Emergence Capital

Are we at the top of the S-Curve for AI? Likely, yes. Performance is plateauing. AI companies are scavenging for data. For example, with Whisper, OpenAI has transcribed a million hours of YouTube videos for GPT-4. Model providers can keep following this path but they won’t escape the flattening S-Curve this way. The next great training source for AI models is data produced in a work context. It's of far higher quality than what’s left of public data for training purposes, especially compared to running the dregs of the internet through the transformer mill. We have written a deeply researched piece about how startups can help us make the jump to the next S-Curve: By helping companies tap the brilliance of their people, and by building the next great AI tools with business data. We've identified four areas for startups who want to solve for breaking out of the AI plateau to build something new, useful, and powerful. Take a read, and if you're working on this problem, please get in touch! https://lnkd.in/g6M8UvzF

Congratulations to Matt Calvano, Henry Bradlow, Francisco Enriquez and the entire team at Adaptive! We're beyond excited to support their ongoing work of building a world class financial solution in the construction industry, which sorely needs it. Adaptive has already helped 280+ customers speed up and simplify payments, and there is still so much more to come.

We're glad to have you here!

Really interesting piece by Gordon Ritter and Wendy Lu arguing that tech development is not linear, but is a series of (hopefully) connected S curves. They argue that currently technology development in the AI world has plateaued due to reliance on public data. The next leap will come if companies can leverage proprietary business data, which is of higher quality and relevance. To achieve this, they recommend that startups should engage experts, leverage latent data within business apps, capture data in context, and secure proprietary models. On a different note. Really like the S Curve framing of technology development. Worth thinking about how that relates to the Gartner Hype cycle which impacts both investor, provider and customer behavior. It may help explain the infamous "trough of disillusion" in the Garner model as expectations from the hype often peak at about the time the technology hits its first plateau. #AI

If your feelings on GenAI have gone from enthusiasm to skepticism, then this really thoughtful piece from Emergence Capital is worth a read. My personal takeaways are: 1) Of course we've hit a plateau with GenAI; history is simply repeating itself and this is to-be-expected. 2) Access to (and proper use of) high-quality unstructured data is the key to unlocking more value from a company's GenAI initiatives. 3) Organizations need new ways to assess the accuracy and relevance of their unstructured data if they're going to be able to use GenAI in production and at scale. Thanks to Gordon Ritter and Wendy Lu for the great article and for highlighting Shelf as a vendor that can help organizations tap into the value of their latent unstructured data! https://lnkd.in/egfEu3id

The majority of companies have already released a #GenAl product or feature, and most of those who have not plan to do so in 2H 2024. As a #founder, if you are not currently thinking about how your team can leverage GenAl to build your company or how GenAl can power your product, you may be falling behind. For the companies that infused GenAl into their product, most have done so as a feature that exists within or complements an existing product. Only 40% have released a new product or SKU. Learn more about #AI in #B2B #SaaS in #BeyondBenchmarks: https://lnkd.in/gg5eCasX

Only 58% of companies that have released a GenAl product or feature have chosen to specifically monetize it. This is a potential challenge as the cost of GenAl inference is significantly higher than traditional software. Almost 50% of the companies that are monetizing are pricing based on usage which aligns revenue to cost, and enables companies to have predictable gross margins. One-third of companies that are monetizing are charging a flat fee. This promotes customer adoption, but makes cost forecasting more difficult. That being said, on average, we observed companies who charged a flat fee had higher gross margins than those with a usage based pricing model. Learn more about AI enterprise adoption, pricing, ROI and model selection in our Beyond Benchmarks report. Free to download, link in comments.

Thank you, Colin Kennedy!

This was a great piece by Emergence Capital highlighting the appetite for training of OS models on enterprise data. I keep saying it, all organizations should - and will - own and customize their own LLM's on their own data. This will unlock massive value and ROI compared to hitting a closed-source API.

When ChatGPT was first released in November 2022, it created a global shockwave. For a while, each subsequent release, and releases of models from other companies like Anthropic, Google and Meta, offered drastic improvements. These days, the incremental progress of each new LLM release is limited. Consider this chart of performance increases of OpenAI’s flagship model. Although every benchmarking system has shortcomings, clearly the pace of change is no longer setting the world on fire. What’s needed now, and what we hope is coming, is the jump to the next S-Curve. We believe we understand what’s caused AI to plateau, and what is needed to make the next jump: Access to the next frontier of data. Link in comments.