Next Week in The Sequence:

• Edge 307: Our popular series about new techniques in foundation models continues with an exploration of program-aided LLMs(PAL). The original PAL paper from Carnegie Mellon University and a review of the NLP Test framework.

• Edge 308: We deep dive into LMQL, a query language for LLMs that explores a new paradigm in language model programming.

📝 Editorial: “The Other” Enterprise Generative AI Platforms

When we think about enterprise generative AI platforms, we immediately gravitate towards tech incumbents such as Microsoft, Amazon or Google or extremely well-financed startups such as OpenAI Anthropic or Cohere. It kind of makes sense as those companies have been dominating the headlines related to generative and we can assume that the data and compute costs required for enterprise generative AI would make this a game of a handful of companies.  However, it would be a mistake to think that the generative AI bonanza will be constrained solely to these platforms.

The enterprise generative AI is certainly heating up and there are several companies with market leadership positions on different sectors that have started making moves to be competitive in this new landscape. Here are some of the ones that I find with very strong potential to be highly competitive with the big guys. The criteria here includes not only a credible technical offering but also strong distribution and financial capabilities.

• Salesforce: The SaaS giant has already incorporated generative AI capabilities across its different products and has been super active in the research space. Just last week, they release XGen-7B, another high performance LLM.

• Databricks: The datalakehouse platform made clear its generative AI intentions with the release of Dolly2 followed by the whopping acquisition of MosaicML for $1.3B.

• Snowflake: Databricks competitor has been building an interesting set of capabilities with acquisitions of <>.

• Anyscale: Don’t sleep on these guys. Anyscale is the the platform behind many of the top ML workloads in the industry for companies like Uber, Microsoft, Intel and, yes, OpenAI. The recently ventured into the LLM space with an impressive set of capabilities.

• Hugging Face: Has all the ingredients to create a very robust enterprise generative AI offering.

• Scale AI: One of the best funded AI platforms in the market has steadily been making inroads in the generative AI space.

• NVIDIA: The GPU giant has serious ambitions in the generative AI software space and a very strong research arm.

There are some other interesting players in the enterprise generative AI space but I think the previous list brings a unique combination of factors. As you can see, these “other” platforms can be extremenly competitive in the enterprise generative AI space. The race is certain on!

💡Webinar: Designing & Scaling FanDuel's Machine Learning Platform

Join us on July 11 at 9 a.m. PT to discover FanDuel's journey in building a powerful ML platform for personalized experiences, learn how they scaled their platform and the lessons learned!

REGISTER NOW

🔎 ML Research

Computationally Universal LLMs

Google Brain researchers published a paper demonstrating that LLMs augmented with read-write memory are computationally universal. This means that they can simulate any algorithm on any input of any size. Pretty remarkable —> Read more.

Grounding LLMs to Images

AI researchers from Carnegie Mellon University published a paper detailing a technique to ground pretrained LLMs in image data. The combination allows the models to generate arbitrarily interleaved image-to-text data —> Read more.

Parsel

Researchers from Stanford University published a paper detailing Parsel, a framework for algorithmic reasoning with LLMs. Parsel takes hierarchical function language descriptions as input and it able to decompose it to reasoning steps —> Read more.

LEDITS

Hugging Face published a research paper detailing LEDITS, a text-guided real time image editing method. LEDITS combines DDPM inversion technique with Semantic Guidance to edit images without requiring modifications in the model architecture —> Read more.

Phi-1

In a paper under the catchy name “Textbooks is All You Need”, Microsoft research unveiled phi-1, a LLM for coding pretrained in high quality datasets from textbooks. Phi-1 outperforms models like StarCoder and Codex despite being significantly smaller —> Read more.

🤖 Cool AI Tech Releases

GPT-4 API GA

The GPT-4 API reached general availability —> Read more.

XGen-7B

Salesforce open sourced XGen-7B, an LLM with 8k sequence length that matches the performance of models like Falcon, MPT and OpenLLaMA —> Read more.

🛠 Real World ML 

LLMs at Thoughtworks

Global software development firm Thoughtworks shared best practices learned while building Boba, a co-pilot for product strategy and ideation à Read more.

📡AI Radar

• Cloud platform Digital Ocean announced the acquisition of AI/ML cloud provider Paper Space for $111 million.

• Notion Capital announced a $300 million fund with a strong focus on generative AI.

• Secretive AI startup Humane unveiled its first device for personal AI compute.

• Generative AI platform for SEO content SpeedyBrand came out of stealth mode with $2.5 million in funding.

• YouTube started testing AI-generated quizzes in educational videos.

• Neko Health, the AI-healthcare company started by Spotify founder Daniel Ek, scored a $65 million in a new funding round.

• Celestial AI, a networking platform that can increase memory and bandwith for large AI models announced that it has raised $100 million.

• Kinnu, a platform that AI-powered learning content, just raised $6.5 million in new funding.