2024 AI Predictions

2024 is here, and I have so many exciting things planned for us for this year. If you thought 2023 was a big year for AI, wait to see what's coming in 2024. This year will be an inflection point for AI, and in that spirit, here are my ten predictions for AI in 2024.

1 - LLaMA 3

My first prediction is that LLaMA 3 will be released in the first half of 2024. The headline on LLaMA 3’s release is its potential to bridge the gap between open-source models and cutting-edge proprietary ones like GPT4. It hasn’t even been a full year since LLaMA 1 was released - well, leaked - then, add in Meta’s drop of LLaMA 2 and we have an incredible influence on the trajectory of open-source models on the AI community.

Meta's commitment to open-source has established the company as a major player in the AI industry, and while Mark Zuckerberg has already teased the release of LLaMA 3 in 2024, Meta’s current focus is on refining LLaMA 2 for integration into their consumer products, which is highly impressive in itself.  If you didn't think LLaMA 2 was nearly as capable as GPT 4, it’s already proven to be production-level ready by the fact that Meta is building it into their products that serve billions of users. 

Zuckerberg has also teased AI Studio, a platform allowing users to create AI models as easily as user-generated content on Facebook, adding excitement for Meta's contributions to open-source AI in 2024.

 2 - Gemini Ultra

My next prediction, which is obviously not a true prediction, is the release of Gemini Ultra in 2024 by Google.  Gemini made waves at the end of 2023 with several demo videos and research papers that showed it was just as competent a model as GPT 4.  Gemini Ultra came across controversy, though - it was leaked that the video was highly edited and, therefore, not a good representation of what Gemini is capable of.  However, the performances detailed in the research paper are still very real, so I'm extremely excited to see what Gemini Ultra will bring to the table.  I'm all for more competition in the closed-source proprietary model space - GPT4 will have some strong competition from Google.

Gemini comes in three packages - Nano, Pro, and Ultra. Nano is going to live on-device and will be able to run completely from that device with no internet connection required.  Gemini Pro will be more on the level of an LLaMA 2 and will need pretty hefty consumer hardware to run. Gemini Ultra will only be able to be run by Cloud servers.

Google’s big play will be a heavy investment to ensure developers are utilizing the Gemini models - they want developers to build on top of Gemini, as does OpenAI with GPT4.  That is how you build a thriving ecosystem - by offering a very compelling developer product. 

Apple already has a very strong developer community, which will force Google Gemini and Open AI GPT4 to step up their game to capture the mindshare of developers.  Between Apple, Meta, Google, and Microsoft OpenAI the winner will be whoever can lure the most developers over to build incredible apps on top of their AI platforms.  

Google Gemini Ultra will be extremely high-quality - that's my prediction.  I think it’ll come with initial problems, such as hallucinations, vulnerabilities, and bugs, but it will quickly improve once it gets into the hands of consumers and developers.

3 - Robots

My next prediction is that robots will make enormous progress.  Humanoid robots and other types of robots will continue to evolve and improve, and more companies will release robot products in 2024.  Boston Dynamics remains a key player, but Tesla has made notable progress with its humanoid robot, Optimus.

I found a couple of super interesting videos (here and here) by a robotics expert highlighting the importance of a humanoid robot's speed in a factory setting. While Optimus currently achieves around 2 miles per hour, the expert suggests 3 mph as the minimum effective speed for factory use.  The primary challenge with Optimus lies in perfecting the actuators responsible for joint movement, but production is expected to accelerate once this hurdle is overcome, which is likely to result in substantial progress this year.  

Other predictions include potentially dozens of Optimus robots being produced by Tesla in 2024.  While we can’t ignore the ambitious timelines associated with Tesla and Elon Musk,  Tesla's evolution into more than just an electric vehicle company must be emphasized.  We already knew that they were an energy storage and production company - we’ll see Tesla quickly emerge as a robotics company as well this year.  Despite other robotics companies releasing products in 2024, Tesla's Optimus will evolve the fastest.

SPONSOR:

The Fastest Way To Build AI-Powered Apps

Create custom AIs for consumers, SMBs, and enterprises - no coding required.

Try MindStudio from YouAI for free today: https://bit.ly/3MTpbmc

4 - Open Source Acceleration

My next prediction: open-source large language models will accelerate in performance. This is the thing I’m most excited about this year.  In the video below, I showcase a trend curve highlighting the closing gap between open and closed-source models, predicting that the release of LLaMA 3 will further narrow the difference.

There are over 325,000 open-source AI models on Hugging Face, and I anticipate continued growth here with a major emphasis on the importance of improving quantization techniques.  I also predict the Mixture-of-Experts to become the gold standard for open-source models because it allows a huge model to perform really efficiently, as you don’t need to use the entire model when running inference. So, while it's a large model, you can run it on consumer hardware.  Then, when you add in quantization techniques, that is where it really gets exciting.

Meta is still leading the charge when it comes to open-source AI, with Apple also getting into the game, not to mention NASA and IBM with geospatial data.  We should also expect to see a lot of industry-specific models this year, such as Finn GPT and Finance - some that might actually be completely new models themselves. Apple’s release of their multimodal model ML Faret last year, which mostly flew under the radar, is completely open-source. This - along with the amount of software that’s come out recently that’s able to leverage the power of Apple Silicon - leads me to believe that Apple wants AI and LLMs to live on device.

