Edited by Aeneas Momoko
After 23 years of generative AI, what new breakthroughs will there be in AI in 24? The big guy predicts that even if GPT-5 is released, LLM is still limited in nature, and in 24 years, basic AGI is not enough to achieve.
23 years is a well-deserved "year of generative AI".
What breakthroughs will be made in AI technology in 24 years?
Jim Fan, a senior scientist at Nvidia, said that 2024 will be the year of video. While robots and embodied agents are just getting started, I think video AI will have a breakthrough moment in the next 12 months.
Greg Brockman, co-founder of OpenAI, predicts that 2024 will be a breakthrough year in terms of AI's capabilities, security, and potential impact.
Of course, in the longer term, this is just another exponential year that will make everyone's life better than it is today.
In the new year, will artificial intelligence shine like in 2023?
AI Guru 2024 Predictions
Meta researcher Martin Signoux made 8 predictions for AI in 2024, and even LeCun said he deeply agreed.
First of all, artificial intelligence smart glasses have become popular. With the rise of multimodal technology, leading AI companies will redouble their efforts to develop AI wearables. What could be better than a glasses form factor to host an AI assistant?
ChatGPT is to AI assistants what Google is to search. In 2023, ChatGPT began to shine, with Bard, Claude, Llama, Mistral, and thousands of spin-offs coming out one after another.
As productization continues to advance, ChatGPT will no longer be the only reference standard in this field, and its valuation will also face revision.
Goodbye to large model models, hello multimodal models. LMMs will continue to emerge and replace LLMs in the debate over multimodal evaluation, multimodal safety, multimodal this, and multimodality. In addition, LMM is a stepping stone towards a true general AI assistant.
There were no major breakthroughs, but improvements were made on all fronts. The new model will not lead to a real breakthrough (GPT-5), LLMs are still limited in nature, and are prone to hallucinations. We won't see any leaps that make them reliable enough in 2024 to "solve the basic AGI".
Improvements in RAG, data wrangling, better fine-tuning, quantification, etc., will make LLMs robust/useful enough for many use cases to drive adoption of a wide range of services across industries.
Small models (SLMs) are already emerging, but cost-effectiveness and sustainability considerations will accelerate this trend. Quantitative technology will also be greatly improved, thereby driving a wave of device integration for consumer services.
The open-source model defeated GPT-4, and the battle between open-source and closed-source gradually subsided. Looking back at the vibrancy and progress of the open source community over the past 12 months, it's clear that the open source model will soon close the performance gap.
Benchmarks remain a challenge. No set of benchmarks, leaderboards, or evaluation tools can be a one-stop-shop for model evaluation. Instead, we will see a series of improvements (such as HELM) and new initiatives (such as GAIA), especially when it comes to multimodality.
The risks that exist are not much discussed compared to the existing ones. While X-risk makes headlines in 2023, public discussion will focus more on existing risks and controversies related to bias, fake news, user safety, election integrity, and more.
Lightning AI的创始人William Falcon对2024年的预测是：
- The 1B model will outperform the 70B.
- Deploying the model on the CPU is almost free, not an API service.
- Data quality will result in a 10x performance boost.
- A combination of open-source models will beat the best private models.
- The compiler will speed up the model (training and inference) by at least 80%.
- Legislation will support content creators, not model developers.
开源工具平台LlamaIndex的创始人Jerry Liu表示，
- RAG will continue to be a big focus
- Every AI engineer still needs a strong foundation in software engineering.
- Vector database began to develop SQL-like interfaces and support multimodality
- Multimodal models are being used more in document processing (but first, computational cost/latency needs to be reduced)
- Full capabilities like GPT-4 become open source and faster/cheaper.
- If this is the case, the development of agents will flourish again.
- Hints are as important as they were before, but the importance of hint engineering decreases
In 2023, ChatGPT will have the highest number of visits in the world
Over the past year, AI has become ubiquitous and has even redefined entire industries.
Writerbuddy AI, an online content writing company, has researched 3000+ AI tools by scraping AI tool data using SEMrush, a well-known tool in the SEO industry.
It was found that from September 2022 to August 2023, the top 50 AI tools generated a staggering 24 billion visits, with an average monthly increase of 236.3 million.
Among them, ChatGPT exclusively accounts for 14 billion traffic, accounting for 60% of the analysis traffic.
Here are the key findings from the report:
- The AI industry grew by an average of 236.3 million visits per month. The 50 AI tools analyzed experienced a 10.7x growth rate, with an average increase of 236.3 million visits per month.
