From pre-training reach AI agent teams: Why Jensen Huang four scaling loops dey change di computing question
Introduction: Wetin dey grow when artificial intelligence get more ability?
Discussion about artificial intelligence [AI] plenty times dey start with how big one model be. How many parameters e get? How much information dem use train am? How big di computing system wey dey behind am be?
But those questions describe only part of di system. Jensen Huang wider argument be say AI fit grow for different stages: when e dey learn im foundation, when e dey improve im behaviour, when e dey work through one particular problem, and when e dey share work among agents [Agents]. If we understand how these stages connect, e go change how we think about computing demand.
Di important difference na between to build model and to use intelligent system. Di two activities connect, but dem no be di same workload [Workload].
1. Pre-training dey build foundation, no be worker wey don ready finish
Pre-training [Pre-training] dey develop di general abilities of model by exposing am to plenty information. Di model dey learn patterns and relationships wey fit help later tasks.
More data and more computing fit improve those foundations. But bigger training run no be complete explanation of useful intelligence. Model fit get plenty general knowledge and still need better judgment, experience for di particular task, or access to information wey e no see during training.
Na why di question no fit stop for whether human-written material go eventually become limitation. Di next stages fit change both how di system dey learn and di work wey e dey do.
2. Post-training dey change how dem use di foundation
Post-training [Post-training] dey improve model after im first learning. Di focus dey move towards behaviour, practical tasks and di kind responses wey people need.
Synthetic data [Synthetic Data] fit help this process. System fit generate examples, explore different versions and produce material wey go useful for more learning. Di value dey inside di quality of experience wey come out, no be just because machine produce am.
This difference matter. To generate more material no automatically prove say di material correct. Di useful loop dey select experiences wey worth keeping and use dem improve later behaviour. Plenty material without selection fit repeat weaknesses instead of correcting dem.
3. Test-time computing dey make inference a harder problem
Some people dey imagine inference [Inference] as one simple last step: di model don already learn, so to produce answer suppose need only small extra work.
Test-time scaling [Test-time Scaling] dey challenge that picture. Difficult task fit need di system to break problem into smaller parts, compare possible approaches, search for information and reconsider attempt wey no work.
Di computing requirement then depend on wetin di system must do before e return useful answer. Short response fit still follow plenty work. On di other hand, long response no be proof of careful reasoning.
For business, this one make response delay [Latency], answer quality and running cost become choices wey connect. Di cheapest immediate response no always be di cheapest task wey dem successfully complete.
4. Agentic scaling dey add coordination join ability
Agentic scaling [Agentic Scaling] introduce another route: system fit share work among several agents [Agents] instead of asking one model to do everything one after another.
One agent fit inspect records while another dey compare options. Wetin dem find fit support one shared result. This growth different from making one model bigger or giving am more time.
To compare am with organisation useful. More workers fit increase capacity, but only when responsibilities, communication and di final decision clear. More agents likewise dey create work for coordination, settling differences between findings and checking. Their number na input, no be measure of success.
5. Di four stages dey form learning loop
Di most important connection na wetin happen after system work. Useful experiences from agents fit become material for later training. Improved models fit then support better reasoning and more effective agents.
So pre-training, post-training, test-time work and agentic activity no need remain inside separate boxes. Dem fit feed one another.
This one still change di infrastructure question. To train model, search through files and coordinate several workers dey place different demands on processors, memory, storage and networking. System wey dem design around only one narrow benchmark fit no match di complete workflow well.
Conclusion: Measure di work, no be only di model
Huang four-part framework dey move discussion beyond di idea say AI progress mean only bigger model wey dem train once on bigger dataset.
Di practical lesson na to identify where extra computing dey produce better result. Di task need better foundations, more useful improvement, longer problem-solving, or coordinated sharing of work?
Enterprise suppose judge di answer through completed tasks, quality, time and total cost. That one na more useful starting point than to count only parameters or agents. Di future computing question na about di whole system wey dey learn and work.
Powered By K7