In the below counterargument by Jaron Lanier, John cautions against the misconception that open-source leads to decentralization due to network effects. Lanier suggests that the exchange of free resources may result in monopoly-like centralization, and companies may protect their data, making it harder for open-source models to access datasets. A potential solution to overcome this challenge could be using synthetic data, which I cover in number 7 below.

 5 - AI Agents

My next prediction: AI will be the year of AI agents.  I anticipate significant improvements this year driven by advanced models and enhanced collaboration capabilities. AI agents will find real-world applications, generating templates for solving use cases in coding, research, and dataset creation. Emergent behavior from AI agents exhibited through habits and relationships, will become more prominent, raising questions about the nature of consciousness and what it means to be human.

I expect AI agents to assist in predicting outcomes in various situations, including testing game theory scenarios like the prisoners' dilemma. These agents could be instrumental in advancing social science, accurately reflecting human behavior, and predicting future actions within communities.

The potential applications of AI agents range from advertising and political polling to psychiatry, psychology, and dating. They could serve as predictive tools, eliminating the need for human trials and allowing scalable experiments. I also predict increased tooling in 2024 to aid in managing AI teams. Right now, the hardest part of running AI Agent teams is not necessarily implementing them but finding the right definitions of system messages, prompts, roles, etc., where they perform well together and get the output you expect.  Keep an eye out for a new AI agent project I’ll review soon.

6 - No AGI

My next prediction is that there will not be Artificial General Intelligence (AGI) in 2024. Mark Zuckerberg and Lex Friedman discussed the concept of superintelligence in a clip below, where they both acknowledge a breakthrough in the past year yet express uncertainty about when a singular AI system with general intelligence will be created. Zuckerberg emphasized the existence of organizations and structures that already exhibit intelligence greater than that of an individual, such as companies that have a singular brand and act as an entity and the stock market.

While the response to Sam Altman’s recent poll on X proved that by far, people do want AGI, he predicted that it likely would not be deliverable by this year.  Ray Kurle and Elon Musk seem to think AGI is coming around 2029, so seeing these two predictions play out will be interesting.

SPONSOR:

Monster API is the best place to integrate the latest AI models at a fraction of the cost.

https://www.monsterapi.ai

7 - Synthetic Data 

Next, I think 2024 is going to be the year of synthetic data now that we see how data is becoming much more valuable and much harder to come by. As companies guard their valuable datasets, synthetic data emerges as a solution, involving large language models in creating new data for training future models. While the effectiveness of synthetic data creation remains a question, I anticipate a growing trend in its use for model training.

A great example here is Tesla’s advantage in full self-driving technology due to the extensive real-world data collected from cameras on Tesla vehicles over the years. In contrast, other automakers face challenges in acquiring such data. They are presented with three options: installing cameras and collecting real-world data, purchasing third-party datasets (difficult to obtain), or creating synthetic data.

Synthetic data is expected to play a crucial role in industries where privacy concerns are paramount, such as healthcare and finance. In these sectors, using user data poses legal risks, making synthetic data an attractive alternative. Solving the synthetic data challenge is crucial for developing Artificial General Intelligence (AGI), as humans may not generate enough data manually, and companies are unlikely to share their valuable datasets.

8 - Multi-Modal

Next, I believe we’ll shift away from text-only models, and multi-modal capabilities will become the default for models in 2024.  Models like Gemini, future versions of GPT-4, and Apple's ML Faret are examples of this trend. These models will be trained to handle a variety of inputs, including images, videos, audio, and text.

Multi-modal won’t come without challenges, though. The below Standford lecture clip discusses issues such as one modality dominating others, added noise from additional modalities, and the lack of full coverage for specific modalities in certain datasets. 

Regardless, I remain optimistic about the impact of multi-modal capabilities and anticipate significant advancements in this area in 2024.

9 - Evil Bots

Detecting bots will become nearly impossible in 2024, and synthetic data could be a key contributor to this challenge. Data proliferation improves models and malicious bots, particularly those designed for scams and spam. Bots are already proficient at solving CAPTCHAs and other "prove you’re human" tests, making distinguishing between AI and human users difficult. 

Elon Musk's proposed solution to combat bots involves introducing a small monthly payment for using platforms like X. The rationale behind this approach is to make running a bot network economically unfeasible, as the cumulative cost of running numerous bots becomes significant. By introducing a financial barrier, the speaker argues that it disrupts the incentive system for bot networks, making it less profitable for malicious actors.

Going into this election year with deep fakes, with bots being very difficult to detect, and with all of the other deceptive practices during the 2024 election cycle, we’re going to need to put a huge emphasis on education to help people access information online and make informed judgments about its credibility.  This education is a necessary measure to empower individuals to determine whether they should believe what they’re reading or not.

10 - GPT 4.5

Last but not least, I predict the release of GPT 4.5 in either Q1 or Q2 of 2024, and it’ll be a significant evolution in various aspects.  However, it won't represent a leap to GPT 5, as Sam Altman has reportedly stated that they are not currently working on GPT 5.  I anticipate that GPT 4.5 will offer performance, speed, and cost-effectiveness improvements, but it will still be based on the existing GPT 4 model and architecture.

At the end of 2023, there were a lot of rumors that GPT 4.5 was already out in the field, but it turns out that was AI hallucination and speculation.  Individuals had apparently manipulated AI to generate statements indicating the use of GPT 4.5, although it wasn't the case.

2024 is bound to be an exciting one, and for me personally, I’m excited to bring you a series of tutorials, reviews, and discussions on broader AI topics.  Please let me know what your thoughts in the comments section.