| |
| Welcome to Inside AI! Here's a look at today's headlines: - AI use is on the rise in online influence campaigns.
- Arthur AI studied top LLMs in areas like math, hallucinations, and U.S. presidents.
- Prominent ex-Google researchers founded Sakana AI.
If you find this newsletter useful, please consider sharing it with your colleagues and friends. We appreciate your support. Beth
p/beth-duckett | |
| 1 | | Mandiant, a Google-owned cybersecurity company, has found "numerous" instances of AI-generated content in political online influence campaigns over the past four years. The campaigns were led by groups linked to countries like Russia, China, Iran, and others. However, AI has not played a key role in any threats, and Mandiant expects AI use in digital intrusions to be limited in the short term. More: - According to Mandiant, the rise of generative AI can allow resource-constrained groups to create higher-quality content for large-scale influence campaigns.
- One example is Dragonbridge, the pro-China information campaign that targeted pro-democracy Hong Kong protesters.
- Since 2019, Dragonbridge's growth has surged across 30 social platforms and 10 languages, according to Mandiant VP Sandra Joyce.
- Still, the effects of these campaigns have been limited. "They really haven't changed the course of the threat landscape just yet," Joyce told Reuters.
|     | |
| 2 | | Arthur, a platform for monitoring machine learning models, has released new research gauging how top large language models perform in areas like mathematics, so-called "hedging," and their knowledge of U.S. presidents. What the numbers say: According to Arthur, OpenAI's GPT-4 performed best on questions involving combinatorial (counting) mathematics and probability, followed by Anthropic's Claude 2. Cohere's model performed the worst in math, with zero correct answers and 18 hallucinations, which occur when models generate inaccurate or nonsensical information. 
With U.S. presidents, Claude 2 got the most answers right, with GPT-4 in second place. Overall, the models produced over 120 hallucinations but didn't avoid any questions. 
In Arthur's other experiment, the researchers examined how often the AI models respond with "hedging" answers, where their responses contain warning phrases to avoid generating offensive or false content (like "unfortunately, I cannot answer that question"). GPT-4 showed a 50% relative jump in hedging answers over GPT-3.5, which backs up some user experiences about GPT-4 being more "frustrating" to use than its predecessor, according to Arthur. Notably, all of Cohere's responses lacked any hedging language. 
The takeaway: GPT-4 outperformed all models tested, with fewer hallucinations than GPT-3.5, particularly in math questions where it reduced hallucinations by up to 50%. In contrast, Meta's Llama 2 experienced a higher overall rate of hallucinations compared to both GPT-4 and Anthropic's Claude 2. Adam Wenchel, co-founder and CEO of Arthur, told CNBC that the key for LLM users is to make sure they "understand how it performs for what you're trying to accomplish." Question for readers: Have you encountered an AI model showing hallucination-like behavior? Feel free to share your experience here. | | | |
| A message from our sponsor, Interview Kickstart. |  | | Is your career safe from AI disruption? AI is transforming the jobs market. Engineers who can build, maintain, and fine-tune AI systems are in exceptionally high demand. Companies like OpenAI, Hugging Face, Google, & Amazon are hiring aggressively and offering top compensation, in many cases upwards of $300,000-$500,000. Don’t get left behind. Make the switch today to an exciting ML career with Interview Kickstart. - Learn live from FAANG+ AI/ML engineers
- Capstone project for hands-on experience
- Up to 15 mock interviews with FAANG+ AI/ML engineers
- Individual coaching and 1:1 help
The results speak for themselves: Alumni who consistently bag > $300K job offers. Highest compensation received: $1.28 million. This could be you! Register for the free webinar to learn more. Register Now | |
|
| 3 | | Two former Google researchers, David Ha and Llion Jones, have founded Sakana AI, a Tokyo-based AI startup aiming to build AI models based on evolutionary computation. Sakana's research is inspired by nature's principles like evolution and collective intelligence. More: - Sakana AI plans to build a generative AI model, or nature-inspired neural network, that's more adaptable than today's models.
