Op-Ed: DeepSeek V4 strikes – China accused of ‘distillation’ but some heavy duty metrics as well

DeepSeek has prompted some experts to advise tech entrepreneurs to make sure they are in position to change which artificial intelligence model they rely on as costs and capabilities evolve rapidly – Copyright AFP/File Hector RETAMAL

AI-saturated headlines notwithstanding, the fan has been hit hard by DeepSeek V4 in multiple contexts. This thing is generating a lot of disingenuity at the national level, with China being accused of “distilling”, aka swiping and making cheaper and/or mini versions of AI.

Ancient readers clutching stone tablets and flying to safety on their Pteranodons will recollect when the original DeepSeek upended the whole idea of AI costs and performance. This time, it’s much more worrying and comes with some interesting specifications built in.  

Let’s start with the plodding “distillation” routine, though. This is a more or less standard and utterly useless form of defence. Competition alone can encourage distillation or even, believe it or not, better research. Common sense demands cost efficiency. Why wouldn’t you try to make something better and cheaper?

In the bottomlessly sincere, dewy-eyed Disney-like theme park imaginatively called IPLand, there’s more at stake than just better runtime. The IP is hardwired, pun intended, into the new values created by these cost and operational efficiencies.

A bit “Duhh-esque”, you think?  Too damn straight. This is where the obvious gets dangerous. A little digging, meaning surprisingly little, says a bit more about DeepSeek V4:

Ghacks.net indicates DeepSeek V4 has a range of performance and cost initiatives, many or which simply look like natural evolutions. Never mind the innuendo. Check out the specs and ask yourself how and why anybody wouldn’t add these values. These specs are sales pitches of themselves.

Using 9.5x less memory is a kick in the teeth for the rather placid image of AI performance, too. You’d swear granny was making cupcakes in her Easy Bake LLM. Transfer those reduced operational runtime costs and hardware impacts to any operation. Good value? Undeniably so.  

Less obvious and perhaps more threatening is that this iteration of DeepSeek comes with an existing support framework, Huawei Ascend, and runs on Huawei chips. The support framework is a game-changer in the sense that it’s a workable business model for commercial uptake.

AI is still largely in its mystic stage for performance demands. DeepSeek has just made a big point for the future of AI overall.  

Equally to the point, the US has now been directly and incisively challenged regarding its now-rather-monotonous AI supremacy jingle. Squeaking about distillation is meaningless.

DeepSeek has just reminded the world that there are other big players in AI. What’s the response?