AI predictions 2024 // LLMs, agents AGI, super-intelligence, copyright, regulation & beyond
- Competition between the main AI products: ChatGPT, Copilot, Bard, Claude, Llama, Mistral, and derivatives. New names will come to the top league // open source AI will surpass proprietary due to the speed of innovation.
- Multimodal LLMs (MLMs) usage is a norm, you just easily can mix different data formats in one context: text, images, audio, video.. // one step closer to models capable of perceiving this world the way humans/animals do.
- New models, more efficient, better quality, better reasoning, less hallucinate, can run (inference) on smartphones and low power devices // yet another step to autonomous machines.
- Opensource autonomous vehicles autopilot on $150 smartphone // forget $40K Teslas
- Generative video will blow your mind // pika, runwayml, etc.
- Copyright paradigm shift due to generative AI // new norms will come
- No real AGI or super-intelligence // sorry, to build something it must be properly researched/defined, we still don’t know what intelligence is, but hype continues
- Regulation, over-regulation and alignment will slow down progress // technology misunderstanding & risks overestimation
- Magic “emergent” properties marketing // nothing personal just business
What a year it’s been for AI.
Anticipating what’s next sounds very perilous, but I’ll give it a try.
Here are eight AI predictions for 2024🔮
— — — — — — — — — — —
1) AI smart glasses become a thing 😎
As multimodality rises, leading AI companies will double down on AI-first wearable devices
And what’s better than the glasses form factor to host an AI-assistant ?
Branches are close to the ears to deliver audio, cameras are close to the eyes to capture ego-centric input, plus they’re hands-free and comfortable
Meta is leading the way with Ray Ban, but think about the recent Open AI x Snapchat rumor 👀
We’re only getting started…
2) ChatGPT won’t be to AI assistant what Google is to search
2023 started with ChatGPT taking all the light and ends with Bard, Claude, Llama, Mistral and thousands of derivatives
As commoditization continues, ChatGPT will fade as THE reference ➡️ valuation correction
3) So long LLMs, hello LMMs
Large Multimodal Models (LMMs) will keep emerging and oust LLMs in the debate; multimodal evaluation, multimodal safety, multimodal this, multimodal that
Plus, LMMs are a stepping stone towards truly general AI-assistant
4) No significant breakthrough, but improvements on all fronts
New models won’t bring real breakthrough (👋GPT5) and LLMs will remain intrinsically limited and prone to hallucinations. We won’t see any leap making them reliable enough to “solve basic AGI” in 2024
Yet, iterative improvements will make them “good enough” for various tasks.
Improvements in RAG, data curation, better fine-tuning, quantization, etc, will make LLMs robust/useful enough for many use-cases, driving adoption in various services across industries
5) Small is beautiful
Small Language Models (SLMs) are already a thing, but cost-efficiency and sustainability considerations will accelerate this trend
Quantization will also greatly improve, driving a major wave of on-device integration for consumer services
6) An open model beats GPT-4, yet the open vs closed debate progressively fades
Looking back at the dynamism and progress made by the open source community over the past 12 months, it’s obvious that open models will soon close the performance gap.
We’re ending 2023 with only 13% left between Mixtral and GPT-4 on MMLU
But most importantly, open models are here to stay and drive progress, everybody realised that. They will coexist with proprietary ones, no matter what OS detractors do.
7) Benchmarking remains a conundrum
No set of benchmarks, leaderboard or evaluation tools emerge as THE one-stop-shop for model evaluation.
Instead, we’ll see a flurry of improvements (like HELM recently) and new initiatives (like GAIA), especially on multimodality.
8) Existential-risks won’t be much discussed compared to existing risks
While X-risks made the headlines in 2023, the public debate will focus much more on present risks and controversies related to bias, fake news, users safety, elections integrity, etc
<END>
Building upon your comprehensive and insightful predictions, here’s a smart addition that encapsulates the overarching trajectory of AI development in 2024:
1) Visual AI Explosion: Expect a surge in AI for generating images and videos, revolutionizing media production and creative industries.
2) AI-Blockchain Fusion: We’ll see groundbreaking integrations of AI with blockchain, enhancing security and trust in automated systems.
3) Ethical AI Frameworks: Ethical considerations and governance for AI will become paramount, with significant moves towards global regulatory standards.
To predict the future, look what’s going on now and near past.
To close out 2023, here are 10 of the most exciting AI advancements from researchers at Meta this year and where you can find more details on the work.