- The startup aims to create new architectures for foundational AI models rather than just scaling up existing ones.
- Instead of building a single large model, the team could use many smaller models that communicate and work together to solve problems, according to Ha.
- The duo chose Tokyo for the company's HQ due to its technical infrastructure, educated workforce, and appeal to foreign talent.
About the founders: - Ha, Sakana's CEO, is the former head of research at Stability AI and a former Google research scientist.
- Jones, who is CTO, was the fifth co-author of the landmark 2017 "Attention Is All You Need" paper, which introduced the Transformer deep learning architecture that paved the way for chatbots like GPT-4 and other generative AI models.
- With Jones' exit, all of the authors have now left Google. Among their new ventures is Character.AI, led by Noam Shazeer, and Cohere, co-founded by Aidan Gomez. Ashish Vaswani and Niki Parmar founded Essential AI.
| | | |
| 4 | | AI-powered data engineering solutions firm Databricks is reportedly in talks with investors to raise "hundreds of millions of dollars," per SiliconAngle. Databricks intends to use the fresh capital to further develop AI applications. More: - Talks of the new funding come as the firm recently crossed the $1B revenue mark in the most recent fiscal year, buoyed by a 70% YoY growth.
- The firm reported a $380M operating loss in the same period.
- Despite the losses, the firm is close to breaking even.
- The San Francisco-based startup has raised $3.5B in total funding to date, per Crunchbase.
- The most recent funding round for the firm was a $1.6B Series H round in August 2021 led by Counterpoint Global.
This story first appeared in today's issue of Inside Venture Capital. Subscribe to receive the newsletter here. | | | |
| A message from our sponsor, MindsDB. |  | | Upskill your developer team with AI MindsDB’s game-changing AI database empowers world’s developers to build, manage and automate Generative AI models for data-intensive applications. MindsDB offers three primary benefits to its users: - First, it makes the creation of the next wave of AI-centered applications readily available and accessible for all developers, by programming and managing AI models through an “enhanced SQL” abstraction.
- The second, MindsDB brings the AI development process much closer to the data, streamlining production of generative AI while bridging the gap between AI and the data needed to power its potential.
- Lastly, MindsDB removes the complexity or need for managing multiple AI frameworks - in a single platform, organizations can run and automate multiple AI frameworks through the platform’s hundreds of integrations with databases.
Learn more | |
|
| |
| 6 | | QUICK HITS - The AWS Cookbook — packed with practical insights and hands-on examples — is a must-have for your cloud toolkit.*
- Microsoft published an AI-generated travel piece about Ottawa, Canada, which notably suggests tourists visit the Ottawa Food Bank as the third recommendation on the list.
- Musician Hozier said he is open to participating in a strike against AI within the music industry, saying that AI "can't create something based on a human experience. So I don't know if it meets the definition of art."
- South Korean game developer NCSoft announced four new AI large language models. Collectively named VARCO, the models aim to streamline game development and could help the company create digital humans and other conversational and generative AI platforms.
- IPOs are a rare chance to invest in VC-backed startups, but this AI company opened its round to retail investors. Invest and learn more about Jurny.*
*This is a sponsored post. | | | |
| Upcoming Events | | OCT
16 | | Meet Our Fund 5, Building and Investing in AI | | | | | | * This is a sponsored event | | | |
Term of the Day Internal rate of return: VC firms use internal rate of return (IRR) to analyze a fund’s performance and give the annualized percentage returns earned by the fund. Question of the Week Do you share details about your salary with your coworkers? Join the conversation |
INSIDE AI LEADERBOARD (30 DAYS) |
 | AI and technology writer | | Beth is the curator and writer of the AI and Tech newsletters at Inside. With a career in journalism and editing that began in 2007, she has written for publications including USA Today, The Arizona Republic, Arizona Business Gazette, and more. You can reach her at Beth.Duckett@yahoo.com. | | This newsletter was edited by Aaron Crutchfield | |
| |
 | Don’t get left behind. Make the switch today to an exciting ML career with Interview Kickstart. | |
|
| |
|
| |
Comments
Post a Comment