From: redpointai
The relevance and adoption of open source models have been a significant point of discussion and surprise in the AI community. While hobbyist and local Llama communities show enthusiasm, enterprise adoption has been surprisingly low and declining.
Surprising Lack of Enterprise Adoption
The general sentiment among enterprise users regarding open source models is that they are “cool” but have not fundamentally altered the adoption path of AI within businesses [00:02:46]. A specific estimate from Anker of Brain Trust suggests that open-source model usage in enterprises is around 5% and potentially decreasing [00:02:59].
This limited enterprise adoption is attributed to several factors:
- Use Case Discovery Mode: Enterprises are primarily focused on discovering practical use cases for AI, prioritizing the most powerful models available, regardless of their open-source status [00:03:11].
- Rapid Model Evolution: The constant emergence of new model generations necessitates continuous re-evaluation and discovery, making it difficult for enterprises to commit to specific open-source models [00:03:22].
- Licensing Debates: While discussions around licenses are common, they haven’t translated into significant shifts in AI adoption patterns [00:02:50].
Open Source Catch-Up and its Nuances
Despite the adoption challenges, there was initial surprise at how quickly open source models caught up to proprietary ones in terms of performance [00:03:32]. It was initially unclear if closed-source models would maintain a compounding advantage, but the gap has felt like it’s shortened [00:03:37].
However, this “catch-up” is nuanced:
- DeepSeek’s Role: The rapid advancement attributed to open source is largely due to the efforts of specific entities like DeepSeek, rather than a collective “team open source” [00:04:11]. There is now evidence that DeepSeek may cease open-sourcing its models [00:04:16].
- Distillation vs. Original Training: Much of the perceived “catching up” involves distilling knowledge from models like DeepSeek, which is an easier task than training competitive models from scratch [00:04:30].
- Replication vs. Innovation: Replicating existing models is significantly cheaper than creating fundamentally new ones [00:04:48]. While DeepSeek was a strong new entrant, it didn’t fundamentally introduce new concepts [00:04:59].
Unique Contributions of Open Source Models
Despite these challenges, some open-source models have offered unique contributions. The R1 model, for instance, was noted for having “full traces,” which is considered a net unique feature in the open-source ecosystem [00:05:15].
Challenges for New Model Players
The continued emergence of new companies focused on training models is viewed as “overhyped” given the current landscape [00:16:14]. While there might be opportunities for new model players, a clear rationale for what they would offer that is currently missing is not apparent [00:16:26]. Many new model companies aim to pursue AGI or hit all benchmarks, which is not sustainable for all [00:16:39].
“If there is something like that that these companies can latch on to and being their sort of known for being the best at maybe there’s a case for that. Largely though I do agree with you that I don’t think there should be at this point more model companies.” [00:17:21]
The prevailing sentiment suggests that a general-purpose model is the optimal approach, and training hyper-specific models, while potentially cheaper and faster, typically does not yield higher quality [00:19:16]. Historical examples, such as the Bloomberg GPT model, showed that general-purpose models often surpass specialized ones, leading to the conclusion that “closed models were better” for many applications [00:20:00].
[!INFO] Exception: Specialized Models for Unique Datasets There is still perceived opportunity for specialized models when they leverage unique datasets, such as those found in robotics, biology, or material science, where substantial new data is required [00:17:34]. These fields involve wet labs and extensive experimentation, providing distinct data advantages [00:17:53]. However, the core LLM agent space remains challenging for new entrants due to competition from larger players [00:18:02].