Amidst the rush to create prototypes using large language models and releasing them on Hacker News, where none of them have seen integrated use cases, a startup is working on offering feasible solutions. Plunging right into a market, where LLMs can be leveraged for sales solutions projected to reach $770M by 2032, are former Google DeepMind employees through their startup Glyphic AI — an AI copilot for the sales team.
“I think one of the biggest challenges of Glyphic, as well as any LLM-based project, is handling hallucinations,” said Devang Agrawal, cofounder and CTO of Glyphic AI, in an exclusive interview with AIM. “One of the advantages of Glyphic is that, since most of our tasks are quite grounded, i.e. we are not just answering questions randomly, we’re trying to go through questions, calls, emails, and things like that, the scope of hallucinations is minimised. The answer is always based on something.”
In June, Glyphic AI officially came out of stealth mode and raised a pre-seed funding of $5.5 million.
Navigating Large Language Models
“I think ChatGPT has brought to everyone’s mind the capabilities of AI, and everyone is now thinking about how it could be useful within their particular use case,” said Agrawal.
Speaking of other challenges with LLMs, he said, “Most people are using large language models these days with some prompt engineering, but soon they will realise is that this approach is not scalable. You can have evaluation sets or tests, but it is difficult to know if you’re getting better or worse, and the sort of iterating on large language models is very challenging. For instance, by changing your prompt, you can completely break everything but you won’t truly be able to detect what broke. You can trip the model and the accuracy can come down, but it’s hard to test these models.”
At Glyphic, a mix of a few small models and large language models are used. “We, kind of, cleverly decide between GPT-4, Claude or Cohere based on the context and tasks we are trying to accomplish. For instance, Claude is able to understand long context and can understand 100,000 tokens at one go. This is something you can’t do with GPT-4 which can understand only up to 16,000 tokens. Claude is also trained in a much conversational way, whereas GPT-4 can be direct. So, based on the sort of context, we use one or the other model depending on what behaviour we actually want.”
Glyphic AI will sufficiently validate its product and use cases, and then move towards optimising them. After which, in the future, the company would work towards building their own language model. Agrawal also spoke about a recent trend that has picked up. “So, everyone is building these pretty-looking prototypes with large language models and putting them on Hacker News. While they look nice, we still haven’t seen deeply integrated use cases, which are of high quality, high fidelity, and are being used everyday — because it is really challenging to do so.”
DeepMind Harvesting Entrepreneurs
Glyphic AI co-founders Devang Agrawal and Adam Liska
Agrawal completed his graduation from Cambridge University and had always been fascinated with AI. He worked as a machine learning engineer at Apple on the Siri project. Switching from a product-focused role, Agrawal desired to move to the research and academic side of things. He thus joined the ‘research-heavy’ Google DeepMind as a research engineer, where he worked for two years before moving on to start Glyphic AI with another DeepMind senior researcher Adam Liska.
“DeepMind is one of the most innovative companies, and one of those few places where you can be working for years and still be learning so much, because you are always on the forefront. It was a tough decision to leave as the multimodal project we were doing was going in a really exciting direction,” said Agrawal. “But, Adam and I were always clear about wanting to build a startup. Right from my college days, I knew what I wanted, and was doing jobs to get the correct skills so when you do have a startup, it becomes more effective.”
Google DeepMind supports people by giving them the flexibility to work on projects that they desire and even allows them to take risks. “It is a very supportive organisation where you can take speculative bets which sometimes work out and sometimes don’t. It is a great platform to develop as researchers and helps us inculcate critical thinking, which is pertinent for research.”
Agrawal believes that having expertise and being at the centre of transformational technology helped with the switch. “Many people are leaving DeepMind to set up startups, as they have expertise in this new transformational technology, which is now ready to be used in a product.”
There were two waves of exits fuelled by technology change at Google DeepMind. “When reinforcement learning had come to the market, a number of people were leaving DeepMind to set up startups using the same, but what we realised was that it was really hard to bring deep reinforcement learning in a product context, and that wave slightly died out. This was around 2017. Now, having specialist skills in large language and multimodal models, people want to build products on it.”
A number of people have also gone on to build research-driven startups to solve large problems such as cancer. “Designing new molecules that might work better for treating certain sorts of cancer, etc. is a different sort of startup that requires a different sort of thinking. We’re just focusing on slightly more open-ended things and solving the problem in a fundamental way.”
Roadmap for Glyphic
Currently, with ten employees, Glyphic focuses on applying large language models and generative AI to transform B2B sales processes. “Right now, we’re just focusing on improving and optimising sales processes but we want to build on top of it to optimise the entire go-to market and product strategy with everything else.”
In line with the company’s vision in a year’s time, Agrawal is looking to have a research team as well. “As we grow and scale, we would invest deeply into research and this would be one of the key differentiators.” Glyphic currently provides services for software companies. “They’re so innovative with their sales processes and are willing to try out new products. Once we prove our products and models, we can expand into enterprises.”
Glyphic also works with a few Indian companies headquartered or registered in the US. “Microsoft is building something in this area, and so are other companies which are building copilot for sales, but I think this is one of the spaces where startups have a huge advantage. The technology is moving fast and you need to be able to completely rip out everything and change it within a couple of weeks if you want to stay ahead of the game. I think we definitely have an advantage because of our kind of technology background.”
