📝 Editorial 

Is the first day of a new year and with that all sorts of new hopes and expectations for AI. For the last few days, I am sure you have been inundated with predictions about AI in the new year so let’s try to avoid those. After all, like the great Neils Bohr always said: “Prediction is very difficult, especially if it's about the future.” 😉 . Instead, lets discuss a single prediction that I have reflecting upon in the last few weeks and it is related to a imminent shift in the value of AI technologies from infrastructure to applications.

The bulk of the value creation in every technology market fluctuates between infrastructure and applications. Mobile computing started as an application first market and it went through cycles that produced many mobile infrastructure startups. Cloud computing started as an infrastructure trend and the value steadily transition into application layers and then back to infrastructure. For the last decade, the bulk of the value in AI has been accrued at the infrastructure layer with movement such as MLOps but 2023 might be the first year of the shift towards application-first value.

The X factor: generative AI.

Generative AI has been the first AI trend that has bring incredibly powerful super models to mainstream developers. As a result, the generative AI movement is going to trigger a new generation of startups that leverage large foundation models in consumer and enterprise use cases. Many of those use cases are transitioning from being an AI problem to a product problem. All signs indicated that 2023 is going to see an explosion in generative AI mega models and, with that, we will see a tremendous value creation race to apply those models to real world scenarios. In that sense, 2023 is going to be the first year in which AI applications might capture more value than infrastructure providers. The opportunities are massive as we might be entering the industrial era of AI.

Happy New Year and thanks for the continued support!

🗓 Next week in TheSequence Edge:

Edge #257: Our series about ML interpretability continues with an overview of local model-agnostic interpretability methods, IBM’s ProfWeight research that combines interpretability and accuracy in a single model and the InterpretML framework.

Edge #258: Deep dives into OpenAI’s new Point-E text-to-3D model.

🔎 ML Research

LLM Benchmark

Meta AI published a paper detailing OPT-IML, a benchmark for over 2000 tasks for language pretrained models —> Read more.

Zero-Shot Logo Recognition

Amazon Science published a research paper detailing a zero-shot learning technique for logo recognition —> Read more.

High Quality Text Summarization

Meta AI published a set of research papers that detailed techniques, benchmarks, training and fine tuning of high quality text summarization models —> Read more.

New DP Algorithm

Google Brain published a new research paper discussing a new differential privacy algorithm —> Read more.

🤖 Cool AI Tech Releases

PALM+RLHF

An open source version of a method that combines the PALM language model with a reinforcement learning with human feedback(RLHF) is now available and shows a similar architecture to ChatGPT —> Read more.

QuickVid

QuickVid is a new generative AI platform to create short form videos —> Read more.

🛠 Real World ML 

Mode Serving at Meta

Meta AI discusses a TorchServe-based architecture and optimizations used to serve models in the  Animated Drawings app —>Read more.

💸 Money in AI

• Virtual meeting data intelligence startup Recall.ai raised a $2.7 million seed round.

• Data quality startup Qualytics announced a $2.5 million seed round.

• AI-Driven Microbiome-Based Therapeutics Biomica raised $20 million series A.