| LurkerJ said: Yes, those advancements happen in cycles and sometimes they happen outside the bigger companies you listed… by startups that quickly get acquired with no meaningful pushback, that’s why “the market” isn’t usually spooked. But a Chinese company that we can’t buy, open source since day 0, a model that can be run locally if desired, this is highly unusual and the closest we have to an actual openAI, ironically. Even when the dust settles, this will always be huge. No wonder so many American CEOs had to say something before they went to sleep. |
The Chinese part is what is unusual, yes, although it was always a good possibility Chinese companies would catch up. But open-source isn't unusual. Open-source base-models have been in a "good enough" phase for a while now (including Deepseek's models up until this point.) I've been using them in my job (as an MLE/Data Scientist) for nearly two years, because they allow for more control over data-security in a way connecting to a web-based API doesn't and they are usually (not always, Google has a very cheap API, even cheaper than Deepseek's actually) cheaper.
For the use-cases that most people are using LLMs for in the business world such as converting unstructured data -> (semi-)structured data, automating the building of knowledge graphs, documentation, transcription, machine translation, text summarization, RPA, etc these smallish base-models have been "good enough" for a while now.
This is why the cloud-based industry that are trying to sell AI as a service have been moving towards two general goals: 1. building tools that make the base models more useful within an ecosystem (video chat, operators/agents, projects, MCP, etc) -- basically becoming what Google Search is for web-search but for AI, and 2. moving on to reasoning A.I that can help create feedback loops for research.
Most of these large-scale advancements are going into #1. Providing a set of services to hundreds of millions of customers costs a lot of inference and therefore a lot of compute, much more than training the models do, but it is these ecosystems that are going to be monetizable in the cloud (or so these for-profit companies hope.) #2 (outside of OpenAI, as they are the leaders and want to gate-keep) has been a much more collaborative effort between different actors. Anthropic, Google, Nvidia, Meta, Mistral, Deepseek, Microsoft (outside their partnership with OpenAI), Qwen (the other big Chinese open-source group, a subsidiary group of Alibaba), etc all release (to varying degrees) quite a bit of details about what they're doing. #2 belongs to no country and likely won't be monetizable other than the fact it pushes the industry forward. Think of #2 like compiler of old. Compilers used to be primarily proprietary, but eventually open-source won.
As for yesterday's reaction, I think it is just that a reaction from low-information investors following the herd. Regardless of who is building these models and how efficient they get unless the CUDA-competitor problem is solved, Nvidia is going to be in high demand. More efficient models still follow the neural scaling laws and Jevons Paradox applies here as much as it does to any other basic infrastructure.
Essay Pro