1. Segment Anything (SAM)
A promptable segmentation system with zero-shot generalization to unfamiliar objects and images, without the need for additional training. Details: https://bit.ly/3H0AuWo2. DINOv2
The first method for training computer vision models that uses self-supervised learning to achieve results matching or exceeding industry standards. Details: https://bit.ly/4aBAM3o3. Llama 2
The next generation of our open source large language model, available for free for research and commercial use. Details: https://bit.ly/3H51sfi4. Emu Video & Emu Edit
Generative AI research for high quality, diffusion-based text-to-video generation & controlled image editing with text instructions. Details: https://bit.ly/48ex4LE5. I-JEPA
Self-supervised computer vision that learns to understand the world by predicting it. The first model based on a component of Yann LeCun’s vision to make AI systems learn and reason like animals and humans. Details: https://bit.ly/3GYrP6D6. Audiobox
Our new foundation research model for audio generation. Details: https://bit.ly/4ayXFVb7. Brain decoding — toward real-time reconstruction of visual perception
Using MEG, this AI system can decode the unfolding of visual representations in the brain with an unprecedented temporal resolution. Details: https://bit.ly/3tyW2pQ8. Open Catalyst demo
A service that allows researchers to accelerate work in material sciences by enabling them to simulate the reactivity of catalyst materials faster than existing computational methods. Details: : https://bit.ly/3tybOBo9. Seamless Communication
A new family of AI translation models that preserve expression and deliver near-real time streaming translations. Details: https://bit.ly/3H0ykpC10. ImageBind
The first AI model capable of binding data from six modalities at once. A breakthrough that brings machines one step closer to the human ability to bind together information from many different senses. Details: https://bit.ly/3vhS5Xn
I’ll play the game too. My top AI predictions and anti-predictions for 2024 are:
1. Hallucinations, or confabulations, will be solved technically, but at a meaningful trade-off between cost and reliability. Prompt Engineers will have to make a new design decision in 2024 between the cost and the reliability of their products. Reliability will eventually win, though I’m uncertain about how quick.
2. Language User Interfaces (#LUI), which include both speech and text, will become almost instantaneous for all widely used AI products. Users will begin to expect instant, and this will lead to extremely engaging AI products. This will substantially grow the use of AI products, but at meaningfully higher cost per user. Consequently, leadership teams of AI companies will face hard choices between growth and profitability.
3. #AIsuperiority will become a priority issue for most nations, with the most capable AI models viewed as the new stealth fighter that you can have as many copies of as you have compute, the new oil, to run it on. This will become universally clear to governments as the big political winners in 2024 will be the ones who most cleverly used AI. Unlike previous international chess games, AI incursions can be difficult to detect and attribute, which may lead to a “warm” conflict.
My anti-predictions for 2024:
1. We won’t create Artificial General Intelligence
2. AI won’t be trusted with shopping decisions, let alone big decisions, but it will be used to verify important decisions and to coach people.
3. AI will displace specific jobs, but in 2024 it won’t result in a decrease in total jobs. The churn in jobs will be very frightening and frustrating to those going through it.My hell-if-I-knows for 2024:
1. Will Google catch up to Microsoft/OpenAI?
2. Will AI solve a complicated, previously unsolved, famous math problem?
3. Will an AI beat the Go Make $1M test? Since the Turing test has been crushed, the new test is, “Go make $1M dollars.” If 1M AIs are given $1 each, I’m not even sure one of them would make $1M.
2024 will be the year of useful AI 🙏🏻
2023 was about exploration, 2024 will be about exploitation (in the sense of utilizing and improving what’s already there).
One of my resolutions is to use the knowledge and expertise that I’ve gained during the last five years for an application that’s both dear to me and potentially beneficial for society.
I’ve decided to start connecting the dots: as an ex-pyschologist and LLM expert, I hope to be well positioned to tackle the domain of LLMs for Mental Health. And I’ll do it so publicly, for anyone to be able to learn from it. Think of it as an experimental content series over the course of several months.
While the focus will be on Mental Health, specifically Depression, the insights from my content will be transferable to any other domain. I’ll try to walk you through everything, ruthlessly, exposing even my own errors and potential flaws in my workflow.
The goal will be to do this truly end-to-end, starting with problem statements, data collection and all these other things that are too rarely talked about, eventually leading to a deployable prototype.
At worst, we’ll all learn something together. At best, the prototype will help at least a single person.
Please comment if you’re working on something similar or know someone that does.
For anyone interested in the topic, I’ll leave a link to a recent article from a lab affiliated with University of California, Berkeley about the current failures and future directions:
