AI Brief #19: the IRS now uses AI to catch tax evasion

Today is September 13, 2023.

The last two weeks have seen quite a few interesting developments in the AI world. We’re starting to see a global race on the AI model front, with Chinese ChatGPT competitors finally in public release.

And in the US, the historically stodgy IRS has revealed they’re already adopting cutting edge AI to investigate tax fraud.

Meanwhile, Meta’s AI team continues to open source more interesting AI tech, this time for video-tracking.

In this issue:

• 🪝 IRS deploys AI to catch tax evaders

• 🏃‍♂️ Meta wants their next chatbot model to rival OpenAI’s GPT-4

• 🇨🇳 China’s ChatGPT rivals make public debut

• 🍎 Duet for Google Workspace launches for $30 per month

• 🔎 Call of Duty polices gamers with AI, and other quick scoops

• 🧪 The latest science experiments, including Meta’s new point-tracking model (with code and demo)

Did someone forward this to you? Subscribe here!

🪝 IRS deploys AI to catch tax evaders

The IRS is using cutting-edge AI to assist its investigations into complex cases, reports the New York Times.

With its new AI tools, the IRS is specifically focusing on hedge funds, private equity groups, real estate investors and law firms – complex organizations where IRS teams have in the past found themselves under-resourced to police.

Why this matters: AI is increasingly finding a fit in serving as a co-pilot to smart humans. The ability of artificial intelligence to sort through large data sets, spot trends, and identify patterns is highly valuable in complex projects ranging from medical to legal and now finance work.

With just 1% of large, complex partnerships currently audited by the IRS, AI tooling could significantly improve the coverage of these organizations in the years to come.

🏃‍♂️ Meta wants their next chatbot model to rival OpenAI’s GPT-4

Meta’s next generation language model is intended to rival GPT-4’s abilities, the Wall Street Journal reports. After releasing Llama 2, Meta intends to further accelerate their AI progress by training a model several times more powerful.

Training won’t start until early 2024, sources familiar with Meta’s plans revealed, as the company is still building up the required data center capacity to develop this model.

Why this matters: sources revealed that Meta likely intends to open-source this model as well, which will continue to put pressure on closed-source AI model vendors OpenAI and Google. But the release of an even more powerful model is also likely to raise further questions from critics about the ethics of sharing such powerful technology so freely.

🇨🇳 China’s ChatGPT rivals make public debut

Chinese search engine and AI company Baidu officially launched their ChatGPT competitor model last week, marking an important milestone in China’s own race to dominate the artificial intelligence landscape. Two other companies, Baichuan and Zhipu AI, also launched AI models the same day.

Dubbed “Ernie,” Baidu’s app quickly topped the iOS App Store charts in China upon release. The model will now benefit from wide-scale human use to further improve its outputs.

Why this matters: China has heavily stressed its goal of overtaking the United States in AI advancements, and the generative AI space is an important sector to showcase these gains. At the same time, US restrictions of advanced AI chips to China have introduced geopolitics into this global tech race, as Chinese companies seek to find workarounds and stay competitive in their AI progress.

These same Chinese AI companies are also navigating a thorny regulatory environment for AI models, as they must pass review by various government bodies before they launch their services.

🍎 Duet for Google Workspace launches for $30 per month

Duet is an all-in-one personal assistant that seamlessly integrates with the Google suite of products (Docs, Spreadsheets, and more), and is designed to enable busy professionals to have a strong AI partner. Imagine how “a last-minute request that once called for an all-nighter, can now be completed before dinner time,” Google’s own marketing buzz touts.

Why this matters: AI models are expensive to run, and Microsoft already announced their 365 Copilot at a cost of $30 per month. Google’s competitive product will help retain customers on their work software suite while also showing off their AI prowess as Microsoft touts OpenAI’s foundation models. And successful wide-scale adoption could represent billions in revenue for Google as well.

Google is currently offering a free trial to any existing Workspace users curious about Duet’s capabilities.

🔎 Quick Scoops

Call of Duty is using AI to identify toxic behavior in real-time voice chats. The AI helps flag instances of toxic behavior for a human moderation team. (PC Gamer)

OpenAI is on pace to generate over $1 billion per year, as demand for artificial intelligence stays strong. The company generated just $28M in 2022. (The Information)

AI startups are booming as the latest batch from Y Combinator features 160 AI companies. The entire class has just 229 companies in total. (Democratizing Finance)

US rejects copyright protection for an award-winning AI-generated art piece, despite the fact that the text prompt was revised 6254 times and the resulting image was then edited in Photoshop. (Ars Technica)

Elon Musk readies X to train its AI on user data, as changed terms of service for X users recently reveal. (Decoder)

🧪 Science Experiments

Meta releases CoTracker, a model for point tracking in videos

• The AI model understands how various points in a video can shift, potentially saving hours of time for humans who normally do this by hand.

• You can try it out for yourself here by uploading your own video.

VideoGen: A Reference-Guided Latent Diffusion Approach for High Definition Text-to-Video Generation

• Baidu’s AI team releases their video generation model, which touts impressive improvements in fidelity and composition over rival models.

• See their project page here.

Blended-NeRF: Zero-Shot Object Generation and Blending in Existing Neural Radiance Fields

• Editing a local region or a specific object in a 3D scene represented by a NeRF is challenging, mainly due to the implicit nature of the scene representation. Blended-NeRF enables editing a specific region of interest in an existing NeRF scene, opening up a realm of possibilities.

• See their project page here.

👋 How I can help

Here’s other ways you can work together with me:

• If you’re an employer looking to hire tech talent, my search firm (Candidate Labs) helps AI companies hire the best out there. We work on roles ranging from ML engineers to sales leaders, and we’ve worked with leading AI companies like Writer, EvenUp, Tome, Twelve Labs and more to help make critical hires. Book a call here.

• If you would like to sponsor this newsletter, shoot me an email at [email protected] 

As always — have a great week!