AI Weekly: 05/29/23

Good morning and welcome to this week’s edition of AI Weekly! In another week of disruptive headlines, Adobe has integrated their image augmentation and editing tool, Firefly, into Photoshop and it is blowing people away. On the other hand, Microsoft has taken another step in the AI race by adding their search engine/virtual assistant, Windows Copilot, to the Windows sidebar.

In healthcare news, Neuralink has received clearance from the FDA to conduct human trials on its brain-implant technology, and a group of London-based scientists have developed an AI model that can predict the spread of a certain type of breast cancer before it reaches other organs.

In funding news, Sound made headlines again as part of the $450 million that Anthropic recently raised. Big things ahead! Enjoy news on the rest of last week’s AI happenings below.

- ZG

Here are the most important stories of the week:

TEXT

AI research company Anthropic has raised $450 million in funding to advance its development of safe large language models and compete with OpenAI. Link.

• The funding round was led by Spark Capital, with participation from Google, Salesforce Ventures, Sound Ventures, Zoom Ventures, and others.

• Anthropic had previously raised $1 billion in funding and was valued at $4.6 billion after its last funding round in March.

• The new funding will support Anthropic's work on "helpful, harmless, and honest" AI systems, including product development, safety research, hiring, and support for businesses using its LLMs and assistants.

• Anthropic aims to raise up to $5 billion in the next two years to challenge OpenAI and expand into various industry verticals, providing API access to its LLMs and tools.

• Anthropic's Claude AI assistant is positioned as a rival to OpenAI's ChatGPT, offering a larger context window for analyzing information. Anthropic's founders have backgrounds at OpenAI, including leading LLM development and safety functions.

Google has started rolling out its AI-powered search experience, called Google Search Generative Experience (SGE), in the US. Link.

• Users can sign up for the waitlist and will receive an email when it's their turn to access the new search feature.

• Google's SGE integrates conversational AI into the traditional search experience, with generative AI results appearing below the search bar but above standard web results.

• A snapshot feature provides key information and links to explore further, and users can expand the snapshot for more details.

• Google is also opening access to other Search Labs, including Code Tips for programming guidance and Add to Sheets, allowing users to insert search results directly into Google Sheets.

• To access Search Labs, users can tap the Labs icon in Chrome for desktop or in the Google search app on Android or iOS, but availability and access are currently limited.

Microsoft is integrating AI into Windows 11 with the introduction of the Windows Copilot tool. Link.

• Windows Copilot, residing in the Windows sidebar, functions as a search engine and virtual assistant, allowing users to perform tasks within the OS and obtain summarized documents.

• The tool is deeply integrated with Windows and aims to empower users, making them more productive and creative.

• Windows Copilot will be available for preview in June and has the potential to reach a vast user base, given Windows' large number of monthly active devices.

• Microsoft recognizes the importance of responsible AI usage and is working on addressing concerns such as privacy, security, and AI misinformation.

• The company aims to learn from user feedback during the preview phase to improve the tool and build trust with customers.

TikTok is testing its own AI chatbot called "Tako" in select markets, appearing on the right-hand side of the TikTok interface. Link.

• Users can ask Tako questions about videos using natural language queries or ask for content recommendations.

• The bot is being publicly tested on iOS devices, but TikTok clarified that it is not currently available to the public in the U.S.

• TikTok is using a third-party AI provider customized for its needs, not in-house AI technologies.

• Tako is labeled as "experimental" and not reliable for medical, legal, or financial advice.

• Conversations with Tako are reviewed for safety purposes and user experience enhancement, and users have the option to delete their chats. Privacy aspects and data retention policies are unclear.

IMAGE/VIDEO

Adobe is introducing a new Generative Fill feature to Photoshop, leveraging Firefly-powered generative AI to add, remove, and extend visual content based on natural-language text prompts. Link.

• The feature, equivalent to DALL-E 2's inpainting and outpainting, allows users to prompt the AI with specific requests like altering the sky to look surreal or extending the scene beyond the image's borders.

