Report: The current state of AI in 2023 from the perspectives of research, industry, policy, etc

Original title: "2023 State of Artificial Intelligence Report"

Author: 36Kr's compilation team Divine Translation Bureau BONI

Image source: Generated by Unbounded AI

Maybe next year, AI-generated songs will hit the Billboard Top 10 chart

Editor's brief: For our increasingly digital, data-driven world, AI is a force multiplier of technological advancement. Therefore, it is important for our work to understand the current state of development of artificial intelligence. This "2023 State of Artificial Intelligence Report" summarizes the current state of artificial intelligence from the aspects of research, industry, politics, security, etc., and makes predictions for the development of artificial intelligence in the next 12 months, hoping to help you understand the development dynamics of artificial intelligence. The article is from the compilation.

Research

2023 is certainly the year of large language models (LLMs), and OpenAI's GPT-4 has shocked the world, successfully beating all other LLMs – both on classic AI benchmarks and on exams designed for humans.

GPT-4's capabilities crush other large models: OpenAI has not only tested it with classical natural language processing benchmarks, but also with some tests that evaluate humans (such as bar exams, GRE, force deductions, etc.); GPT-4 also performed better than the previous model on hallucination issues

Due to concerns about security and competition, we have found that AI has become less open to it. Regarding GPT-4, OpenAI only released a technical report with very limited information, Google did not disclose much about PaLM2, and Anthropic did not disclose any technical information, whether it was Claude... Or Claude 2.

Whether it's a tech giant or a start-up, leading companies are getting shy about the details of their AI technology

However, Meta AI and other companies have stepped up to keep the flame of open source burning by developing and releasing an open-source LLM that rivals GPT-3.5's many features.

Meta open-sourced LLaMa, thus setting off an open-source race for large models, with the help of open-source models, some people began to fine-tune the models and develop applications for vertical fields

Looking at Hugging Face's leaderboards, open source is more active than ever, with downloads and model submissions soaring to all-time highs. Notably, the LLaMa model has been downloaded more than 32 million times on Hugging Face in the last 30 days.

Hugging Face has become a temple for open-source AI, with significant growth in the number of datasets, spaces, and models in 23 compared to '22

While we have a lot of different (mostly academic) benchmarks for evaluating the performance of large language models, the biggest thing these different evaluation criteria seem to have in common, and the biggest science and engineering benchmark, is this: "resonance"

With the increase of open-source and closed-source language models, and with the similarity of training data, there is a lack of differentiation between LLMs, which makes it difficult to evaluate models. The current mainstream benchmark for comparing the model's capabilities is Stanford's HELM leaderboard with Hugging Face's LLM Benchamark, but users seem to prefer a more subjective approach: resonance.

In addition to the exciting atmosphere at LLMs, researchers, including Microsoft, have been exploring the possibilities of small language models, finding that models trained with highly specialized datasets can rival competitors that are 50x larger.

Microsoft has found that small language models can compete with models that are up to 50 times larger when trained on very specialized and carefully selected datasets.

If the team at Epoch AI gets it right, this work could become even more urgent. They predict that we run the risk of running out of stock of high-quality language databases in the next "two years," leading labs to explore alternative sources of training data.

Some research teams believe that human-generated data is running out, with low-quality linguistic data estimated to run out between 2030 and 2050, high-quality linguistic data aggregated by 2026, and visual data between 2030 and 2060.

Look at the status quo at a higher level – although it has diminished in recent years, the U.S. still leads the way, and the vast majority of highly cited papers still come from a small number of U.S. institutions.

China ranked second in the field of artificial intelligence research

Industry

All of this work means it's a great time to get into the hardware business, especially if you're NVIDIA. GPU demand has propelled them into the trillion-dollar club, with chips being used 19 times more in AI research than "other alternatives combined."

Artificial intelligence research mainly uses NVIDIA's chips. Note: The scale of the y-axis is an exponential change

While NVIDIA continues to introduce new chips, their older GPUs are showing exceptional lifetime value. Released in 2017, the V100 is the most popular GPU in AI research papers of 2022. This CPU may be discontinued in 5 years, which means it has been in service for 10 years.

NVIDIA's V100 chip shows strong vitality

We've seen demand for NVIDIA H100 grow rapidly, and labs are rushing to build large computing clusters—and there may be more on the way. However, we have heard that these construction projects have not come without major engineering challenges.

The AI computing power cluster built with NVIDIA's latest GPU H100, the largest Google A3, uses 26,000 GPUs

The "chip wars" have also forced the industry to adjust, with NVIDIA, Intel and AMD all building special, sanctions-compliant chips for their large Chinese customer base.

Some of the chips in this picture have been added to the control list by the United States

Perhaps the most unsurprising news of all time is this: Chat-GPT is one of the fastest-growing internet products of all time. It's especially popular among developers, and has replaced Stack Overflow as a new place for developers to find solutions to coding problems.

The rise of ChatGPT stands in stark contrast to the decline of Stack Overflow

But according to Sequoia Capital, there are now reasons to doubt the staying power of generative AI products — with inconsistent retention rates across everything from image generation to AI companions.

The general public's interest in AI products seems to be on a whim

In addition to the consumer software space, there are signs that generative AI can accelerate progress in the field of physical AI. Wayve GAIA-1 has demonstrated impressive versatility and can be used as a powerful tool for training and validating autonomous driving models.

