In an era where everything seems to be getting bigger\u2014from wider\nscreens and taller buildings to the occasionally inflated egos of\nofficials\u2014the world of artificial intelligence is taking a quiet,\nopposing path: shrinking itself to become stronger.<\/p>\n
A discovery from Google\u2019s research team introduces something called\nTurboQuant, which, if translated into everyday coffee shop lingo,\nroughly means: \u201cA slim AI, but with an even sharper mind.\u201d The results\nare not just intriguing. They nearly defy common sense.<\/p>\n
Let us start with the most basic elements, so as not to get lost in\nthe thicket of technical terms that can sometimes be denser than the\nAmazon rainforest.<\/p>\n
AI, in essence, never reads words. It does not know if you write\n\u201cking\u201d, \u201clove\u201d, or \u201cnational debt\u201d. What it sees are numbers, sometimes\nthousands of them, forming what is called a \u201cvector\u201d.<\/p>\n
A simple word can become thousands of coordinates in a\nhigh-dimensional mathematical space. The more complex its meaning, the\nlonger the list of numbers. Thus, every conversation we have with AI is\nactually a transaction of vast numbers.<\/p>\n
The problem is straightforward: those numbers are expensive. Not\nmorally expensive, but expensive in terms of memory. Every conversation\nis stored in what is called a KV cache\u2014a kind of \u201cdigital cheat sheet\u201d\nso the AI does not have to reread the entire record each time it\nresponds.<\/p>\n
But like a student\u2019s notebook that is too diligently filled with\nnotes, these notes eventually cover the desk, then spill onto the floor,\nthen make the system gasp for breath. This is where the biggest\nbottleneck of modern AI hides. Not in its brain, but in its memory.<\/p>\n
Until now, the known solution has been quantisation, which simplifies\nthose numbers. A high-precision number like 16.738291 is rounded to 17.\nSimilar to compressing a high-resolution photo. Some details are lost,\nbut the face is still recognisable.<\/p>\n
The problem is that this old technique has a built-in flaw. To\nperform the compression, it requires \u201cadditional costs\u201d in the form of\ncalibration parameters that actually consume memory. Like dieting while\nsnacking. Weight drops a little, but rises quietly.<\/p>\n
This is where Google\u2019s researchers offer an almost philosophical\napproach. They do not just shrink the data, but eliminate the hidden\ncosts of the shrinking process itself.<\/p>\n
How it works sounds like a magic trick, but it is actually advanced\nmathematics that happens to be very elegant.<\/p>\n
The first stage is called PolarQuant. Vectors that were previously\nstored in ordinary coordinates are now randomly rotated to simplify\ntheir structure, then transformed into polar form, consisting of a\ncombination of \u201cmeaning strength\u201d (radius) and \u201cmeaning direction\u201d\n(angle).<\/p>\n
Imagine you no longer give an address as \u201c3 blocks east and 4 north\u201d,\nbut simply \u201c5 blocks at a certain angle\u201d. The destination is the same,\nbut the way it is stored is far more efficient.<\/p>\n
With this approach, the AI system no longer needs expensive\nnormalisation. Data is mapped to a \u201ccircle\u201d with fixed boundaries, not a\n\u201cbox\u201d that changes. This is like moving from a traditional market to a\nlogistics warehouse. The goods are the same, but the arrangement makes\neverything faster and more efficient.<\/p>\n
However, like all compressions, there is always a small residue,\nsometimes in the form of slight errors that remain. This is where the\nGoogle team applies the second stage called QJL, or Quantized\nJohnson-Lindenstrauss.<\/p>\n
This name, which sounds like an ancient incantation, is actually a\nclassic mathematical theorem that allows high-dimensional data to be\ncompacted without damaging the relationships between its points.<\/p>", "url": "https:\/\/jawawa.id\/newsitem\/ai-slim-down-total-1774763448", "image": "" }, "sponsor": "Okusi Associates", "sponsor_url": "https:\/\/okusiassociates.com" }