• Adobe claims that Generative Fill maintains the original scene's perspective, lighting, and style, enabling significant image alterations with minimal effort.

• To address ethical concerns, Adobe's current-generation model only learns from Adobe Stock images and copyright-free public domain content. AI-generated images in Photoshop will be encoded with a digital signature indicating their origin.

• Generative Fill is available in the Photoshop desktop beta, and Adobe plans to make it generally available in the second half of 2023.

• Additionally, Generative Fill is accessible on the web as part of the Firefly beta module.

A deepfake image depicting an explosion at the Pentagon circulated on Twitter, causing a temporary dip in the stock market. Link.

• The image was quickly debunked by the Arlington Police Department, stating that there was no incident at the Pentagon.

• The image exhibited characteristics of an AI-generated image, such as blurred elements and inconsistencies.

• Twitter's verification system compounded the issue, with a verified account impersonating a Bloomberg news feed sharing the image.

• The incident highlights the risks of misinformation associated with generative AI and the challenges of detecting deepfakes.

• It serves as a reminder of the potential real-world consequences of fake images shared on social media platforms.

Google announced at its I/O 2023 event that its generative AI Google Bard, similar to ChatGPT, is now available for free in over 180 countries and supports images in its responses. Link.

• Bard can display photos from the web, enhancing the user experience and making research more visual.

• Microsoft Bing Chat also supports photos in AI responses, offering a more visual experience compared to text-only conversations.

• Google enabled support for images in Bard's search results, allowing users to receive visual responses from the AI.

• Users can ask Bard to show them images directly or during conversations, and each image is accompanied by a source.

• While Bard can show images, the number of images per response may vary, and users may need to refine their prompts or explore different versions of the answer to achieve the desired outcome.

Google has launched Product Studio, a free tool powered by generative AI that allows Google Shopping merchants to edit and customize their product images. Link.

• Product Studio aims to help sellers diversify their product and lifestyle images, making their listings more appealing and standing out against competitors.

• The tool can generate different backgrounds for seasonal campaigns, remove existing backgrounds, and enhance the quality of small or low-resolution images.

• It is built on one of Google's text-to-image models, though it is not specified which AI model is used.

• Product Studio is designed for businesses of all sizes, particularly benefiting independent sellers and small businesses with limited budgets for professional photography.

• While similar generative AI tools exist, Product Studio offers convenience for merchants already using Google's Merchant Center, eliminating the need to switch between platforms.

SPEECH/AUDIO

Amazon aims to leverage generative AI to enhance its virtual assistant Alexa and create a more conversational and personalized user experience. Link.

• The use of generative AI could enable Alexa to have more natural conversations, answer complex questions, and even create content such as stories or song lyrics.

• Amazon's goal is to make Alexa the ultimate source of knowledge and provide a conversational experience comparable to the "Star Trek" computer.

• Despite cost-cutting measures and layoffs at Amazon, the Alexa division remains a priority with thousands of employees working on its development.

• Amazon recently introduced new Echo smart speaker products that offer incremental updates and hints at the future direction of its AI efforts.

• Generative AI continues to be a key focus for Amazon, with plans to integrate it into various consumer experiences, including e-commerce search capabilities and potentially a home robot.

CODE/DEVTOOLS

Microsoft has introduced Copilot, an AI-powered assistant for its low-code website creation tool, Power Pages. Link.

• Copilot can generate text, forms, chatbots, web page layouts, and design themes based on prompts provided by users.

• It enables users to describe their needs in natural language and use Copilot's suggestions to design web pages and create content.

• The AI assistant can build complex business data forms, automatically generating the necessary backend database tables.

• Microsoft emphasizes that Copilot is not intended for generating spam sites and has safeguards in place, such as offensive content filtering.

• Users have control over the suggestions provided by Copilot and can choose to apply them, modify them, or discard them according to their preferences.

Arc, a jobs platform for remote software developers, has launched HireAI, an AI-powered tool to simplify recruitment. Link.

• HireAI, powered by OpenAI's GPT-4, automates tasks like resume screening and mass outreach to find suitable candidates.

