BENGALURU: The low-cost AI model from DeepSeek, R1, is now the most downloaded model across the globe and AI experts are drawing comparison with OpenAI's o1. R1, a reasoning model, is trained with fewer chips compared with other models and so, it is cheaper.
Reasoning models
Big companies are now coming up with reasoning models that cost huge money as these require computing resources to run. These models think and consider various approaches to generate a response. Be it OpenAI o1 series models, R1 or Alibaba Cloud's reasoning AI model QwQ, these perform complex reasoning and arrive at solutions within seconds or minutes. Alibaba has now come up with a new version of its Qwen 2.5 model and it claims that the model outperforms across the board GPT-4o, DeepSeek-V3 and Llama-3.1-405B.
Tech experts point out that many companies will now launch new versions that will be more accurate and also cheaper. Anil Kumar, CTO, Exotel, says, "What we saw with DeepSeek, though shocking, was expected. It follows the familiar pattern of efficiencies coming in after any tech gets popular and crosses the schism of value proposition. This shows that GenAI is working and we are just scratching the surface (think of it as we are in the Nokia 3310 days of the smartphone). The best is yet to come!"
Another reason DeepSeek garnered attention was because of its verbose reasoning that wowed people just like ChatGPT impressed users in its initial days. "OpenAI’s o1 model already did this chain-of-thought-reasoning without being verbose. With models getting cheaper to build and run, value shifts away from chips to higher in the chain to use-case builders," Kumar adds.
All open source models need to be trained on large datasets so as to achieve performance. Are there any risks associated with this? Dan Schiappa, chief product and services officer at Arctic Wolf, says people are already concerned about how much data social media firms have access to, most recently shown by rulings on TikTok. "Just imagine what the risks could be with Chinese Foundational models being trained on all your data. Considering DeepSeek is already limiting its registrations due to a cyber attack, you have to wonder whether they have the appropriate security and policies in place to maintain your privacy,” he adds.
A new Sophos survey showed that 89% of IT leaders are concerned about potential flaws in Gen AI cybersecurity tools that could put their organisation at risk.
Chester Wisniewski, director and global field CTO, Sophos, says, “DeepSeek’s ‘open source’ nature opens it up for exploration - by both adversaries and enthusiasts. Like llama, it can be played with and largely have the guardrails removed. This could lead to abuse by cybercriminals, although it’s important to note that running DeepSeek still requires far more resources than the average cybercriminal has."
He adds that more pressing for companies, however, is that, due to its cost effectiveness, we are likely to see various products and companies adopt DeepSeek, which potentially carries significant privacy risks.
As with any other AI model, it will be critical for companies to make a thorough risk assessment, which extends to any products and suppliers that may incorporate DeepSeek or any future LLM. They also need to be certain they have the right expertise to make an informed decision, Wisniewski says.
Though there are various concerns around privacy and reasoning skills of LLMs that are generally said to be overestimated, tech experts pointed out that one of the highlights of DeepSeek is its ability to build a low-cost model and other companies too will follow the same in coming days.