- Over the past 12 months, AI applications have received an average of 2 billion monthly visits. Over the past six months, the average number of monthly visits has surged to 3.3 billion.
- ChatGPT、Character AI和Google Bard的访问量分别净增长了18亿次、4.634亿次和6800万次。
- Most visited AI chatbots: ChatGPT is in the absolute lead, accounting for 76.31% of the total visits to the AI chatbot category. It was followed by Character AI with 19.86% of visits.
- Craiyon、Midjourney和Quillbot面临最大的流量下降。
- The United States contributed 5.5 billion visits, or 22.62% of total visits, while European countries together contributed 3.9 billion visits.
- AI chatbot tools are the most popular, with 19.1 billion visits.
- More than 63% of AI tool users access via mobile devices. There is a gender difference, with 69.5% of users being male and 30.5% being female
In addition to ChatGPT, which has swept the world, these technologies in 23 years are also amazing
23 years have passed, and the key word for this year is undoubtedly "generative AI".
The launch of ChatGPT at the end of '22, and the release of GPT-4 in March '23, allowed the world to see the widespread availability of large language models, making '23 a year of text, audio, and video generative AI.
In addition to this year's "darling" ChatGPT, other companies should not be overlooked, such as the company that released the first open-source language model, and several new AI startups, including Mistral, which released the best open-source language model currently available, Mixtral 8x7B, at the end of the year.
In addition to this, there are also the following impressive technologies.
The town of Stamford and the robot cat
"Stanford Town", which showcases an impressive application of text and coding tasks.
The team created a Sims-inspired sandbox environment with 25 AI agents, each with their own profession and personality, that can interact autonomously.
These agents exhibited credible personal and emergency social behaviors, including making plans and attending Valentine's Day parties. This work shows how LLM-based agents interact with each other and yields interesting results.
This idea has been adopted by other research and open-source projects, such as Auto-GPT and BabyAGI, while OpenAI has greatly simplified it through the Assistant API.
Basic models such as GPT-4, which have also been used in robotics, have made some progress, such as Google's robots RT-2 and RoboCat.
RT-2 is an AI model for robot control that can be Xi from robot and network data. The model can process text and image inputs and leverage its extensive network knowledge to perform tasks that have not yet been explicitly trained.
In more than 6,000 tests of the robot, the RT-2 was almost twice as successful in untrained tasks as its predecessor.
RoboCat, on the other hand, is an AI that generates training data to improve the robot's control.
Other companies' technologies, such as NVIDIA's multimodal VIMA model, also use the base model in robotics.
DreamerV3和FunSearch
In the field of intensive chemical Xi, researchers have also achieved many important results.
An example of this is the DreamerV3, which can handle completely different problems without any adjustments.
Without a human model, DreamerV3 will learn Xi how to mine diamonds in Minecraft.
Earlier this year, DeepMind also demonstrated AdA (Adaptive Agent), a foundational reinforcement Xi model for DeepMind.
AdA follows the classic recipe of the base model, trained on tasks with large amounts of data. AdA is significant because it shows that scaling in intensive chemistry Xi can make the model perform better on other tasks.
In various fields of science, deep science Xi is showing an increasing role.
DeepMind has developed AlphaTensor, a new algorithm for fast matrix multiplication.
At the same time, DeepMind's latest version of the AlphaFold protein structure prediction system overcomes many of the weaknesses of previous versions and opens up new possibilities for computational structure prediction.
In addition, Google DeepMind showcased FunSearch, the first to use code-generated language models combined with evolutionary search algorithms to find previously unknown solutions to mathematical problems.
OthelloGPT、Q-Star和AI法案
2023 is also a year of AI regulation and a year of warnings about the risks of AI existentialism.
This trend will undoubtedly spur research into the industry to better understand the inner workings of LLMs.
There were some interesting papers during this period, such as OthelloGPT, Microsoft's GPT-4 embodies the AGI spark, and Google's paper on the "epiphany" of large models.
The field of prompt engineering provides insight into LLMs.
François Chollet explains prompt engineering as finding the right vector program and Promptbreeder, suggesting that prompts may become more automated in the future.
At the end of the year, rumors spread about Q-Star, accompanied by people's fear of AI, the hype of AGI, and the farce of OpenAI's palace fight, which was reversed many times in just a few days.
In 2024, perhaps we will see less speculation and more negotiations.
In what categories is the data used in AI training reasonable? The recent lawsuit filed by the New York Times against OpenAI has aroused widespread discussion in the whole society.
