AI (artificial intelligence) is a multidisciplinary field of science and engineering aimed at creating intelligent machines. As our world becomes increasingly digital and data-driven, it may be a force multiplier for technological progress. This is because we can find products of intelligence everywhere, whether in a culture or consumer products.
This time, we provide you with a report about the State of AI. If you want to read it whole, check out this page.
Key Takeaways from Report
Research
- The diffusion model revolutionized computer vision by generating text-to-images with impressive accuracy.
- Using artificial intelligence, scientists are tackling more science problems, including plastic recycling, nuclear fusion reactor control, and the discovery of natural products.
- Scaling laws refocus on data: perhaps model scale isn’t enough. So a single model to rule them all is being developed.
- Large models are being open-sourced at breakneck speed, enabling collectives to compete with large labs.
- As AI research becomes more neuroscience-inspired, its approaches are beginning to resemble cognitive science.
Industry
- Can upstart AI semiconductor startups compete with NVIDIA? According to usage statistics in AI research, NVIDIA is 20 to 100 times ahead of its competitors.
- Large tech companies expand their AI clouds and partner with AGI startups.
- Many startups, including DeepMind and OpenAI, were formed as a result of hiring freezes and the dissolution of AI labs.
- There are 18 clinical assets owned by major AI drug discovery companies, and the first CE mark for autonomous medical imaging diagnostics has been awarded.
- Big tech and startups are quickly translating AI for code research into commercial developer tools.
Politics
- Academia and industry are at odds in large-scale AI work: almost 0% of work is done in academia.
- Decentralized research collectives funded by non-traditional sources are taking over from academia.
- In earnest, the Great Reshoring of American semiconductor capabilities has begun, and geopolitical tensions are at an all-time high.
- A greater number of defense product categories incorporate AI, and defense AI startups receive even more funding.
Safety
- Despite increased awareness, talent, and funding, AI safety research still lags behind capabilities research.
Research
The research shows DeepMind’s breakthroughs in the physical sciences. The company has since made significant advancements in both mathematics and materials science.
In its proposal, DeepMind described an iterative workflow involving mathematicians and supervised machine learning models (usually a NN). As data becomes available, mathematicians can refine their hypotheses and/or generate more data until the conjecture holds.
What’s more, UT Austin researchers engineered an enzyme capable of degrading PET, a type of plastic responsible for 12% of global solid waste.
Source: State of AI report
Additionally, Minecraft is used by OpenAI as a testbed for agents that interact with computers. Through a small amount of labeled mouse and keyboard interactions, OpenAI trained a model (Video PreTraining, VPT) to play Minecraft from video frames.
Source: State of AI report
It’s also interesting that corporate AI labs invest heavily in AI for code research. For example, a key feature of OpenAI’s Codex, which drives GitHub Copilot, is its ability to complete code on multiple lines or directly from natural language instructions. In response to this success, Salesforce, Google, and DeepMind have conducted more research in this area.
Source: State of AI report
The 2020 State of AI Report predicted that transformers would expand beyond natural language processing to achieve state-of-the-art computer vision. Transformers are now clearly a candidate for general-purpose architectures. Transformer-related papers in 2022 illustrate how ubiquitous this model architecture has become.
Source: State of AI report
Industry
In FY 2021, NVIDIA’s datacenter revenue reached $10.6B. In Q4 2021, they recognized $3.26B, which is more than the combined valuation of the top-3 AI semiconductor startups. In AI research papers, their chips are the most popular…and by a huge margin. In terms of usage, GPUs are 131x more common than ASICs, 90x than chips from Graphcore, Habana, Cerebras, SambaNova, and Cambricon combined, 78x than Google’s TPU, and 23x than FPGAs.
Source: State of AI report
In recent weeks, hyperscalers and challenger AI compute providers have partnered up on major AI compute projects, including Microsoft’s $1 billion investment in OpenAI.
Source: State of AI report
Companies build supercomputers that are larger than national ones in a gold rush for computer power.
Source: State of AI report
All the authors of the landmark paper that introduced transformer-based neural networks have left Google to start their own startups in AGI, conversational agents, AI-first biotech, and blockchain.
Source: State of AI report
The number of AI-driven drug discovery companies in clinical trials has increased from zero in 2020 to 18.
Source: State of AI report
As the broader market slows down, investment in AI startups has slowed down as well.
Source: State of AI report
Politics
In large model AI, industry and academia are separated by a widening computing chasm. In order for the AI community to scale models, this chasm of “haves” and “have nots” creates significant challenges for AI safety, pursuing diverse ideas, concentrating talent, and more.
Source: State of AI report
Stability AI is attempting a new paradigm in commercializing open-source AI. For large-scale projects, there was previously a reliance on ad-hoc compute donations. Stability is pioneering a new approach to providing structured computing and resources to open-source communities, as well as commercializing these projects with revenue-sharing.
Source: State of AI report
As the US lags behind in new fab projects, which take years to build, the Great Reshoring will take time. For example, China accelerated its greenfield fab project output by almost 7x between 1990 and 2020, while the US slowed down by 2.5x.
Source: State of AI report
Safety
In acknowledging these uncertain but catastrophic risks, the UK is taking the lead. Non-aligned AGI poses a long-term threat to the UK and the world, as well as unforeseeable changes.
AI safety is becoming increasingly important to researchers. In recent years, researchers have become more concerned about the risks of superhuman AGI and human-level AI, which were once dismissed as science fiction by mainstream AI research and academia.
Source: State of AI report
The increasing awareness of AI existential risk has led to an increase in headcount, with 300 researchers now working full-time on AI safety. In contrast, there are still many fewer researchers in the broader field (correct chart), which is growing faster than ever.
Source: State of AI report
Over to You
We can talk and talk for hours about AI, but we can read reports like this one, which provides readers with a lot of research conclusions and equips them with a wealth of knowledge.
Sign up for our newsletter and stay up to date with the news!
