GSV's AI News & Updates (08/04/24)
Meta Llama 3.1, OpenAI SearchGPT, Google Gemini 1.5 Pro, AI Friend, Uplimit Series A, Techwolf Series B, Heeyo, EU AI Act, Management Consultants in the AI Boom
General 🚀
Meta releases the biggest and best open-source AI model yet: Llama 3.1 outperforms OpenAI and other rivals on certain benchmarks. Mark Zuckerberg expects Meta’s AI assistant to surpass ChatGPT’s usage in the coming months.
Llama 3.1 405B contains 405B parameters, making it the largest open-source AI model in recent years. It was trained using 16,000 Nvidia H100 GPUs.
Meta claims Llama 3.1 outperforms GPT-4o and Anthropic's Claude 3.5 Sonnet in various performance metrics.
Unlike closed-source models, Llama 3.1's weights are available for download, allowing developers to customize and fine-tune the model. The model has been downloaded over 300M times across all Llama versions.
Open Source AI Is the Path Forward: Mark Zuckerberg outlines the company's commitment to open source AI, drawing parallels between the development of AI and the historical evolution of Unix to Linux.
Zuckerberg says Meta will need 10x more computing power to train Llama 4 than Llama 3
Meta's new AI Studio helps you create your own custom AI chatbots: AI Studio will let you build a chatbot that acts as a virtual extension of yourself.
OpenAI announces SearchGPT, its AI-powered search engine: This new tool marks OpenAI's entry into the search market, potentially challenging AI-powered search competitors like Perplexity and Google's long-standing dominance. The company emphasized that SearchGPT was developed in collaboration with various news partners (The Wall Street Journal, The Associated Press, Vox Media). Publishers will have a way to “manage how they appear in OpenAI search features”.
Microsoft says OpenAI is now a competitor in AI and search: The change comes days after OpenAI announced a prototype of a search engine and marks a significant shift in their long-standing partnership.
Google's new Gemini 1.5 Pro model is available to test and topped AI leaderboards: It’s ranked #1 at solving math problems, #1-2 in instruction-following, and in the top 5 at handling tricky tasks in English and coding.
Why OpenAI could lose $5B this year: The potential $5B loss is a significant increase from the previous year's reported loss of $540M. This rapid escalation in expenses raises questions about OpenAI's path to profitability.
AI-Powered Necklace Will Be Your Friend for $99: Friend is a pendant that will talk to you via text when you're lonely or any time you want conversation.
OpenAI has released a new ChatGPT bot that you can talk to: This new feature, powered by the GPT-4o model, allows users to engage in real-time, natural conversations with the bot, which can interpret different tones of voice and respond appropriately.
On speaking to AI: Ethan Mollick compares the new voice capabilities of ChatGPT and Siri, highlighting the fundamental differences in their design philosophies.
Education and the Future of Work 📚
Uplimit, the AI-Powered Enterprise Learning Platform, Announces Series A Funding Round Led by Salesforce Ventures: Uplimit has developed a unified enterprise learning solution that leverages AI to enhance cohort management, learner support, and content authoring.
TechWolf raises $43M Series B to take an AI-sized bite out of the internal recruiting game: TechWolf's AI engine analyzes data from internal workflows to understand employees' skills and interests. This information helps managers and internal recruiters make better decisions regarding project assignments, training, and salary management.
YC alum Fluently’s AI-powered English coach attracts $2M seed round: Key features include real-time call analysis, AI practice sessions, personalized feedback and privacy (local audio storage, encryption, and anonymized data).
Heeyo built an AI chatbot to be a billion kids’ interactive tutor and friend: The startup offers children between the ages of 3-11 an AI chatbot and over 2,000 interactive games and activities, including books, trivia and role-playing adventures.
Pear Deck’s New ‘Instant Pear Decks’ Feature: Instant Pear Decks save teachers time by leveraging generative AI technology to easily and quickly create customized lessons
Pluralsight Launches AI Assistant Iris to Accelerate and Personalize the Learning Experience
AI Is a Services Revolution: "AI disruption is happening, but slowly. It may seem like the race is over, as we see copilots emerge for every industry under the sun. But it’ll take time for the true winners to emerge, and for metrics to catch up to the hype."
AI will never substitute teachers. This failed experiment shows why: AI attempts to scale education fly in the face of all we know about learning—that learning is social, and that the more human attention, the better the outcomes.
The AI job interviewer will see you now: AI interview services say they’re eliminating bias — but not everyone agrees.
Management consultants find sweet spot in the AI boom: In a gold rush, you want to be the person who sells shovels. That’s not a new observation and helps explain why Nvidia has come from nowhere to a $2.6T market cap. The next best job, it seems, is telling everyone how the shovels should be used. Enter the consultancy industry — an early beneficiary of the AI revolution.
AI is both the problem and its own solution: An irrational fear of running afoul of AI laws is paralyzing decision-making and preventing more companies from tapping the tech’s true potential.
Instructure to Be Acquired by KKR for $4.8B
GSV’s breakdown of the recent take-privates of PowerSchool (June 2024) and Instructure (July 2024)
Tech 💻
AI achieves silver-medal standard solving International Mathematical Olympiad problems: This accomplishment marks the first time an AI has reached this level of performance in such a prestigious mathematical competition. Solution times varied significantly, with one problem solved within minutes (19 seconds for the geometry question), while others took up to three days.
AI models collapse when trained on recursively generated data: AI models can experience "model collapse" when trained on recursively generated data, leading to a loss of information about the original data distribution and compounding errors over generations. This phenomenon results in narrowed distributions, reduced variance, and highlights the importance of using genuine human-produced data to maintain model performance and mitigate data pollution.
Regulation and Policy ⚖️
White House says no need to restrict ‘open-source’ artificial intelligence — at least for now
The EU’s AI Act is now in force: This starts the clock on a series of staggered compliance deadlines that the law will apply to different types of AI developers and applications.
AI Video Generator Runway Trained on Thousands of YouTube Videos Without Permission: A leaked document obtained by 404 Media shows company-wide effort collected thousands of YouTube videos and pirated content for training data.
Websites are Blocking the Wrong AI Scrapers (Because AI Companies Keep Making New Ones): Despite efforts to block AI crawlers, many websites are mistakenly blocking outdated bots while leaving the actual active scrapers unchecked, highlighting the chaos and confusion in the current robots.txt landscape.
Argentina will use AI to ‘predict future crimes’ but experts worry for citizens’ rights
Who will control the future of AI?: Sam Altman presents a strategic imperative for the US and allies to lead a global AI initiative based on democratic values that suggests robust security, infrastructure investment, commercial diplomacy, and collaborative norms development to maintain AI supremacy over authoritarian regimes.