There is a useful feature in DeepSeek that isn't present in other commercial LLMs. It displays its internal "thinking" process. I wonder what technological aspect makes this possible. Do several LLMs communicate with each other before providing a solution? Are there different roles within these LLMs, such as some proposing solutions, others contradicting or offering alternative viewpoints, or reminding of overlooked aspects?
Comments
123yawaworht456•7mo ago
>Do several LLMs communicate with each other before providing a solution?
no
>I wonder what technological aspect makes this possible.
one of its training datasets (prioritized somehow over the rest of them) contains a large number of examples emulating the thinking process within <think></think> tags before providing an output. the model then emulates it at runtime.
JPLeRouzic•7mo ago
Thank you for taking the time to answer. However I am not sure the answer is "NO" because DeepSeek has a particular technique in their architecture. To cite this blog [0]:
"Modern large language models (LLMs) started introducing a layer called “Mixture of Experts” (MoE) in their Transformer blocks to scale parameter count without linearly increasing compute. This is typically done through top-k (often k=2) “expert routing”, where each token is dispatched to two specialized feed-forward networks (experts) out of a large pool.
A naive GPU cluster implementation would be to place each expert on a separate device and have the router dispatch to the selected experts during inference. But this would have all the non-active experts idle on the expensive GPUs.
GShard, 2021 introduced the concept of sharding these feed-forward (FF) experts across multiple devices, so that each device"
any model, MoE or not, can be sharded over multiple devices (separate GPUs on a single machine or separate machines via network), yeah. but your question was "Do several LLMs communicate with each other before providing a solution?", and in this context (Deepseek's thinking), the answer is definitely "no".
models can communicate with one another via tool calling, sure, and there are hypothetical workflows where agents delegate tasks to other agents (with inference being done on different models), but that simply isn't the case here.
123yawaworht456•7mo ago
no
>I wonder what technological aspect makes this possible.
one of its training datasets (prioritized somehow over the rest of them) contains a large number of examples emulating the thinking process within <think></think> tags before providing an output. the model then emulates it at runtime.
JPLeRouzic•7mo ago
"Modern large language models (LLMs) started introducing a layer called “Mixture of Experts” (MoE) in their Transformer blocks to scale parameter count without linearly increasing compute. This is typically done through top-k (often k=2) “expert routing”, where each token is dispatched to two specialized feed-forward networks (experts) out of a large pool.
A naive GPU cluster implementation would be to place each expert on a separate device and have the router dispatch to the selected experts during inference. But this would have all the non-active experts idle on the expensive GPUs.
GShard, 2021 introduced the concept of sharding these feed-forward (FF) experts across multiple devices, so that each device"
[0] https://www.kernyan.com/hpc,/cuda/2025/02/26/Deepseek_V3_R1_...
123yawaworht456•7mo ago
models can communicate with one another via tool calling, sure, and there are hypothetical workflows where agents delegate tasks to other agents (with inference being done on different models), but that simply isn't the case here.