• Companies upload their job description, and HireAI provides a shortlist of candidates, refining preferences with each match for more accurate results.

• The tool automates the onboarding process by generating job descriptions through AI recruiter conversations and providing instant matches.

• AI-powered tools like HireAI free up recruiters' time for more strategic recruitment tasks, such as assessing cultural fit and building relationships.

• The use of AI tools can potentially promote diversity, equality, and inclusion by removing biases related to age, ethnicity, and gender during the initial matching process.

HEALTHCARE

Neuralink has received clearance from the FDA to conduct human trials on its brain-implant technology. Link.

• The company aims to treat neurological disorders and enhance human performance using their chip.

• Previous safety concerns and rejections from U.S. regulators have surrounded Neuralink's human testing proposals.

• The approval is seen as an important step towards helping people with paralysis regain independence and restoring vision.

• Neuralink's objective for the upcoming trial has not been specified.

• Congress members have called for an investigation into possible conflicts of interest and animal welfare violations related to Neuralink's animal testing oversight.

Researchers have used generative AI to reconstruct "high-quality" video from brain activity, according to a new study. Link.

• The researchers developed a model called MinD-Video, which utilizes fMRI data and the Stable Diffusion AI model to generate videos based on brain readings.

• The model was trained using a dataset of videos and fMRI brain readings from test subjects.

• The reconstructed videos closely resemble the original videos shown to the subjects, with slight differences in subjects and color palettes.

• The researchers found that the reconstructed videos were of "high-quality" in terms of motion and scene dynamics, with an accuracy of 85%.

• The study highlights the potential applications of generative AI in fields like neuroscience and brain-computer interfaces, while providing insights into the visual cortex and the hierarchical operation of the fMRI encoder.

London-based scientists have developed an AI model that can detect changes in lymph nodes to predict the spread of triple negative breast cancer, a more aggressive form of the disease. Link.

• The AI model, based on generative AI, aims to help doctors plan treatment and provide patients with peace of mind about the likelihood of cancer spreading.

• By analyzing immune responses in lymph nodes, the AI model can predict the spread of cancer even before it reaches other organs.

• The research team tested the AI model on more than 5,000 lymph nodes donated by 345 patients, establishing its ability to predict the likelihood of cancer spreading.

• Clinical trials are planned to further validate and refine the AI model, with the aim of developing AI-powered software for pathologists to benefit patients with triple negative breast cancer.

• Triple negative breast cancer accounts for around 15% of all breast cancers in the UK and is more prevalent among women with altered BRCA genes, black women, and premenopausal women.

POLICY/LAW/ETHICS

Google's Sundar Pichai has agreed to collaborate with lawmakers in Europe on an "AI Pact" to establish voluntary rules or standards for AI applications. Link.

• The EU aims to proactively work on the AI Pact ahead of incoming legislation that will regulate AI in the region.

• The chatbot initiative called Jugalbandi, developed by Microsoft, OpenNyAI, and AI4Bharat, is making progress in providing information access to villagers in India.

• Jugalbandi uses WhatsApp as a platform and offers responses in local languages using language models and reasoning models.

• The EU has previously established voluntary agreements with tech giants to address online hate speech and disinformation, but self-regulatory initiatives lack legal enforcement.

• The EU is leading in digital rule-making and has drafted regulations for AI, but they are still under discussion and may take time to be implemented, leading to a push for stop-gap measures.

Adobe has set strict user guidelines on its recently released tool Firefly, including a prohibition on creating pornographic material or explicit nudity. Link.

• Adobe has released a beta version of its generative AI tools called Adobe Firefly, aimed at transforming content creation and monetization.

• This restriction is likely a response to concerns about the potential for AI tools to facilitate the creation of deepfake pornography and abusive content.

• The distinction between porn and art is complex, and Adobe's outright ban on explicit nudity raises questions about the role of digital art in their products.

• It remains unclear how Adobe plans to enforce these guidelines and prevent the creation of prohibited content.

• The situation highlights the ethical challenges surrounding generative AI and the tension between companies' business interests and the potential for misuse of AI tools.

