I'm researching the interface between artificial intelligence, natural language processing and human pondering as director of the Laboratory for the further development of human and machine pondering on the University of South Florida. I’m also commercializing this research in a single AI startup which provides a vulnerability scanner for language models.
From my perspective, I observed significant developments in the sector of AI language models in 2024, each in research and in industry.
Perhaps probably the most exciting of those are the capabilities of smaller language models, assistance in combating AI hallucinations, and frameworks for developing AI agents.
Small AIs are causing a sensation
At the center of commercially available generative AI products like ChatGPT are large language models (LLMs) which are trained on massive amounts of text and produce convincing human-like speech. Your size is usually measured in parameterthese are the numerical values ​​that a model derives from its training data. The larger models like those from the large AI firms have a whole lot of billions of parameters.
There is an iterative interaction between large language models and smaller language modelswhich appears to have accelerated in 2024.
First, organizations with probably the most computing resources are experimenting with and training ever larger and more powerful language models. These lead to latest functions for giant language models, benchmarks, training sets and training or input tricks. These, in turn, are used to create smaller language models – within the range of three billion parameters or less – that may run on cheaper computer setups, require less energy and memory to coach, and might be fine-tuned with less data.
So it's no wonder that developers have released a wide range of powerful smaller language models – although the definition of small is continuously changing: Phi-3 And Phi-4 from Microsoft, Lama-3.2 1B and 3BAnd Qwen2-VL-2B are only a number of examples.
These smaller language models might be specialized for more specific tasks, equivalent to quickly summarizing a series of comments or checking facts against a particular reference. You can work along with their larger cousins to supply ever more powerful hybrid systems.
Wider access
Increased access to high-performance large and small language models is usually a mixed blessing. With many follow-up elections all over the world in 2024, the temptation to misuse language models was strong.
Language models may give malicious users the power to generate social media posts and fraudulently influence public opinion. There was one great concern about this threat in 2024, because it was an election 12 months in lots of countries.
And sure enough, a robocall spoofing President Joe Biden's voice asked Democratic primary voters in New Hampshire stay at home. OpenAI needed to intervene disrupting over 20 operations and fraudulent networks that attempted to make use of its models for fraudulent campaigns. There were fake videos and memes created and shared with the assistance of AI tools.
Despite the Fear of AI disinformationIt is What impact these efforts actually had just isn’t yet clear about public opinion and the US elections. Nevertheless, US states have passed a lot of them Legislation in 2024 Regulation of the usage of AI in elections and election campaigns.
Bots misbehaving
Google began recording AI overviews in its search results and returned some results that were weird and clearly improper – unless you prefer it Glue your pizza. However, other results can have been dangerously improper, equivalent to when this was proposed Mix bleach and vinegar to scrub your clothes.
Large language models, that are those mostly implemented, are vulnerable to hallucinations. This means they will say things which are false or misleading, often using protected language. Although I And other While we proceed to put it up for sale, many organizations will proceed to learn the hard way in 2024 in regards to the dangers of AI hallucination.
Despite extensive testing, a chatbot plays the role of a Catholic priest advocated baptism via Gatorade. A chatbot Advice on New York City laws and regulations incorrectly said it was “legal for an employer to fireside an worker who complains about sexual harassment, fails to reveal a pregnancy, or refuses to chop off her dreadlocks.” And OpenAI's speech-enabled model forgot whose turn it was to achieve this speak and answered a human in her own voice.
Fortunately, in 2024 there have been also latest ways to alleviate and live with AI hallucinations. Companies and researchers are developing tools to make sure secure AI systems Follow the given rules before deploymentin addition to environments to judge them. So-called Guardrail scaffolding Inspect the inputs and outputs of huge language models in real time, although often using a special layer of huge language models.
And the conversation to AI regulation acceleratedresulting in the main players in the big language model space updating their policies scale responsibly And Use AI.
But although researchers keep checking out Ways to cut back hallucinationsin 2024, research convincingly shown this AI Hallucinations will at all times exist in some form. This could also be a fundamental feature of what happens when an entity has finite computing and knowledge resources. After all, this can also be known in humans confidently misremember and speak untruths infrequently.
The Rise of the Agents
Large language models, especially those based on variants of Transformer architectureare still driving probably the most significant advances in AI. For example, developers are using large language models not only to create chatbots but additionally to function the idea for AI agents. The term “agentic AI” became famous in 2024some experts even call it that third wave the AI.
To understand what an AI agent is, imagine a chatbot augmented in two ways: First, give it access to tools that provide this Ability to take motion. This might be the power to question an external search engine, book a flight or use a calculator. Second, give him more autonomy, or the power to make more decisions on his own.
For example, a travel AI chatbot can perform a seek for flights based on the knowledge you provide, but a tooled travel agent could plan a whole itinerary, including finding events, booking reservations, and adding them to the calendar.
In 2024, latest frameworks for developing AI agents emerged. To name just a number of: LangGraph, CrewAI, PhiData And AutoGen/Magentic One were released or improved in 2024.
Companies are fair begins to adopt AI agents. Frameworks for developing AI agents are latest and evolving rapidly. Additionally, security, privacy and hallucination risks are still a priority.
But global market analysts predict that it will change: 82% of organizations surveyed Plan to make use of the lively ingredients inside 1-3 yearsAnd 25% of all firms currently use generative AI are expected to introduce AI agents in 2025.
