AI Weekly: 04/02/23

Good morning and welcome to this week’s edition of AI Weekly! This week comes with some major news in the AI world. The most relevant of which is an open-letter led by many prominent AI researchers to pause AI development so that the potential risks can be researched and studied. We are constantly talking about the insanely fast pace of AI innovation that is taking place and to me, it says a lot that those profiting the most from this wave are the ones most loudly suggesting we take a step back.

In other news, Anthropic launched Claude for Slack and Google is one step closer to allowing us to never write our own emails again.

Oh, and Italy banned ChatGPT.

Enjoy more about all of that and more below. Happy AI reading!

- ZG

Here are the most important stories of the week:

TEXT

Perplexity.ai raises $25.6 million in Series A funding, led by New Enterprise Associates and featuring participation from prominent investors. Link.

• Company has experienced rapid growth, reaching 2 million monthly active users within four months of launching its answer engine, Ask.

• Long-term mission: to become the go-to platform for answers and information, evolving into a comprehensive knowledge hub.

• Perplexity.ai is committed to providing citations with every answer, ensuring proper attribution and verifiability of information.

• Expansion from web to mobile includes the launch of the Perplexity AI iPhone app, offering an improved mobile experience.

• Supported by users and investors, the company aims to redefine search and access to information, exploring the frontier of interactive possibilities.

Got It AI releases ELMAR, an enterprise-ready, cost-effective large language model (LLM) for chatbot Q&A applications. Link.

• ELMAR is smaller than GPT-3 and can run on-premises, providing a more accessible option for enterprise customers.

• The LLM includes truth-checking on responses and post-processing to mitigate incorrect response rates.

• ELMAR's hardware requirements are less expensive than those needed for OpenAI's GPT-4.

• A Got It AI study found that ELMAR performed as well as larger LLMs on dialog-based use cases.

• Got It AI plans to commence ELMAR's beta program across multiple industries, aiming to improve its speed, accuracy, and cost-effectiveness.

Google is expanding its generative AI features for Gmail and Google Docs to the next round of "trusted testers." Link.

• This brings the features to a select group of public testers for the first time, including consumer, enterprise, and education users in the United States.

• Testers must be invited by Google, sign up, and opt in, with the option to leave the program at any time.

• Google plans to bring AI features to other Workspace apps, such as Slides, Sheets, Meet, and Chat, in the future.

• AI features will help with drafting, replying, summarizing, prioritizing emails, proofreading, writing, generating images, audio, video, and more.

• It is unclear when these AI features will be available to all Google Workspace users or when Google Bard, the competitor to ChatGPT, will be fully launched.

Anthropic launches Claude App for Slack, a virtual teammate, now in beta, offering summarization, Q&A, and more. Link.

• Claude was developed using Anthropic research focused on creating helpful, honest, and harmless AI systems.

• Claude can summarize Slack threads, answer questions about website content, and assist across various organizational functions.

• Users can interact with Claude in channels and group DMs by mentioning @Claude, or through 1-to-1 direct messages.

• Anthropic has utilized Claude internally for tasks like code editing, memo critiques, and drafting customer communications.

• Claude aims to boost human productivity and communication, with Slack's Steve Wood praising its natural conversational skills and ability to retain context.

IMAGE/VIDEO

Read, a meeting intelligence tool, introduces a feature that condenses hour-long meetings into two-minute clips with key points. Link.

• The company uses large language models and video analysis to identify notable parts of meetings, including participant reactions.

• Read works with Zoom, Google Meet, Microsoft Teams, and Webex, providing analytics like sentiment and engagement scores.

• Co-founded by former Foursquare CEO David Shim, Read has raised $10 million in seed funding from investors like Madrona Venture Group and PSL Ventures.

• During a preview period, agency clients experienced a 30%+ increase in productivity using Read's features.

• Read plans to incorporate text summaries in the videos using captions in the coming weeks, differentiating itself from competitors like Otter and Zoom.

Midjourney temporarily disables free trials due to extraordinary demand and trial abuse. Link.

• The service gained attention for generating convincing images of Pope Francis in fashionable clothing and other controversial images.

• Midjourney's V5 update enhanced its output, making images appear more coherent and lifelike.

• Despite concerns over generated misinformation, CEO David Holz claims profit was the main motivator for ending free trials.

• Holz says the service experienced issues with throwaway accounts and a temporary graphics card shortage.

• Midjourney is working on improving its moderation policies and has new systems planned for release soon.

Emad Mostaque, CEO and founder of Stability AI, hinted at plans for an IPO in the next few years. Link.

• Mostaque dismissed the possibility of Stability AI being acquired.

• Stability AI gained attention for developing Stable Diffusion, Dance Diffusion, and open source music.

• Mostaque signed the open letter calling for a six-month pause and increased regulation in the AI space.

• Reports suggest Stability AI is seeking funding at a $4 billion valuation, up from a previous $1 billion.

• Mostaque believes AI is not a bubble and encourages founders to build good products and solve problems.

Deep Render has raised $9 million in a Series A funding round led by IP Group and Pentech Ventures for its AI-powered video compression technology. Link.

