Things got a whole lot more interesting in the AI wars this week.
It was through we were years away from open source models rivaling what companies such as OpenAI are putting out. The Chinese proved this wrong.
Deepseek R1 was released and it rivals GPT o1 for 98% less money. This is a prime example of the deflationary nature of this entire industry.
While tens of billions are being invested, the cost per unit of artificial congnition is dropping like a rock. This echoes throughout the entire structure, when it is each generation of GPUs from Nvidia or newer models that roll out with more advanced algorithms that train on less compute and data.
These factors are driving companies to keep investing more money, believing in the first mover advantage. Many consider it the race to AGI, something that is hotly debated.
Who will end up winning? To determine that, we probably still have to define the exact parameters of AGI and what we mean.
Can AI "think"?
It seems this is the latest question. These are considered reasoning models. Deepseek R1 was trained utilizing reinforcement learning. It is a method that is similar to how humans learn.
This realm was entered when OpenAi started to discuss their Strawberry project. Many jumped on the bandwagon claiming it was AGI in the making. So far, that has not turned out to be the case. Nevertheless, progress is being made.
At stake is a lot of money. If we consider the implication of an AI system that can make reason through problems and make decision, we can only imagine what that means to jobs. It also opens up a massive possibility for monetization, something Sam Altman and OpenAI were after.
I under the past tense since that idea was shot to hell.
Deepseek, according to many who tested it, rivals OpenAI o1. This difference is it does not come with a $200 per month price tag. Here is where OpenAI's path to riches was thwarted.
This is in alignment with my thinking that these models will simply become commoditized. When that happens, the money goes out of something.
Why is someone going to pay $200 per month when something is just as effective for a lot less money?
Of course, the eventual results remain to be seen. In fairness, it is rumored that OpenAI will release o3 soon, sparking a belief that it will jump back to the top of the list.
The cat is now out of the bag.
We are witnessing massive steps in the open source community. Meta took the lead by opening up the weights on its model, Llama. Now we have a Chinese company that pushed the value system of openness and public use across the world.
It is actually making AI accessible to more people. As more models show up, developers will grab then and tinker around. Additions will be released, expanding the entire ecosystem.
The AI game is still very costly. Training LLMs requires billions of dollars in hardware. It also needs a great deal of data, something that comes with a huge price tag.
Fortunately, advancement is algorithms is getting better results with less data. Some believe the future power of the technology is in the small language models as compared to the behemoths Big Tech is producing.
Things should get very interesting when applications and platforms are built atop many of these models. The open source world offers the availability to anyone with the talent. How will this compare to what Big Tech builds on their LLMs should be interesting to watch.
The pace of this sector is so rapid that what I write today has a shelf life of a few months. Deepseek is the top banana for the moment. There is a good chance that another open source model will arrive within the next 6 months. Of course, to get noticed, it is going to have to usurp what this generation achieved.
