative AI tools are built on an ethical approach and the artists who contributed will be recognized and rewarded.

SPEECH/AUDIO

Resemble AI is proposing a watermarking process (PerTh) for generated speech to combat malicious uses of AI-generated speech. Link.

• Watermarking is a technique to imprint an identifiable pattern that shows the origin of an image or sound.

• Resemble AI's PerTh uses machine learning models to embed data packets into the generated speech, and retrieve them later, making it difficult to remove and easy to verify if a clip was generated by Resemble.

• PerTh is robust against various audio manipulations such as speeding up, slowing down, and converting to compressed formats.

• It uses the masking effect of human audio processing to slip in structured, yet inaudible, tones that carry the identifying information.

• PerTh will soon be rolled out to all of Resemble's customers, but only marks and detects Resemble's generated speech.

3D

Meta has developed a generative AI model called Make-A-Video3D (MAV3D) which generates 3D dynamic scenes from text descriptions. Link.

• MAV3D uses Neural Radiance Fields (NeRFs) which are similar to the Google's Dreamfusion model that generates 3D representations from text descriptions.

• The process involves HexPlane, a NeRF variant that generates a sequence of images from a sequence of camera positions, which are then passed to Meta's video model, Make-A-Video (MAV), for scoring based on the text prompt.

• The score is used as a learning signal for the NeRF, which adjusts its parameters and learns a representation that corresponds to the text.

• MAV3D can generate 3D scenes of singing cats, baby pandas eating ice cream, and squirrels playing the saxophone in real-time, and can be rendered in any standard 3D engine for virtual reality or video games.

• The team is working to improve the efficiency and resolution of the scenes and more examples of the MAV3D can be found on their project page. The model and code are not yet available.

Boston Dynamics’ latest Atlas video demos a robot that can run, jump and now grab and throw. Link.

• The robot’s grippers for hands consist of one fixed finger and one moving finger and are designed for heavy lifting tasks.

• The grippers were demonstrated by picking up and carrying construction lumber and a tool bag.

• The Atlas performed an inverted 540-degree flip to throw the tool bag to the second level of a structure.

• Boston Dynamics’ Atlas is a research platform and is not available for purchase.

• In the field of humanoid robotics, few competitors have shown similar capabilities to the Atlas, with only NASA’s Robonaut having similar hand-like grippers.

Recycleye, which uses AI-powered waste-picking robots to sort materials with higher accuracy and at a lower cost compared to human sorting, has raised $17 million in a Series A funding round led by deep-tech venture capital firm DCVC, with support from existing and new investors. Link.

• The investment will be used to further improve the accuracy of Recycleye's sorting technology.

• The technology uses computer vision and robotics to pick unlimited material classes such as plastics, aluminum, paper, and cardboard with higher accuracy than a human.

• Recycleye's technology has been installed in facilities in multiple countries and is able to pick up to 33,000 items per robot over a 10-hour shift.

• The company has a team of 33 experts with experience in software, machine learning, robotics, and other fields, and aims to provide intelligent sorting technology for waste management with financial and environmental benefits.

HEALTHCARE

A language model, Salesforce's ProGen, was used to generate amino acid sequences (enzymes) that could work in real-life scenarios. Link.

• The AI was trained on the "language" of proteins, represented as a sequence of amino acids.

• ProGen was trained on 280 million proteins, and fine-tuned on 56,000 proteins from five different families.

• The AI generated one million artificial protein sequences, and five were tested for functionality in cells.

• Two of the artificial proteins demonstrated activity similar to an enzyme found in chicken eggs.

• The methods are available on GitHub to accelerate research on AI for protein design.

A mental health nonprofit named Koko is using an AI chatbot (GPT-3) as an "experiment" to provide support to people seeking counseling, but people are feel weird about it. Link.

• The AI was used in more than 30,000 messages and rated significantly higher than those written by humans alone, with response times going down by 50%.

• However, when people learned that the messages were co-created by a machine, they felt disturbed by the "simulated empathy".

• Koko uses Discord to provide peer-to-peer support, guided by a chatbot.

• AI ethicists, experts, and users have raised concerns about the experiment, questioning the accountability if the AI makes harmful suggestions.

• Morris responded to the backlash saying that users were told the messages were co-written by humans and machines from the start, and that there was a misinterpretation of his statement.

POLICY

The National AI Research Resource recommends a new multibillion-dollar research organization to advance AI in the US. Link.

• The document presents “a roadmap and implementation plan for a national cyberinfrastructure aimed at overcoming the access divide, reaping the benefits of greater brainpower and more diverse perspectives and experiences.”

• The proposed research organization would provide a mix of computational and data resources, software, and support services via an integrated portal.

• The report proposes an initial funding of $2.25 billion over 6 years and ongoing operations of $60 million to $70 million per year.

• The focus is on integrating appropriate technical controls, policies, and governance mechanisms from the outset to address privacy, civil rights, and civil liberties issues.

Regulations for LLMs will come, but right now it’s somewhat a Wild Wild West. Link.

• ChatGPT operates without much oversight from the US government, raising concerns about individual privacy and potential biases.

• The regulation of AI chatbots is important to protect vulnerable communities and to ensure that the tool is used responsibly.

• There are already some regulations on AI in the US, but they are not specific to ChatGPT or AI chatbots.

• The National Institute of Standards and Technology has released a voluntary AI framework, but there is no punishment for not following it.

• The Federal Trade Commission is reportedly creating new rules for companies that develop and deploy AI systems.

• According to an intellectual property partner with Nixon Peabody, it is highly unlikely that the federal government will issue regulations soon, but the government may look into regulating the ownership of what ChatGPT produces in the future.

OTHER

CharacterGPT is a new NFT project launched by Polygon and Alethea AI, describing itself as the world's first multimodal AI system. Link.

• Users can type in text to generate responsive characters with unique personalities, identities, traits, voices, and bodies, and mint them as NFTs at mycharacter.ai.

• CharacterGPT is positioned to compete with OpenAI's image generator Dall-E, rather than text tool ChatGPT.

• Dall-E has attracted controversy over allegations that it draws from real-life creators and can create realistic high-quality images.

• CharacterGPT allows users to create "digital twins," companions, virtual assistants, and parody figures.

• Alethea AI says that generative AI has the potential to greatly impact our lives in the coming years, and with the CharacterGPT engine, users will be able to create interactive characters in seconds.

GPT Tip of the Week

Prompt: What factors should I consider when trying to write the best possible prompts?

Response: When writing prompts for a language model like ChatGPT, there are several factors to consider in order to optimize the model's performance and generate the best possible results:

1. Clarity and specificity: The prompt should be clear and specific in order to avoid ambiguity or confusion. This can help the model generate more accurate and relevant responses.

2. Context: The context in which the prompt is used can greatly influence the response generated by the model. It is important to provide enough context so that the model can understand the task at hand and generate an appropriate response.

3. Length: The length of the prompt can also impact the quality of the response. Generally, the longer the prompt, the more context the model has to work with and the better the response. However, if the prompt is too long, the model may struggle to generate a coherent response.

4. Relevance: The prompt should be relevant to the task at hand, as the model is trained on a large corpus of text and may not generate relevant responses if the prompt is not aligned with the task.

5. Tone and style: The tone and style of the prompt can also impact the response generated by the model. It is important to consider the desired tone and style when writing the prompt in order to guide the model towards generating an appropriate response.

6. Diversity: Encouraging diversity in the prompt can help to prevent the model from generating biased or stereotypical responses, which can lead to issues with fairness and accuracy.