We all know what the industry’s main character is right now – ChatGPT. But natural language processing (NLP) is in many respects a project as old as tech itself. A ton of companies are working on this stuff, some even before the current round of hype, with the attendant Great Pivot from Web3 to LLM. One such company is deepset.
Founded in 2018 in Berlin by Milos Rusic, Malte Pietsch, and Timo Möller, deepset maintains the open source haystack project, which is designed to make it easier to use Transformers and large language models (LLMs) in your applications. Transformers are a concept introduced by Google in 2017 in the seminal paper Attention Is All You Need – a neural network architecture that has dramatically accelerated the state of the art in AI/ML. deepset wants to make this kind of technology usable and useful by the enterprise, with both on prem and cloud products. Because for all the excitement about LLMs and related technologies there is also a lot of fear, uncertainty and doubt. Who owns the models likely owns the moats. Enterprises and governments are concerned about ownership and business sustainability. Samsung recently had a leak of source code and trade secrets after engineering teams used ChatGPT in a planning meeting. ChatGPT has been banned in Italy because of piracy concerns. So much for data protection – It’s not clear whether the type of crawling and learning approaches pioneered by OpenAI are even compatible with EU law, in the shape of the the General Data Protection Regulation (GDPR). Ant Stanley covers a lot of this in a great post on his new blog, with this post Ask for forgiveness, Not permission
Anyway, when an area is so hot it’s always interesting to talk to folks that are steeped in it. I was lucky enough to catch up with Pietsch recently, for a RedMonk Conversation video. It was funny that we both have stories about moms using ChatGPT. While I am not a fan of the “even my mom can do it” framing, it’s definitely worth paying attention when a technology is crossing over so fast to mainstream adoption. Conversational AI based on LLMs is “haptic” – the feedback loops are just very immediate. Insert Mythic Quest reference here.
Mainstream adoption creates all kinds of challenges for the kind of innovation unleashed by OpenAI and ChatGPT. That’s where data and model sovereignty, compliance, the avoidance of AI-driven hallucinations in content, code and decision-making comes in. Those are the kinds of areas where deepset is focusing its attention. What multicloud was to the last 10 years, multi-model probably will be to the next ten. We’ve already seen AWS start positioning itself accordingly/. Multi sounds good when you’re not the market leader.
OpenAI will be a winner, but not the only one.
A concept you’ll be hearing a lot more about is Retrieval Augmentation – in terms of improving models. Again we cover that in the conversation. So dive in!
So watch the video, and tell me what you think, here or on Youtube, but in the meantime I will leave you with a story from deepset about a gentleman in his 80s that runs a legal publishing firm in Germany. He called deepset just before Christmas last year to insist on a meeting before the end of the year to discuss ChatGPT’s potential implications on his business, and how he could do something similar but without giving his own information away. ChatGPT only launched on November 30th 2022. That’s the scale of the challenge, and the opportunity.
disclosure: AWS, Google and Microsoft are all clients. deepset sponsored this video.