2023 AI Predictions check and predictions for 2024

I published my predictions in an article in January 2023. (see here ) . I included my last year's text, which is in Italic. This year's new text is in bold.

"2022 (can be replaced with 2023) was an amazing year for technology. I believe it was a milestone year, especially for the use of AI in mainstream technology. The democratization of AI was one of the most important trends, the access to this technology by citizen developers became a normal fact. I think that the big COVID era hype about the Metaverse was replaced by this trend, the metaverse didn’t find yet a similar place in mainstream technology, it is still mainly oriented towards the entertainment industry.

Here are a few thoughts about what will happen in 2023 (can be replaced with 2024) in AI."

In January 2023, I wrote:

1.??????The Generative AI will become the basis for the development of many new systems and platforms. AI and ML in general will be used in a very high percentage, every company will need to invest in the usage of the new technology in order to keep up with the global trend. The LLMs and image and visual generative systems will be incorporated into a lot of technologies. The Generative AI’s Foundational Models will become the basis to develop systems for different industries. The main challenge is that Generative AI cannot be used out of the box.

Similarly with a house, after you build the foundation and the pipes, you need to make the space livable and build on it something that will fulfill the living needs.

In 2023 the big LLMs (Large Language Models) will continue to be developed, and in my opinion those which are not Open Source yet will become Open Source. This migration towards Open Source will democratize the Foundational Models.

For example the open source Bloom model developed by a Hugging Face project, has a similar architecture with Open AI GPT3 but has been trained in more than 50 languages and 13 programming languages. You can read about this here .

CONCLUSION: Everything I mentioned happened. This trend will continue in 2024. There are more and more Open Source models, the Open AI "revolution fiasco" just strengthens this movement.

2.??????The “last mile” deployment of Generative AI.

This terminology is mine, it is a little bit exaggerating a similarity with the last mile delivery. During COVID, the food delivery business boomed, many companies made fortunes with it. The same was with package deliveries. The problem was to insure the “last leg “of the delivery to the destination, to the customers. The same with Generative AI, you cannot use the ChatGPT type of system out of the box, without further developing it for the specific vertical needs of the place where is used. It needs some work.

This is done with further training, using humans. One of these type of methods is the RLHF (Reinforcement Learning with Human Feedback). You can read about this here .

This method introduces other models, for example, a Reward model that is also used on top of the initial language ?model in order to further train the models with specifics of the vertical of interest. The information for this training is supplied by domain experts who are participating in this process.

This type of training requires much fewer samples and consists of rating the responses to some prompts and even entering in the system conversations between SMEs about a subject. (SME – Subject Matter Expert).

Many companies will be created to build specific systems for different industries, probably most of them will be at the beginning for customer support systems.

This, in fact, proves that the ML for a user is nothing more that a combination of a good UX and Human-Computer Interaction. As the founder of Scale.com said, “Labelling is the new programming”.

If you want to see a system like this, the UiPath Automation Platform’s Communication Mining System (formally known before the acquisition as re:infer) is exactly that. An SME is working in tandem with the AI system to label and confirm the automatic labeling of the customer communication messaging. The AI also suggests tasks to be done to improve the trustworthiness of the AI. You can read about this here .

CONCLUSION: I didn't know about RAG last year. RAG was in stealth mode. In 2023 RAG became the big thing, people talk less now about RLHF, they talk more about RAG, vector databases. In 2024 RAG will be further developed, and used on a large scale, for cost reduction and accuracy reasons also. UiPath also further developed Active Learning as an AI-assisted optimized labeling process, that reduces the cost of fine-tuning by combining supervised learning with unsupervised one.

3.??????Prompt Engineering

One of the characteristics of Generative AI is that you need to ask the AI the right question and provide the right context in order to get the right answer. This fact created 2022 a new domain that nobody heard about before , better said only a few people knew it.

In 2023, I am sure that many companies will create tools to be used to create the right prompts, specialized in different verticals and subjects.

This is essential in every Q&A system, including the customer support systems. Just think about a use case like this. Everybody knows how complicated is to create content when a company launches a new version of a product. It takes months to prepare all the content required for the go-to market, including training materials for support and customers. I am sure that this technology will allow the upload of the specifications of the changes for the new version, and the AI will learn the new features and will instantly create significant support for the new version’s GTM (Go To Market).

CONCLUSION: Prompt engineering got swallowed by RAG. It will further develop in 2024, with focus on multimodal prompts and models.

4.??????Trustworthy AI

One of the biggest problems of the new Generative Models is the fact that these systems can provide false information and/or can be biased. I am sure that in 2023 many companies will create solutions to increase the trust factor and eliminate false answers and biases.

The main factors of making an AI system trustworthy are privacy, robustness, explainability, fairness, and transparency. Many companies in 2023 will provide solutions for all these factors that will increase the trust in the Generative Models and will prevent losses and dangers due to false negatives or false positives in the AI provided information.

CONCLUSION: Many companies focused on hallucination detection, bias elimination. Human in the loop became the de facto standard to implement guardrails that will prevent loses or negative actions based on wrong AI creations. UiPath deals with these using Human in the Loop as mandatory guardrail in critical processes, and also Bias detection and curration in the Active Learning process.

Unfortunately, the current global geo-political situation is very complex. The conflicts in different regions of the world, and the existence of rogue authoritarian governments (China , Russia, Iran, North Korea) add a layer of real danger to the usage of these technologies.

CONCLUSION: I would add the complexities created by the Israeli war.

Just think about the usage of technology to influence elections in democratic countries. ?The creation and training of the Generative Models with biased content toward of political party or another can create serious problems. Or the use of technology for military applications can create serious damage in the world.

I am sure that the AI-related laws will be updated with new features related to the mal usage of Generative AI. Also, one of the big challenges is to protect the new technology to get into the hands of bad guys. This is hard to realize on the software side, but on the hardware side, preventing the export of essential components required by the use of large models can be achieved.

CONCLUSION: We wil see this in 2024, which is a critical election year in the world. Most of the important countries in the global geo politics will have elections. In 2023 many companies provided technologies to detect deep fakes, detect text created by generative AI, and allow content issuers to "water mark" their content. Several export embargoes implemented by the US related to AI technologies (hardware and software) brought significant measures to protect the use of AI by the rogue nations. The real test of these measures will be during the elections that will happen in 2024.

More and more Legislative measures will be taken by different countries and global organizations. Enforcements of the EU and US legislation for AI will be started.

New trends in AI for 2024, didn't exist in 2023:

1. Further cost reduction of the ML training costs, through RAG and edge computing - reduce the dependency on very costly cloud providers. AI will be used on a large scale on the smart phones, LLMs will end up on them.
2. Further improvement of Dev Ops related to AI
3. Creation of Constitutional AI type of systems that will use AI to enforce the usage rules (AI enforced guardrails)
4. Usage of quantum computing in AI
5. Elimination of dependency on China or Taiwan of the chip industry (US is building twin factories in the US to the Taiwan ones - to reduce the strategic risk created by China)
6. Increase the use of AI in the military
7. Spectacular development of Intelligent AI based assistants, swarms of agents, drones

CONCLUSION

We are living in very exciting times related to the use of AI in everyday life. In 2023 (I can change this to 2024) almost everybody will feel the impact of the new technologies, but this will bring new challenges that will need to be overcome.

Happy New Year !!!!