📝 Editorial 

The quest to achieve artificial general intelligence (AGI) is one of the most fascinating endeavors in the entire technology industry and one that is making rapid progress. Many experts believe that we might be just two or three technological breakthroughs away from the first forms of AGI. Models like GPT-3, AlphaFold, and DALL-E clearly exhibit initial signs of intelligence that resembles human cognition. Despite the progress, the quest for AGI is still full of existential challenges, such as the alignment with human values and intent. If we can’t guarantee that AGI systems are aligned with human values and do what humans want, we might be creating systems that pose fundamental risks to humanity.  

The idea of AGI alignment with human values is well understood but the formal frameworks to ensure such a goal are still in very early stages. Last week, OpenAI unveiled some efforts on AGI-human alignment research following an empirical approach. In general, OpenAI’s methodology follows three key principles:  

1. Training AI systems using human feedback: Start by training systems to follow human goals.  

2. Training AI systems to assist human evaluation: As AI systems grow, it can assist humans in the evaluation of AI agents.  

3. Training AI systems to do alignment research: Past a certain scale, AI agents will be required to ensure that other AI agents align with human values.  

Following the three previous steps, we can see a progression of AGI alignment research that starts with humans and continues with AI. The last point basically states that, at large scale, AI will be more efficient than humans at finding alignment with human values. That’s quite a statement! OpenAI alignment research is not without limitations but its certainly one of the most complete works tackling one of the existential challenges for our path towards AGI.   

🔺🔻TheSequence Scope – our Sunday edition with the industry’s development overview – is free. To receive high-quality content about the most relevant developments in the ML world every Tuesday and Thursday, please subscribe to TheSequence Edge 🔺🔻

🗓 Next week in TheSequence Edge:

Edge#221: we explain what Diffusion Models are; discuss Imagen, Google’s massive diffusion model for photorealistic text-to-image generation; explore MindEye that allows you to run multiple generative art models in a single interface.

Edge#222: we deep dive into Axion, the feature store architecture powering ML pipelines at Netflix.

Now, let’s review the most important developments in the AI industry this week

🔎 ML Research

AGI Alignment  

OpenAI discussed the main frameworks for ensuring human alignment in future AGI systems →read more

ML and Google Meet Backgrounds 

Google Research published a paper presenting a video segmentation technique used to improve background images in Google Meet →read more

Galaxies and Graph Neural Networks 

Carnegie Mellon University published a paper discussing a graph neural network and GAN-based technique used to predict the orientation of galaxies in the universe →read more

Time-Series Transformers 

Salesforce Research published a paper detailing ETSformer, an exponential smoothing transformer technique for time-series forecasting →read more

💎 We recommend: Out-of-the-box vector search experience, designed for AI, delivered on Zilliz Cloud*

Vector embedding is the most powerful technique that can encode almost any kind of data into vectors. This technique is widely adopted in various AI-driven applications, including recommendation engines, reverse image search, video similarity search, intelligent QA chatbots, fraud detection, data deduplication, molecular structure analysis, and more. However, traditional databases are not designed to handle vector embeddings effectively, necessitating a purpose-built data infrastructure to manage and process them at scale. Enter vector database.

Zilliz Cloud is a fully-managed cloud vector database built atop Milvus, the most popular open-source vector database that was also created at Zilliz. Milvus is widely recognized for its high scalability, ease of use, and high performance. Zilliz Cloud further simplifies the process of deploying and scaling vector search applications by eliminating the need to create and maintain complex data infrastructure.

Zilliz Cloud is currently in private preview for early access to select customers. Find out if you’re eligible today!

APPLY NOW

*We thank Zilliz for their ongoing support of TheSequence.

🤖 Cool AI Tech Releases

MoCapAct 

Microsoft Research released MoCapAct, a dataset for different tasks in humanoid robotic movements. One of the demos teaches a robot to dance like Mick Jagger! →read more 

🛠 Real World ML 

Multi-Task Learning at LinkedIn 

LinkedIn provides some insights about the architecture used to enable multi-task learning models for different scenarios →read more

Real-Time Analytics and Uber Freight 

Uber discusses the architecture powering real-time analytics for the Uber Freight solution →read more

Time Constrained Recommendations at Netflix 

Netflix presents a reinforcement learning approach for recommendation systems that operate on a time-constrained budget →read more

💸 Money in AI

• Ray’s developer Anyscale raised $99 million in Series C funding co-led by existing investors Addition and Intel Capital. Hiring in San Francisco/US.

• Vector database company Zilliz raised a $60 million series B extension to expand its operations in Silicon Valley. Hiring in San Francisco, CA (US) and remote.

• Cross-game avatar platform Ready Player Me raised $56 million Series B round led by a16z. Hiring globally.

• Virtual character developer platform Inworld AI raised a $50 million Series A round. Hiring in the US and Canada.

• “Retail AI” startup Lily AI raised $25 million in a Series B financing round. Hiring remote.

• Food supply chains solution Lumachain raised $19.5 million in Series A funding, led by Bessemer Venture Partners.

• “Cultural AI” startup Qloo raised $15 million in a Series B funding from Eldridge and AXA Venture Partners. Hiring in New York/NY.

• Meeting AI startup Headroom raised a $9 million investment round led by Equal Opportunity Ventures. Hiring in San Francisco/US or remote.