OpenAI's leadership believes that rapid AI development necessitates the establishment of an international regulatory body similar to the one governing nuclear power. Link.

• Existing authorities are insufficient to effectively regulate AI due to its fast-paced innovation.

• AI, exemplified by OpenAI's ChatGPT, poses both a unique threat and a valuable asset.

• An international authority for superintelligence efforts, akin to the International Atomic Energy Agency (IAEA), should inspect systems, set safety standards, and impose restrictions.

• Tracking compute power and energy usage dedicated to AI research can be an objective measure to report and monitor.

• OpenAI's call for regulation is a conversation starter in the industry and shows support from the largest AI brand, but the design of such a regulatory mechanism is still unknown. OpenAI acknowledges the need for public oversight but wants to maintain the potential benefits of AI while mitigating risks from bad actors.

OTHER

FlowX.ai, a startup based in Romania, has raised $35 million in a Series A funding round led by Dawn Capital, with participation from PortfoLion, SeedBlink, and DayOne Capital. Link.

• The company's platform allows enterprises in the financial sector to integrate and run applications and services based on legacy and newer software.

• FlowX.ai primarily serves customers in the financial sector, including BNP Paribas, OTP, Banca Transilvania, and Alpha Bank, and aims to expand its customer base over time.

• The startup focuses on helping traditional financial institutions launch new services quickly to compete with neobanks and other innovative players in the market.

• FlowX.ai leverages AI to automate integration, application, and service creation processes, addressing the challenge of integrating old and new systems in the finance industry.

• The startup's open architecture and connector technology enable users to integrate any system and build applications using any programming language, with AI and generative AI supporting natural language queries and fine-tuning by engineers.

Alteryx has introduced a generative AI engine called AiDIN as part of its cloud platform suite. Link.

• AiDIN offers multiple new LLM-powered capabilities to help enterprises drive actionable insights from data and improve productivity.

• The engine includes features such as Magic Documents, which automates the summarization and delivery of analytical insights in plain language for different communication mediums and audiences.

• Workflow Summary uses ChatGPT to generate concise summaries of Alteryx workflows, enhancing documentation for governance and auditability.

• Alteryx Designer will have an OpenAI connector to implement generative AI into data and analytics workflows for various use cases.

• Alteryx plans to bring more capabilities under the AiDIN engine and expand its offerings to reimagine analytics in 2023.

Uber is set to launch Waymo self-driving cars in the Phoenix area, potentially signaling the end of the feud between the two companies. Link.

• This marks Uber's return to the autonomous vehicle industry after selling its AV research unit earlier in 2021.

• Waymo gains access to passenger rides and food delivery for expanding its self-driving car testing.

• The partnership will start with a limited number of robotaxis in a specific area later this year.

• Self-driving cars are seen as crucial to Uber's future, and ride-sharing platforms are expected to be the first to deploy autonomous vehicles on a broad scale.

• Waymo is considered one of the most advanced autonomous vehicle companies, while other players like Cruise and Tesla are also pursuing self-driving technology.

Nvidia partners with Microsoft to integrate Nvidia AI enterprise software into Azure machine learning and bring deep learning frameworks to Windows 11 PCs. Link.

• The partnership aims to accelerate enterprise and individual AI efforts and streamline the development and deployment of AI applications.

• Azure ML users gain access to Nvidia AI enterprise's software platform, AI frameworks, pre-trained models, and development tools.

• Nvidia AI enterprise and omniverse cloud will be available in the Microsoft Azure marketplace.

• Nvidia's GPU-accelerated deep learning frameworks will be enabled on Windows 11 through the Windows subsystem for Linux (WSL), allowing AI models to run on Windows laptops.

• Users have the option to perform large-scale training on Azure and deploy the models on Nvidia PCs.

Jugalbandi, a chatbot developed by Microsoft, OpenNyAI, and AI4Bharat, is bringing information access to villagers in India. Link.

• It provides insights into over 170 government programs in 10 indigenous languages through the widely used platform WhatsApp.

• The chatbot leverages language models from AI4Bharat and reasoning models from Microsoft Azure OpenAI Service.

