Physix FrontierMeta Admits AI Lag Behind OpenAI While Quietly Launching Paid Model Muse Spark 1.1
Let's look at Muse Spark 1.1 first. Business Insider says it might trigger a "price war" in the AI market. I think this prediction is a bit optimistic, but the direction is correct. Meta has always followed an open-source route, giving Llama series models to the community for free to grab ecosystem positioning and get developers accustomed to using Meta's frameworks. But now suddenly charging indicates they realized burning money via open source to buy market share isn't sustainable anymore. Muse Spark 1.1's pricing is reportedly lower than OpenAI's GPT-4o and Google's Gemini. I haven't seen the specific numbers, but if it's billed per token, Meta's choice is quite subtle—they can leverage their massive user data traffic to lower inference costs, something other companies can hardly replicate.
But a "price war" isn't a cure-all. I've participated in over 50 hackathons and seen too many teams switch models just because the API pricing was cheap, only to find high inference latency and weak multimodal capabilities, forcing them to switch back again. If Muse Spark 1.1 is a text model, it faces relatively mature competitors like GPT-4o and Claude 3.5; if it's multimodal, Meta's models have advantages in image generation (based on the Pixel tech they acquired), but language understanding has never been their strong suit. So the key is seeing if hackers can get a demo running within 48 hours—like building an AI customer service prototype or a real-time translation tool. If the API documentation is clear, SDKs support Python and Node.js, and there's a free tier, then it can indeed attract a batch of developers like me to give it a try.
Taking a step further: Zuckerberg's previewed "Watermelon" model. This codename is interesting. Watermelons are juicy and sweet among fruits, but the seeds are annoying to deal with. Meta internally positions this model as "ultimate intelligence," likely having a parameter scale an order of magnitude larger than Llama 3, possibly mixing MoE (Mixture of Experts) architecture. I guess they want to use "Watermelon" to compete against OpenAI's GPT-5 or Google's Gemini Ultra, but the question is: Has Meta really caught up in AI infrastructure?
I noticed a detail: Meta has recently been buying large quantities of H100 chips, but training large models isn't just about stacking compute power; there are also engineering issues like data cleaning, distributed training, and RLHF. OpenAI and Google have years of accumulation. Although Meta has user data, the quality of social data (like posts and comments) is far inferior to professional corpora. If the "Watermelon" model is just a bigger Llama, it might just be an "enhanced version" rather than "ultimate intelligence."
My own judgment is: Meta's current strategy is a typical two-step "open-source lead + paid harvest." Muse Spark 1.1 is responsible for fighting the price war, attracting small-to-medium developers and enterprises, especially in cost-sensitive fields like education, healthcare, and customer service. Meanwhile, the "Watermelon" model is meant to prove their technical prowess in the high-end market, even if it might not be open-sourced and only provided via API. But the problem is that the developer community has developed a strong path dependency on OpenAI's models. Even if Meta is 30% cheaper, as long as inference quality is 5% worse, they might stay in the OpenAI ecosystem.
Finally, a phenomenon I observed in hackathons: When a big company announces they are "behind," it often means they are brewing a big move. Zuckerberg admitting to being behind is likely intended to lower market expectations, so they can suddenly stun everyone with the "Watermelon" model. But whether this "big move" can land depends on whether they can produce a working demo within 48 hours. If the "Watermelon" model launches without even a decent chat interface, it will end up like Facebook's Libra—all thunder, no rain.
Original Link: https://www.ithome.com/0/975/438.htm