GAIA-1 uses video, text, and action inputs to generate realistic driving scenarios to train AI to respond to extreme situations

In addition to generative AI, we're also seeing significant moves from industries that have been struggling to find suitable applications for AI. Many traditional pharma companies have bet all their bets on artificial intelligence, struck multibillion-dollar deals with companies like Exscientia and InstaDeep.

Mainstream pharmaceutical companies are starting to devote themselves to AI-assisted drug development

As militaries rush to modernize their forces to deal with asymmetric warfare, the AI-first defense market is booming. However, the conflict between new technologies and established players makes it difficult for new entrants to gain a foothold.

Last year, U.S. defense start-ups raised $2.4 billion

In addition to these successes, the VC industry is focused on generative AI, a sector that holds up a slice of tech private markets like Atlas. If it weren't for the market boom in generative AI, investment in AI would be down 40% from last year.

The world's most important investment in artificial intelligence is relatively stable, and generative AI has become the new darling of investment

The authors of the paper that first introduced Transformers' neural networks are living proof that in 2023 alone, the Transformer Gang has secured billions of dollars in funding.

The authors of "Attention is All You Need" have since left Google to start their own businesses, and have collectively raised billions of dollars in funding

The same goes for Baidu's DeepSpeech2 team, an artificial intelligence lab in Silicon Valley. Their work on deep learning for speech recognition shows us the law of augmentation that now underpins large-scale artificial intelligence. Most of the members of this team went on to become founders or senior executives at leading machine learning companies.

The early work of Baidu's Silicon Valley AI Lab is a testament to how scale wins

Many of the most high-profile blockbuster funding rounds were not led by traditional VC firms at all. 2023 is the year of corporate venture capital, with big tech companies making effective use of the war funds they have on hand.

The main force of venture capital has become the tech giants

Politics

Not surprisingly, billions of dollars of investment, coupled with huge leaps in capabilities, have put AI at the top of the agenda for policymakers. The spectrum ranges from loose to tightly regulated, and there are several approaches to regulation around the world.

From lax to strictly regulated, from leveraging existing legal frameworks to formulating targeted policies, countries have different attitudes towards regulating AI

There have been a number of potential proposals for global governance of AI, mainly by a range of global organizations. The UK AI Security Summit, organized by Matt Clifford and others, may help to concretize some of these ideas.

The global governance of AI is still in its early stages

As we continue to see the power of AI on the battlefield, these debates are likely to become even more pressing. The conflict in Ukraine has become a laboratory for AI warfare, demonstrating how even relatively improvised patchwork systems can have devastating effects when cleverly integrated.

Ukraine has become a testing ground for AI warfare

Another potential flashpoint is next year's U.S. presidential election. So far, deepfakes and other AI-generated content have played a relatively limited role compared to the kind of disinformation of the past. But a low-cost, high-quality model could change that, prompting preemptive action.

There is a good chance that AI could be used to interfere in the election

Previous State of Artificial Intelligence reports have warned that large labs may be ignoring the safety of AI. In 2023, the debate about whether humanity is at risk for survival because of AI has intensified, and the debate between researchers over open source versus closed source has intensified, and extinction risk has made headlines.

As artificial intelligence (AI) has demonstrated its amazing capabilities, some experts have become concerned about the risks to humanity's survival

...... Needless to say, while not everyone agrees – Yann LeCun and Marc Andreessen are the main skeptics.

When it comes to extinction risk, experts are divided into two camps

It's no surprise that policymakers are only now alarmed by the potential risks of AI, although they have struggled to understand it. The United Kingdom has spearheaded the creation of a dedicated Frontier AI Taskforce, led by Ian Hogarth, and the United States has launched a congressional investigation.

Governments around the world are starting to pay attention to artificial intelligence

While there are still theoretical disputes, labs have already begun to take action, and Google DeepMind and Anthropic were among the first to elaborate on their approach in more detail when it comes to mitigating the extreme risks of development deployments.

Large labs have already begun to take action to mitigate the risk

Even without touching on the distant future, tricky questions are being asked about techniques such as reinforcement learning based on human feedback (which underpins technologies like Chat-GPT).

Reinforcement learning based on human feedback faces some fundamental challenges

Prediction

As always, in the spirit of transparency, we take care of last year's predictions – we scored 5/9.

✅ LLM training, generative AI/audio, tech giants are fully invested in general AI R&D, alignment investments, and training data.

❌ Doom for multimodal research, biosafety lab regulation, and semi-finished start-ups.

Here are our 12 predictions for the next 10 months! These include:

• Generative AI/Filmmaking

• Artificial Intelligence and Elections

• Self-improvement agents

• The return of IPOs

• Models worth more than $1 billion

• Competition investigations

• Global governance

• Bank + GPU

-Music

• Chip acquisitions

1. Hollywood-grade productions will use generative AI to produce visual effects 2. There will be a generative AI media company under investigation for misusing generative technology in the 2024 U.S. election 3. The performance of self-improving AI agents in complex environments (e.g., AAA games, tool usage, science, etc.) will crush the latest technology 4. The tech IPO market will be unfrozen, and at least one AI-focused company will go public (e.g. Databricks) 5. The cost of training large models for generative AI could soar to more than $1 billion 6. The FTC in the U.S. or UMA in the U.K. will launch a competition investigation into the Microsoft/OpenAI deal 7. With the exception of volunteering, there will be limited progress in global AI governance 8. Financial institutions will replace venture capital equity capital and launch the GPT debt fund to finance computing power 9. AI-generated songs will hit the Billboard Top 10 Hits chart or Spotify Hits 2024 10. As inference loads and costs skyrocket, a large AI company (such as OpenAI) acquires an AI chip company for inference