• The initiative aims to bridge the language and literacy gaps that hinder access to government programs in smaller towns and villages.

• Microsoft sees potential for Jugalbandi to expand its reach and address a broad spectrum of needs in India.

• Microsoft is also collaborating with Gram Vaani, a Delhi-based firm, to democratize information access for the broader Indian populace, particularly farmers.

Red Hat has announced Lightspeed, a generative AI-driven tool for its Ansible IT automation tool. Link.

• Lightspeed, developed in collaboration with IBM, is trained on Ansible's automation language to make it easier to build automations.

• The tool creates workflows based on user descriptions and shows the sources of its work for transparency and trust.

• Red Hat plans to release a public version of Ansible Lightspeed for training the data model, followed by an enterprise version in October.

• The use of generative AI in Ansible aims to close the skills gap and make subject matter experts more productive.

• Red Hat also announced Event-driven Ansible, which automates advanced developer or SRE tasks and is now generally available.

Builder.ai has raised $250 million in a Series D funding round led by the Qatar Investment Authority (QIA). Link.

• The round also saw participation from existing and new investors, including Iconiq Capital, Jungle Ventures, and Insight Partners.

• With this funding, Builder.ai has raised over $450 million and achieved 2.3x revenue growth.

• The company aims to focus on using human conversation to enable app development on its platform, moving away from traditional no-code/low-code systems.

• Builder.ai has offices in multiple countries, including the UK, US, UAE, Singapore, and France.

• The company's innovative technology and growth potential align with QIA's strategy of supporting forward-thinking companies shaping the global economy.

GPT Tip of the Week

Prompt: What are the different ways people can ensure privacy as voice-driven products like Alexa increasingly implement LLM technology and monitor our every word?

Response: As voice-driven products like Alexa continue to evolve and implement language learning models (LLMs), it's understandable that privacy concerns may arise. While complete privacy is difficult to achieve in such scenarios, there are several steps individuals can take to enhance their privacy:

1. Review privacy settings: Familiarize yourself with the privacy settings and options provided by the voice-driven product. Opt-out of any data collection or sharing features that you find intrusive. Regularly check for updates to these settings as new options may be introduced over time.

2. Mute or disable the device: Consider muting or disabling the voice-driven product when you don't need it actively listening. This can provide peace of mind by preventing accidental activations or recordings.

3. Be cautious with sensitive conversations: Avoid discussing highly sensitive or personal information near voice-driven devices. If you need to have a private conversation, move to a different location or turn off the device temporarily.

4. Routinely delete recordings: Check if the voice-driven product offers the option to review and delete your recorded conversations. Regularly go through the recorded data and remove any information you don't want stored.

5. Use a separate voice profile: If available, set up separate voice profiles for different users. This can help to differentiate between individuals and limit the potential for cross-referencing conversations.

6. Consider alternative products: Research alternative voice-driven products that prioritize privacy and implement stricter data protection policies. Look for products that explicitly state their commitment to user privacy and limit data collection and storage.

7. Secure your home network: Ensure your Wi-Fi network is password protected and encrypted. Regularly update your router's firmware to benefit from security patches. Additionally, consider segmenting your network and creating a separate guest network for IoT devices.

8. Read privacy policies: Take the time to read and understand the privacy policies of voice-driven products. Pay attention to how they handle data collection, storage, and sharing. Look for products that prioritize user consent and provide clear explanations of their practices.

9. Explore local processing options: Some voice-driven products offer options for local processing, meaning the voice data is processed on the device itself instead of being sent to the cloud. This can enhance privacy by minimizing the amount of data transmitted and stored externally.

10. Keep up with updates and news: Stay informed about the latest advancements, updates, and news related to voice-driven products. Companies often introduce new features and privacy enhancements based on user feedback or regulatory requirements.

Remember that while these steps can help improve privacy to some extent, they may not provide absolute assurance. If privacy is a top concern, it's important to carefully consider the trade-offs associated with using voice-driven products and make an informed decision about their use in your home or personal environment.