English version published 2026-11-08 with AI assistance. Attribution, source notes, and verification limits are retained in the text. 中文原文
Full text — English translation
Fact-check note (2026-10-28):
1. Med-PaLM (Singhal et al.) scored about 67.6% on MedQA (U.S. medical licensing exam-style questions), above the 60% passing mark. The paper does not claim it performs on par with clinicians. Source:
https://www.nature.com/articles/s41586-023-06291-2
2. GPT-3 training cost estimates: Lambda put it at about $4.6 million; Rethink Priorities estimated about $12 million (range $5 million to $33 million). Sources:
https://lambda.ai/blog/demystifying-gpt-3 https://rethinkpriorities.org/research-area/the-replication-and-emulation-of-gpt-3/
3. Riffusion: released December 15, 2022, fine-tuned from Stable Diffusion to generate music from spectrograms. Source:
https://techcrunch.com/2022/12/15/try-riffusion-an-ai-model-that-composes-music-by-visualizing-it/
4. Science’s 2022 Breakthrough of the Year list included creative AI. Source:
https://www.science.org/content/article/breakthrough-2022
5. Trillion-parameter models: BAAI’s WuDao 2.0 has 1.75 trillion parameters; Alibaba’s M6-10T has 10 trillion (both 2021). Sources:
https://baai.ac.cn/en/research https://arxiv.org/html/2110.03888
6. Based on the sources above, claims without a reliable source were removed and some statements were revised to verifiable versions (4 changes in total).
----
Robin: The opportunity for AI startups isn’t building the underlying platforms. Today’s wave of SaaS startups doesn’t build cloud infrastructure either. They build on top of the big public clouds and grow all kinds of services out of their own specialized industry knowledge.
On top of the platforms the big tech companies have already built, we’re definitely going to see a lot more interesting AI applications that are closer to how people actually work and live.
---
Science magazine’s list of the top ten breakthroughs of 2022 included creative AI. And AI has broken into the mainstream in a big way, from ChatGPT to text-to-image generators. The last time AI was this hot was 2016, when AlphaGo beat Lee Sedol.
This time around, a lot of people have two worries. First, will this wave actually create new business opportunities, or will it be like 2016—hype a concept, then cool off? Second, if there are opportunities, how do we capture them?
Peng Tianfang (a Dedao contributor) gathered a large body of Chinese and international research and reporting from the past two months.
First, are there real business opportunities? Definitely.
That’s because this generation of AI is in a completely different league from AlphaGo in 2016. AlphaGo was great at computation but pretty mediocre at creativity.
The 2022 wave of AI jumped up a whole level in creativity. In technical terms, new algorithms like attention-based Transformers and diffusion models outperform the earlier neural network approaches—convolutional networks (CNNs), recurrent networks (RNNs), generative adversarial networks (GANs), and so on.
You don’t need to remember the jargon. Just remember the conclusion: this wave of AI is remarkably good at creative tasks involving text, images, and audio.
More importantly, this wave is only getting started. A lot of the potential—especially of new algorithms like Transformers and diffusion models—is far from fully tapped. If it all gets unlocked, the opportunity is enormous.
Put simply, the AlphaGo wave sold a concept and built confidence. This wave of AI actually works.
(Robin: What matters even more is that anyone can learn to use it, not just a small circle of software engineers.)
For now, though, these opportunities are in the hands of the tech giants, because the AI models behind these new algorithms are huge. A big model means training it takes enormous amounts of data, a very long time, and a very high cost (massive compute).
Take chatbots. Before 2019, top-tier language models had parameters on the order of a billion. In 2020, OpenAI’s GPT-3 jumped to 175 billion. By 2022, companies in China and abroad were holding launch events for trillion-parameter models.
How expensive is that scale? A model has to be trained many times. For GPT-3 in 2020, industry estimates put the cost of a single training run at somewhere between a few million and more than ten million dollars. And training is uncertain—people in the industry joke that it’s like alchemy, and you need some luck. Only giants worth hundreds of billions of dollars can afford to roll those dice.
So does that mean newcomers have no shot? No.
Shortly after ChatGPT came out, well-known Silicon Valley investor Chamath Palihapitiya argued on his video channel that as large models get more powerful and more expensive, AI could develop a distinctive business model called Model as a Service, or MaaS. Academics and entrepreneurs in China and abroad have been discussing the model a lot lately.
Here’s how MaaS works. A big company develops an AI model as the base layer. Think of it as a new grad with an excellent general education—perfect scores in math, physics, and chemistry, rock-solid fundamentals. Great foundation, but still an outsider in any specialized field. If you want it to do specific, highly specialized work—medicine, education, sales—it needs more training.
MaaS splits the training process into two stages.
The general education stage takes massive investment; you have to learn everything. The giants handle this. They’re the model creators, handing you an exceptionally smart generalist with great fundamentals.
In the second stage, you turn that generalist into a specialist who can do real work in a particular field. The tech giants aren’t good at this, so it falls to companies in each vertical. The people who do this vertical training are called model refiners.
This second round of training for specific verticals is the opportunity for new AI players. Under MaaS, the tech giants will open up APIs to their models. Smaller players who understand their industries more deeply will pay to use those APIs and refine the models into applications that truly meet their industry’s needs.
That way everyone wins, from giants to newcomers. The giants carry the huge cost of building the models, but they collect API fees from many specialized industries and avoid the risk of betting too heavily on any one of them. Meanwhile, people in each vertical don’t have to build from scratch and can use AI models at much lower cost.
AI startups have already started putting this into practice. Peng Tianfang looked into it and found a few examples of MaaS in action.
The first is in music. Two engineers released Riffusion, a music tool that generates a clip of music from a text prompt. What’s more interesting is that they didn’t build it from scratch. In their spare time, they took an open-source AI image model, fine-tuned it, and turned it into a music tool.
For example, type “electric guitar rock solo” and you get a clip like this.
Type “funky sax jazz solo” and you get a clip like this.
You’re probably wondering how an AI model built to draw pictures ended up making music. The key is the spectrogram. Without getting into the technical details, any piece of audio can be represented as an image called a spectrogram. The two engineers trained the AI on huge numbers of spectrograms, and that gave it the ability to compose.
The second example is in medicine. Google released a model called Med-PaLM, which for the first time passed the threshold on multiple-choice questions styled after the U.S. medical licensing exam. It could play a bigger role in AI-assisted diagnosis in the future and lower the cost of care.
This AI doctor also followed the two-stage training approach. It started with an existing foundation model and added a training process built specifically on medical data. Two rounds of refining produced the result.
Finally, a question to think about: in your industry, what interesting applications could come from combining industry knowledge with large AI models?
#DedaoHeadlines20230117
These articles are personal notes, book excerpts, and collected reprints. Figures and events reflect the time of the original publication, are not current information, and do not constitute investment advice.
Comments
What do you think? Share your experience or questions.
Sign in to comment, reply or like. Sign in / Sign up
Loading comments…