We're in this weird moment right now in the AI cycle where enterprises are not sure to which extent they should keep doing what some perceive to be the "old" thing (call it, data science) but they're not sure either how to do the new thing (Generative AI) just yet, so basically not very much is happening, except for consultants. In reality, the future will be an ensemble of different data analysis and ML/AI tools and techniques in the enterprise, depending on the use case and business objective, and it's more timely than ever to invest in data and MLOps infra while the Generative AI stack matures.
Matt Turck what do you see in terms of provable data on GenAI adoption, pervasiveness, customer excitement?
I believe the "old" approach often fails due to the absence of high-quality data required for successful outcomes. Quantity alone is insufficient. I have seen instances where GenAI serves as a catalyst for traditional data science, enabling projects to succeed that would have otherwise faltered or become prohibitively expensive.
Matt Turck your point in investing in Data and MLOps infra is valid, but I am surprised at your point on - “so basically not very much is happening, except for consultants”. While perfecting GenAI applications might take some time, early results in applying LLMs to both unstructured data and structured data look very promising to me.
It's pretty crazy that companies that haven't solved data, let alone machine learning moved on to the arguably harder Generative AI. I am surprised how dead the MLOps space is. Those problems won't go away and GenAI just adds more to them.
The hype from GenAI is almost overpowering to the point where many conflate GenAI & AI. As you note, AI is an ensemble of methods and model types, each with its own use cases. GenAI, with its natural language interface, is far more visible. But if you want more precise answers, those so-called "traditional" forms of AI (e.g., machine learning, deep learning models) are going to be more reliable.
Our work at AI Infrastructure Alliance focuses on the evolution of MLOps. How agents, large language models and smart apps are demanding MORE not less from AI/ML Infrastructure. The next-gen emerging stack of AI, prompt engineering, open source and closed source generative models, common app design patterns, legal challenges, LLM logic and reasoning. Click up download or free report, https://ai-infrastructure.org/agents-llms-and-smart-apps-report-2023/?utm_source=post&utm_medium=linkedin&utm_campaign=lee
Some solid comments here. a) yes ML Ops still needed; to go from basic prompt engineering to production possibly w fine tuned models, control over data, and more b) the user experience of GPT-4 etc. raised the bar massively for all of us enabling enterprises & communities c) transformers - self attention & positional encodings in concert with GNNs will have a broader impact than many anticipate, yesterday's GNN conference at Stanford indicated as much & I'm seeing much better than benchmark on some anti-fraud use cases d) net / net (aka TL;DR) -> slower and then bigger impact once again, enterprises will be flanked and disrupted, having not anticipated for eg what radically better & decentralized weather predictions, supply chain visibility, anti-fraud, and not to mention LLM++ mean for what is possible.
AI is basically drum machines and samplers and if you’re worried about Terminator you haven’t been using it enough.
co-founder / SVP Operations at Frame AI
1yEvery company's CEO and product team are desperate to work GAI into their products while their own security and legal teams are prohibiting their vendors from using GAI by name (or *gasp* "LLM") in their products without special permission.