Nowadays, all kinds of big model news are overwhelming. We are enveloped in such information confusion, and our thinking is easily divergent and unfocused. In the end, it is difficult to form a focused point of view. We feel that this is right and that is not bad.
In this article, let’s sort it out:
Viewpoint 1
Objective data shows that the usage growth, revenue and profit margin of large artificial intelligence models are not satisfactory.
- According to Sequoia Capital estimates, in 2023, major model R&D companies will spend as much as US$50 billion purchasing hardware such as GPUs from NVIDIA, resulting in revenue of only about US$3 billion.
- The valuations of big-model startups are already higher than they should be, and low gross margins are a fact that cloud providers are cutting back on.
- Some large model startups are facing dissolution. The former CEO of Microsoft even said that the future of the well-known AI large model StabilityAI is not clear.
(On April 9, an autonomous driving company invested by OpenAI has now closed)
- GPTS has gradually been proven to be a complete failure. When people’s enthusiasm subsided, everyone has basically understood this “thing”. It is an interesting thing, but it does not significantly improve productivity. Even if it can improve productivity, it is not stable. , the results are not always as expected.
- The general public only knows ChatGPT. The vast majority of people use the outdated GPT3.5 version and are unwilling to pay to use higher versions such as GPT-4 and Claude3. Among this group of people using GPT3.5, some people unscrupulously use AIGC content to corrupt the content creation of the Internet. This is becoming a big problem!
Does the above mean that the AIGC large model has failed? Maybe, but maybe not!
Viewpoint 2
And some more data:
- Large models are still the main focus of the current news media, and the reading volume is still maintained at a relatively high level.
- Although the valuation of large-model startups is relatively high, it is also a good phenomenon to return to reasonable values.
- The group of people who really use ChatGPT to improve productivity are mainly programmers.
- There are still large models being launched. For example, Anthropic released Claude 3 in early March, which recently surpassed GPT-4 for the first time in the field of LMSys chatbots; more products will be released this summer, which may include OpenAI’s GPT-5 , Apple’s new Siri, Meta’s Llama 3, Google’s Gemini 1.5 Ultra and xAI’s Grok 2.
- AIGC has not stopped. Just last week, Databricks disclosed that DBRX is the best open source model so far; AI21 labs’ Jamba is the best production-level SSM converter (with a novel architecture); OpenAI’s Voice Engine is a text A speech-to-speech model that generates custom realistic speech with 15 seconds of audio.
- In addition, underlying scientific research is still being invested, with Microsoft and OpenAI planning to build a $100 billion supercomputer infrastructure project in the next few years;
So, what stage is this wave of AIGC large models in?
Bottom line: The big model hype has died down, but the revolution continues.
In the historical industrial revolution, printing presses, electricity, and the Internet all developed in this way. For example, in 2000, Internet companies collapsed and online commerce was sluggish, but after many years, it formed a revolution again.
The development of generative AI is likely to be similar: early turmoil mixed with enthusiasm, followed by lull, and finally resurgence.
From the current limited perspective, we may not be able to prove or disprove the revolutionary nature of generative artificial intelligence at all. Perhaps, to put it more cliché: time will tell. It won’t be long~