Los desafíos y oportunidades de la IA generativa

An interview with Dr. Jean-Claude Junqua

It seems like articles about Chat GPT, Bard, and Generative AI (Gen AI) appear almost daily. We caught up with Sigma’s Executive Senior Advisor, Dr. Jean-Claude Junqua, to discuss what goes into the development of these technologies and the future of generative AI.

What are some of the use cases you are seeing within Gen AI that you didn’t see even two years ago?

In my mind this technology can be seen as human augmentation and pushing humans toward what they do best.
If I was asked two years ago, I don’t think I would have mentioned all the use cases for generative AI in sales and marketing and how accessible they are. You can create multi-modal marketing or sales material that incorporates text, image and video and the content is actually pretty good. The use cases in these areas keep evolving.

An area that current models are doing quite well at is coding or creating scenarios to test code and so on.

AI is also being used in medical scenarios where AI can check diagnoses and supplement the knowledge of a doctor. Generative AI will have a big impact on collaboration and decision making. In healthcare, AI can give us an advantage. Humans have bias and are limited by their experiences. AI can act as an assistant that has broader knowledge or could question a diagnosis. It may be wrong sometimes, which is where human knowledge and decisioning comes in. This feedback can then help the model improve.

Another example is language dubbing from one language into a voice of another language. You need speech to text, translation and voice generation. All are difficult technologies, and currently each of them is not perfect, but progress is being made quickly.

Dubbing or speech to speech conversion is a great example. It’s not just translating. There’s tone, emotion, and nuance where an actor pauses in speaking. What would it take for a company like Netflix to dub a Greek movie into Hawaiian?

We are not too far off! YouTube is testing a tool that allows creators to dub content in 70 languages. Google and Meta both have generative AI video tools.

It won’t be too long before the level of quality and accuracy improves.

How are annotation needs evolving as these use cases become more complex?

Going back to the dubbing example, up till now annotation was more about text content, structure and words for the technologies that are used in the process.

The type of paralinguistic information that systems will be able to understand such as tone, emotion, style will change what we ask annotators to label to help machines understand and enrich the experience. In other words, more complex annotation will be needed.

There’s been a lot of commentary about crazy answers people have gotten from generative AI. Why is this?

Some of the current Gen AI models I’ve tested give answers that are not trustable.
For example, I asked one model, “Do you know Jean-Claude Junqua?”
The model said it didn’t know Jean-Claude Junqua.

So I said, “do you know Jean-Claude Junqua who wrote a startup innovation book?
The model responded, “Yes, I know him. He worked at IBM” and so on.

I never worked at IBM. But when I added more information, the model response was correct. It needed a more specific prompt with more context to get the right answer.

As users, we don’t know what data and languages models have been trained on. Different models allow for different levels of error and prediction. This is one important challenge of Generative AI.

What is needed to build high quality foundation models?

Creating large scale foundation models takes a lot of data and a lot of computing power. It has been and will continue to be the domain of only a few very large companies. The next tier of enterprises will create applications and customizations on top of these models or combine and integrate them. They’ll aim at specific use cases by leveraging these large language models. The focus will be on enhancing productivity or on developing new applications. Generative AI has the potential to reach many industries and applications.

There is a lot of focus on these huge models right now, but there are also efforts to make them smaller.

One way to create smaller models is to have better quality data initially instead of doing unsupervised training on lot of data that may not be accurate.

You may want to train some of these models on smaller amounts of highly accurate data which could generate models that are as accurate or even maybe better than large models that have been trained on a combination of high-quality and low-quality data. We’ll see more domain specific models where there may be a smaller amount of more accurate data to train the foundation model.

We’ve been talking about LLMs, but if we look at what is happening today, maybe 90-95% of the models that are being deployed in the field are using supervised training, meaning you establish ground truth by labeling data, you create a model and then you generate inferences on some new data.

The field is moving progressively toward putting more knowledge in the model. Transfer learning has been used for a number of years to avoid training a new model from scratch. Instead of recreating a model based on your new task, transfer learning lets you adapt a model that has already been trained.

For example, you train a classifier to recognize 20 objects and you have a 21st one that you want to recognize. You don’t need to retrain a model from scratch, you can adapt the model that has been trained on 20 objects with some new data that includes the 21st object. To do this, we need additional labeled data but not as much labeled data as if you were training the model from scratch. Generative AI is also following the same direction as for customizing initial models, also called foundation models, only a small amount of data is needed.

What should enterprises be thinking about if they want to adhere to ethical standards around generative AI systems?

Addressing bias and inclusion is an important part of creating ethical AI, in addition to making sure that the technology can be used to make people’s life better.
People who speak less represented languages will have less access to AI-driven technology and automation tools than those speaking more common languages. Bias is built into GenAI because most LLMs have been trained on data from the open Internet and this data is already biased (e.g. there is more content available in English than any other language)

The only way to be able to trust the answer provided by an AI system is to be able to train the system on data that is representative of those who will use it. It’s sometimes difficult to understand this because the system may appear to work. The AI provides an answer, but is it accurate and trustworthy? The only way to be able to trust the answer is to train on data that is representative for the people and cultures using it and keep humans in the loop to assess the model and make sure it is reliable. Outputs should be reviewed, especially when the content is not generated by humans or controls should be put in place to make sure that the data can be trusted.

With the EU AI Act and other countries considering legislation, there’s a lot more focus on enterprise accountability.

There are several things that companies need to be concerned about when they put models into production, especially if the systems deployed are autonomous.

While there’s still no regulation, companies should think about the legal aspects of building models such as copyright, privacy, security and permission for using different data sources. Explainability and traceability of data sources is becoming more of a focus.

There are also considerations from sustainability and ecological impact. In terms of energy consumption, large models require a lot of energy.

Models need to be reliable and accurate. Content moderation is also needed. If the model is not able to filter inappropriate content, or address inaccuracies, this can be a huge problem, especially for large companies that need to protect their brands.

Whether the model is used with a human in the loop approach or autonomously, both approaches need human oversight at some level.

How do you create a model that is accurate and limits bias for that language that the data scientist might not have personal knowledge of?

This is very difficult. You really need people that are native speakers of the target language to understand context, make the right decisions so that the model learns the right thing.

It’s very difficult to train models on cultural differences. When there is no data on a particular topic for a given language, the model will extrapolate from one language to another. This is why you need native speakers with accurate cultural understanding to avoid what is called, in the generative AI field, “hallucinations”.

How do companies like Sigma AI do this differently than what you might get from crowdsourced annotations?

It’s about data quality, consistency, and quality control. If you want quality, you need to have a well-designed quality control process.

Sigma AI works with client teams to establish quality parameters, refine annotation guidelines and continuously assess and improve upon labeling. Human-in-the-loop, and the expertise of Sigma AI’s delivery team help provide a high level of quality control on data and ultimately on the outcome of models.

To learn more about Sigma AI and our leadership team, please visit our company page.

¿Quiere saber más? Póngase en contacto con nosotros ->

Sigma ofrece soluciones a medida para los equipos de datos que anotan grandes volúmenes de datos de formación.