New Generative AI Innovations from Google and Salesforce

Sundays, The Sequence Scope brings a summary of the most important research papers, technology releases and VC funding deals in the artificial intelligence space.

🗓 Next week in TheSequence Edge:

Edge #261: Our series about ML interpretability continues with an overview of the famous LIME method, Meta’s research about how interpretable neurons can affect the accuracy of neural networks and the Alibi Explain framework.

Edge #262: Explores NVIDIA’s Get3D generative model.

📝 Editorial 

Generative AI is at the center of the activity in the AI market these days. While the top headlines are captured by high profile models/apps like DALL-E or ChatGPT, there is a tremendous levels of activity in research and applications across the board. This week, we saw some interesting developments with new generative AI models.

Google Brain has been super active in the text-to-image space with the creation of models like Imagen or Parti. This week, Google unveiled Muse, a new text-to-image transformer model that is able to achieve state-of-the-art results while being more efficient than diffusion and autoregressive methods. Muse’s efficiency gains relative to diffusion techniques is based on requiring fewer iterations. Compared to autoregressive models, Muse seems to improve based on its use of parallel decoding techniques.

Salesforce Research has also been active in the text-to-image area. A few days ago, the research lab unveiled a new algorithm that allows text guidance to edit images. Named  Exact Diffusion Inversion via Coupled Transformations (EDICT), the technique can be super relevant for the inpainting and outpainting capabilities of text-to-image foundation models.

Just another week in generative AI!

🔎 ML Research

New Text-to-Image Model

Google introduced Muse, a new text-to-image generation and editing transformer model —> Read more.

Text Guided Diffusion

Salesforce Research published a paper detailing Exact Diffusion Inversion via Coupled Transformations (EDICT), a model to guide text-to-image generation to achieve consistency —> Read more.

Language Models and Disinformation

OpenAI and Georgetown University published a report quantifying the used of large language models in disinformation campaigns —> Read more.

DP for Speech Recognition

Amazon Science published a paper detailing  Private aggregation of teacher ensembles (PATE), a differential privacy technique for speech recognition methods —> Read more.

🤖 Cool AI Tech Releases

PyTorch HTA

PyTorch released  Holistic Trace Analysis (HTA), a framework developed by Meta to debug performance issues in large scale models —> Read more.

TensorFlow for Intel Processors

TensorFlow released a new build optimized for Intel Xeon processors —> Read more.

🛠 Real World ML 

Accelerating PyTorch Training

Engineers from Hugging Face discuss how to accelerate PyTorch training jogs using a cluster of Intel Rapid servers on AWS —> Read more.  

💸 Money in AI

• Microsoft is in talks to invest as much as $10 billion in OpenAI.

• Microsoft announced the acquisition of DPU vendor Fungible.

• Sales intelligence startup attention raised $3.1 million to extend CRMs with natural language interfaces.

• Autonomous Driving platform Oxbotica raised a $140 million series C led by Tencent.

• Generative AI startup Playstream raised $2 million to streamline AI generated live stream content.

• Data labeling powerhouse Scale AI laid off about 20% of its workforce this week.

• OpenAI is in talks to invest in AI chip maker startups Atomic Semi in a $15 million round.

• Inbenta raises $40 million for its AI-based chatbot platform.

• DeepL, a translation as a service startup is actively fundraising at about $1 billion valuation.

• Data intelligence solutions startup Seek AI  announced a $7.5 million seed round.

• AI-based sales platform Coho AI announced a $8.5 million seed round.

• AI-based low code platform Blaze.tech raised $2.5 million pre-seed round.

• Worlds, a platform that uses AI for building industrial solutions, announced a $21 million series A1.  

• DataStax announced the acquisition of machine learning startup Kaskada.

• Hewlett Packard Enterprise announced the acquisition of ML data pipeline startup Pachyderm.  

• Alphabet’s robot software division, Intrinsic, announced 20% layoffs.