• The startup aims to resolve bandwidth limitations by using AI-based compression to replace traditional methods.

• Deep Render's AI compression algorithm was trained on a dataset of over 10 million video sequences.

• The company claims its algorithm is up to five times "better" than the industry standard codec, HVEC.

• Better compression allows streaming platforms to increase quality without sacrificing bandwidth-related aspects.

• Deep Render has raised a total of $15 million and plans to take on venture debt soon.

CODE/DEVTOOLS

Replit partners with Google Cloud, providing developers with access to Google Cloud infrastructure, services, and foundation models via Replit's Ghostwriter. Link.

• The collaboration aims to accelerate the creation of generative AI applications and expand Replit's mission of empowering 1 billion software creators.

• Replit's Ghostwriter coding AI generates over 30% of the code for developers using the platform.

• Replit offers a built-in AI, instant debugging capabilities, monetization, a strong community, AI on mobile, and collaborative coding features.

• The partnership allows developers using Replit to have global access to Google Cloud infrastructure and Google's foundation models on Vertex AI.

• Replit's Ghostwriter was developed and launched last fall, with a conversational AI interface for generating and proactive debugging added in January 2023.

Fetch.ai has raised $40 million in funding from DWF Labs for its AI tooling and platform. Link.

• The startup focuses on autonomous agents, network infrastructure, and decentralized machine learning.

• Fetch.ai aims to enable communication and actions between AI applications for more actionable results.

• The platform uses blockchain technology and has its own FET token for transactions.

• CEO Humayun Sheikh believes that Fetch.ai can provide a more equitable and traceable approach to AI.

• The company plans to launch commercial services later this year, including turning AI-generated results into transactions.

ROBOTICS

Bill Gates predicts autonomous vehicles (AVs) will reach a tipping point within the next decade. Link.

• Gates believes AVs will change transportation as dramatically as PCs changed office work.

• He recently had a positive experience riding in a Wayve-developed autonomous vehicle in London.

• No personal vehicles are currently fully self-driving, but some companies offer autonomous taxi rides in select cities.

• Gates expects personal vehicles to gain self-driving capabilities after long-haul trucks, deliveries, taxis, and rental cars.

• He believes AVs will eventually save time and help the elderly and people with disabilities, but widespread adoption is still "decades away."

POLICY/LAW/ETHICS

Elon Musk and prominent AI researchers sign an open letter urging AI labs to pause development of large-scale AI systems for at least 6 months. Link.

• The letter, published by the Future of Life Institute, expresses concerns about the profound risks that uncontrollable AI systems pose to society and humanity.

• Signatories include Yuval Noah Harari, Steve Wozniak, Jaan Tallinn, Andrew Yang, and AI researchers such as Stuart Russell and Yoshua Bengio.

• The letter calls for a public, verifiable pause and urges governments to institute a moratorium if necessary.

• The letter highlights growing opposition to the current "ship it now and fix it later" approach in AI research, which could potentially influence political considerations.

• The signatories suggest using the pause to develop shared safety protocols for AI design and development, overseen by independent experts.

Italy bans ChatGPT over privacy concerns and launches an investigation into the AI service. Link.

• Italian authorities claim ChatGPT lacks a proper legal basis for collecting personal user information.

• OpenAI is accused of failing to verify users' ages and enforce rules for those over 13, potentially exposing them to unsuitable content.

• Italy's decision follows calls from experts to halt AI development over potential dangers.

• European consumer organization BEUC urges EU and national authorities to investigate ChatGPT and similar chatbots for potential privacy issues.

• OpenAI has 20 days to respond to regulators' concerns; the company has not yet commented on the block.

The Center for AI and Digital Policy (CAIDP) filed a complaint with the Federal Trade Commission (FTC) against OpenAI. Link.

• The complaint argues that GPT-4, used to power ChatGPT, violates FTC rules on deception and unfairness.

• CAIDP's complaint calls GPT-4 biased, deceptive, and a risk to privacy and public safety, stating it makes unproven claims and lacks testing.

• The complaint quotes OpenAI's own reports on the potential risks of AI, such as disinformation, influence operations, and weapon proliferation.

• CAIDP criticizes OpenAI for not conducting safety checks to protect children during GPT-4's testing period.

• The complaint implies support for greater government regulation of AI, as seen in European regulators' rules-based approach to the technology.

OTHER

Microsoft unveils Security Copilot in limited preview, aiming to offer "end-to-end defense at machine speed and scale." Link.

• Powered by OpenAI's GPT-4 and Microsoft's security-specific model, it aids cybersecurity analysts in threat response and risk assessment.

• Security Copilot collates data from products like Microsoft Sentinel, Defender, and Intune to provide insights and help identify vulnerabilities, ongoing attacks, and remediation steps.

• It can analyze suspicious user logins, create PowerPoint presentations outlining incidents, and accept files, URLs, and code snippets for analysis.

• The security-specific model is privacy-compliant and informed by more than 65 trillion daily signals.