A similar debate will be held in the European Union, which agrees on the EU Artificial Intelligence Act by the end of this year. The details of this bill will be decided next year and will have a significant impact on the AI market in Europe.
2024 AI Outlook
After an explosive 2023, what progress will there be in the field of artificial intelligence in 2024?
Needless to say, in this new year, we will still see leading-edge AI applied in many new and creative ways to drive progress across the industry.
Copilot AI登台:智能体时代来临
OpenAI released GPTs, Assitants and other tools at the first developer conference, Microsoft products were renamed Copilot, etc., and the intelligent body ushered in a big explosion this year.
These tools are already starting to make an impact in one industry after another, but what we've seen so far is tiny compared to what's to come.
Earlier this year, a ReAct paper published by a team from Princeton and Google showed how large models can effectively learn how Xi how to use tools, and has spurred a lot of research in this area.
Companies including OpenAI and Anthropic have spent a year tweaking their models to better use the technology.
For example, OpenAI's function calls, and Anthropic's Claude XML support.
Project Address: https://react-lm.github.io/
There are also research institutes that have trained specialized large models, such as Gorilla LLM in Berkeley.
In addition, open-source code libraries such as Langchain and Rivet have made it much easier for agents.
As you can see, AI agents are easier and cheaper to develop than ever before. They harness human ingenuity while deeply connecting the data that matters most to users and companies.
In 2024, we will see the arrival of the "Age of Agents", which is the beginning of a whole new direction in meeting needs and interacting with technology through software.
Multi-modal large models break through visual barriers
ChatGPT's ability to understand and express human natural language is a groundbreaking feature that appeals to users and developers alike.
However, 2024 will see that AI vision is likely to be even more important and far-reaching.
Words are powerful, but images, video, and audio can convey information and emotion in a more focused way. The spatial expression of ideas is a very powerful tool for simply conveying complex concepts.
LLM can not only train text data, but also visual data, and the multimodal ability is more obvious.
We have already seen the development of wearable devices such as Ai Pin, Apple Vision, etc., which are expected to help us in our daily lives.
For example, they can provide contextual information about the person you're communicating with, work-related visual cues, or real-time suggestions for completing tasks.
It's hard to say where innovation will go and how fast, but being able to interpret images and videos and react instantly to physical changes in the environment adds an extremely important dimension to how intelligent AI can help humans.
AI manipulation reaches a dangerous level
While the explosion of AI has brought earth-shaking changes to various fields, it has also brought us to see that the generation of false information by AI has brought trouble to life.
Never before in human history has AI been so powerful or pervasive to influence and manipulate AI on a large scale.
AI has made it almost impossible to discern "real" social interactions and content, as images and even videos can be easily generated.
AI manipulation is likely to be all the rage in the coming year, from automated extortion and fraud to the spread of conspiracy theories.
All in all, in 2024, AI will bring many incredible things to the world, but it will also challenge us in new ways.
Friends predict
The imagination on this topic is also on the Zhihu hot list.
Zhiyou "Lead Xiaobai" predicts that in 24 years, the model effect will be further breakthrough, and it may be that as long as 7B model inference resources are available, it will be on par with the current GPT-4.
With the significant drop in deployment costs, 24 years may become the first year of AI Agent, and there will be a hit.
It is possible that a model of multi-modality into multi-modality will emerge.
The first AI movie is also expected to appear in 24 years.
According to a Ph.D. student in the Department of Automation at Tsinghua University, "further breakthroughs have been made in multimodal large models, and the ability to generate images and videos has been further improved. More human work, especially one that requires some creativity, is replaced. The emergence ability of large models in some fields is further highlighted, showing some more creative behaviors. 」
AI architect "Chunyang CYang" predicts that 2024 should be the first year for the application of AI large models.
In the whole year of 2023, although large models are popular, there are still very few products that can really land, and they only focus on shallow applications such as rewriting copywriting.
But now, there are many creative products in the field of large models on the ground, and we can expect a wave.
Programmer @Little Fifth Brother predicted——
Large language models will be used to calculate and reason on the mobile phone, agents will do more practical things instead of people, and the best thing is that humanoid robots are likely to help us with laundry, mopping floors, cooking, and tidying up the room!
Resources:
https://writerbuddy.ai/blog/ai-industry-analysis
https://x.com/gdb/status/1741529664856764556?s=20
2023/https://www.zhihu.com/question/635190738
https://www.zhihu.com/question/635190738/answer/3327969527
https://www.zhihu.com/question/635190738/answer/3334551780