What’s going to happen with generative AIs in 2024?

With everything moving in the AI landscape, it’s difficult to make predictions… If we think back to January 2023, would we have predicted that generative AIs were going to be so powerful? Would we have believed hoaxes that we no longer believe now?

Last First Friday, we discussed misconceptions and false expectations around generative AIs with our “expert” friends Watch the video.

For a software and data boutique like ours, we need to stay alert and keep learning.

One of the ways to stay up to date is with our subscription to DataCamp (this is not advertising, we just like these people).

DataCamp is a data science and programming training platform that allows us to stay current with the latest data analysis and AI techniques and tools.

Plus, they certify your knowledge and allow you to demonstrate to your clients that you’re up to date (despite the years piling up)

Let’s get to it:

AI generative predictions for 2024 according to taniwa (and largely DataCamp, assuming this keeps heating up):

The battle between private and open source models intensifies:

Some are worried that chatGPT (and other models like BardGemini, Anthropic, Coehere, EleutherAI) will dominate their AIs’ responses and control the information generated from users.

Some respond that open models are what will save us from the dangers of large corporations’ commercial interests.

Some believe that regulation will protect us, but others think regulation will be a burden on innovation.

I think that open models like LlaMA2 will help prevent monopolies and democratize access to AI, but not entirely: Not everyone has the resources to train an AI model and run it extensively.

Pay attention to HuggingFace, lots of models to use and many datasets. If you want to do something with AI, you can’t ignore it. There will be a post about Huggingface soon.

“Boring” models are the ones that will succeed:

We’re constantly amazed by the new capabilities of generative models, but the reality is that most business problems are solved with more “boring” models like classification, regression, clustering, summarization, Q&A, but applied to company data and specific needs.

Beyond the personal productivity improvements of generative models, companies will focus on process improvement, task automation, time and resource savings using tasks such as:

• Automatic data classification, like customer complaints.
• Text summaries, like project documentation and legal matters.
• Anomaly detection, like fraud in transactions.

Small and specialized models:

We’ve already mentioned it: “build your own MyGPT”. You use a generic model but reduce it to the part that interests you. It’s like taking everything chatGPT “knows” about marine biology and keeping only what it knows about programming.

You can also specialize a model with BERT or PALM in a specific domain, or fine-tune a general model with your data: For example, Med-PaLM 2, which is a medical language model with specialist-level response accuracy.

Check this out: A responsible path to generative AI in healthcare

The importance of data:

The AI explosion makes companies interested in this world, only to realize how important data is.

What’s important is the data or its governance. If you haven’t already, investigate DataMesh, which is a way to manage company data in a decentralized and secure manner.

A generative AI is only as good as the quality of your data.

Learn, experiment, make mistakes, and start over:

With so much media noise and so many expectations, there are many superficial fishermen. It’s hard to differentiate yourself if you stay in the soft layers of all this, in the headlines and obviousness, which are curious for a pundit but not for a professional who wants to provide answers to business problems.

Some things we do at taniwa:

• Copilot for everyone and knowledge and experience exchange meetings.
• Continuous training on DataCamp and other platforms. Certification is even better (I’m at 70% of my Data Scientist certification)
• Everyone has access to PacktPub books and videos.
• Proof of concepts: Right now we’re working on a project to classify headlines as boastful or not (clickbait as they call it)
• Playing with platforms: ollama, HuggingFace, Kaggle, Snorkel (good leads there)
• We organize First Fridays to invite experts and learn from them.

If you want to know more about how we can help you with your data and AI projects, don’t hesitate to contact us at hola@taniwa.es

Photo from Pexels