• Microsoft's recent AI push includes integrating generative AI features into Bing, Edge browser, GitHub, LinkedIn, Skype, and Microsoft 365 Copilot.

Hugging Face reported substantial performance gains using Intel's hardware for training and running machine learning models. Link.

• Nvidia's A100 GPU faces competition from Intel's latest data center technologies, including Intel Xeon 4th generation Sapphire Rapids CPU and Habana Gaudi2 GPU.

• Intel Habana Gaudi2 was found to run inference 20% faster on the BLOOMZ model than the Nvidia A100-80G.

• The latest 4th Generation Intel Xeon CPU showed increased performance compared to the prior 3rd generation version.

• Intel's strategy for generative AI involves both training and inference across various environments, including enterprises, edge, cloud, and telcos.

• Intel is in a "coopetition" with Nvidia, and more performance metrics and hardware, such as the Habana Guadi3 GPU accelerator, are expected to be released soon.

Bloomberg unveiled BloombergGPT, an AI model focused on natural language processing tasks within the financial industry. Link.

• BloombergGPT is trained on massive financial data to assist with risk assessments, financial sentiment analysis, and potentially automating accounting and auditing tasks.

• The model utilizes data from the Bloomberg Terminal, which provides real-time market data, news, and analytics.

• The AI was trained on a 363 billion token dataset consisting of English financial documents and general-purpose datasets.

• BloombergGPT is a 50-billion parameter decoder-only AI program.

• The model aims to improve existing NLP workflows and discover new ways to benefit customers in the financial sector.

DataDome, a bot protection service provider, raised $42 million in Series C funding led by InfraVia Growth with participation from Elephant and ISAI. Link.

• Imperva survey reveals 42.3% of internet traffic in 2021 was from bots with malicious intent.

• Cybersecurity startup funding decreased from $22.8 billion in 2021 to $15.3 billion in 2022.

• DataDome plans to use the funds for commercial rollout and R&D efforts.

• DataDome analyzes traffic requests in real-time to prevent bot-based threats.

• The company has over 300 customers, protecting 400,000 domains worldwide.

Oscilar, a fintech company that focuses on AI-driven fraud and theft protection for online transactions, has emerged out of stealth. Link.

• Co-founders Neha Narkhede and Sachin Kulkarni self-funded Oscilar with $20 million.

• The company has been working under the radar for two years before its public launch.

• Oscilar's AI relies less on first- and third-party data for training machine learning models.

• The platform offers a toolkit for risk operations teams to create, test, and deploy models.

• Despite competition and recent funding challenges, Oscilar is already working with dozens of fintech customers.

Jigso, an early-stage startup creating an AI assistant for business observability, has raised $7.5 million in seed funding led by General Catalyst, Entrée Capital, and Jibe Ventures. Link.

• Jigso's Sidekick Slack bot allows users to access information from various business apps without leaving Slack.

• The product currently integrates with Salesforce, Zendesk, Jira, Excel, ClickUp, Google Sheets, Gmail, and Calendar.

• Founded in late 2021, Jigso is working with around 50 companies and has 10 employees.

• Co-founders Or Shani and Tomer Naveh aim to build a diverse company as it grows.

German-based Parloa has raised €20 million (~$21.67 million) in Series A funding led by EQT Ventures, with participation from Newion and Senovo to support its contact center automation software. Link.

• Over $400 billion spent annually on customer contact centers worldwide, with increasing adoption of AI and automation.

• Contact center automation startups, including Invoca, Replicant, PolyAI, and Observe.ai, have raised significant funding recently.

• Funding to be used for customer acquisition, opening a U.S. office, and product R&D.

• Parloa combines conversational AI tech and low-code tools to assist contact center employees.

• The platform connects various apps and services to enable automation and improve customer service experiences.

GPT Tip of the Week

Prompt: What are the best ways to spot BS when talking to founders building in generative AI?

Response: When talking to founders building in generative AI, it can be difficult to separate hype from reality. Here are some strategies that may help you spot BS:

1. Ask for details: If a founder is making grand claims about their generative AI technology, ask for specifics about how it works, what datasets it uses, and what kind of results it has produced so far. If they can't provide detailed answers, it may be a red flag.

2. Look for evidence: Ask for evidence of the technology's effectiveness, such as case studies, demos, or customer testimonials. If they can't provide any evidence, or the evidence seems flimsy, it may be a sign that their claims are exaggerated.

3. Check their background: Look into the founder's background and experience in the field of generative AI. If they don't have a strong background or have a history of overhyping their technology, it may be a warning sign.

4. Beware of buzzwords: Generative AI is a hot field, and some founders may use buzzwords and jargon to make their technology sound more impressive than it really is. Be wary of buzzwords and ask for clear explanations of what they mean.

5. Be skeptical of "black box" solutions: If a generative AI solution is described as a "black box" that produces amazing results without any explanation of how it works, it may be a warning sign. It's important to understand how the technology works and what data it's using to produce its results.

Overall, it's important to approach any claims about generative AI technology with a healthy dose of skepticism and ask for evidence and details before making any decisions.