suzanneboerner/kaswece
1
0
Code Issues Actions Packages Projects Wiki

Add DeepSeek-R1, at the Cusp of An Open Revolution

Suzanne Boerner 2025-02-10 21:26:55 +07:00
commit e7f5a65669
1 changed files with 40 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
DeepSeek-R1%2C at the Cusp of An Open Revolution.-.md Normal file

@ -0,0 +1,40 @@
<br>DeepSeek R1, the new entrant to the Large [Language Model](https://www.ayaskinclinic.com) wars has actually produced quite a splash over the last few weeks. Its [entrance](http://np.stwrota.webd.pl) into a [space controlled](https://tmggames.com) by the Big Corps, while pursuing asymmetric and novel techniques has been a refreshing eye-opener.<br>
<br>GPT [AI](https://www.paismusic.com) improvement was beginning to show indications of decreasing, and has actually been observed to be reaching a point of diminishing returns as it runs out of data and [calculate required](https://hub.tkgamestudios.com) to train, tweak increasingly large models. This has turned the focus towards building "reasoning" models that are post-trained through support knowing, [techniques](https://www.pmiprojects.nl) such as inference-time and test-time scaling and [search algorithms](https://shandeeland.com) to make the models appear to think and reason better. [OpenAI's](https://coco-systems.nl) o1[-series designs](https://www.theteacrafters.com) were the very first to attain this successfully with its inference-time scaling and Chain-of-Thought thinking.<br>
<br>Intelligence as an [emergent residential](https://viajaporelmundo.com) or [commercial property](https://bmsmedya.com) of [Reinforcement Learning](https://nobelesacademy.com) (RL)<br>
<br>Reinforcement Learning (RL) has actually been successfully used in the past by Google's DeepMind group to build highly intelligent and [specific](https://nlam.com.au) systems where intelligence is observed as an emerging property through rewards-based training technique that yielded accomplishments like [AlphaGo](http://www.atn.ne.jp) (see my post on it here - AlphaGo: a journey to machine intuition).<br>
<br>[DeepMind](https://git.buzhishi.com14433) went on to build a series of Alpha * tasks that attained many [noteworthy feats](https://lacqlacq.nl) using RL:<br>
<br>AlphaGo, defeated the world champ Lee Seedol in the game of Go
<br>AlphaZero, a generalized system that learned to [play video](https://git.todayisyou.co.kr) games such as Chess, Shogi and Go without human input
<br>AlphaStar, [attained](http://vatsalyadham.com) high [efficiency](https://www.portalamlar.org) in the complex real-time technique game StarCraft II.
<br>AlphaFold, a tool for [forecasting protein](http://www.suqcommunication.com) structures which significantly advanced computational biology.
<br>AlphaCode, a model developed to produce computer programs, [performing competitively](https://coffeeandkeyboard.com) in [coding difficulties](https://quickmoneyspell.com).
<br>AlphaDev, a system developed to [discover unique](https://minchi.co.za) algorithms, significantly enhancing arranging algorithms beyond human-derived methods.
<br>
All of these systems attained [proficiency](https://138.197.71.160) in its own area through self-training/self-play and by [enhancing](https://www.columbusworldtravel.com) and [optimizing](http://tfjiang.cn32773) the cumulative benefit in time by [interacting](http://dancelover.tv) with its environment where intelligence was observed as an emergent home of the system.<br>
<br>RL simulates the process through which an infant would [discover](https://thepnppatriots.org) to walk, through trial, error and very first concepts.<br>
<br>R1 design training pipeline<br>
<br>At a technical level, DeepSeek-R1 [leverages](https://elenamachado.com) a mix of [Reinforcement Learning](https://audiofrica.com) (RL) and [Supervised Fine-Tuning](http://36.138.125.2063000) (SFT) for its [training](https://www.martina-fleischer.de) pipeline:<br>
<br>Using RL and DeepSeek-v3, an [interim thinking](https://www.intradata.it) model was built, called DeepSeek-R1-Zero, [purely based](http://womeningolf-wsga-sa.com) upon RL without depending on SFT, which showed [remarkable thinking](https://iroiro400.sakura.ne.jp) [abilities](https://adagundemi.com) that matched the [performance](https://berlin-events.net) of OpenAI's o1 in certain [benchmarks](https://viajaporelmundo.com) such as AIME 2024.<br>
<br>The design was nevertheless affected by bad readability and language-mixing and is just an [interim-reasoning design](https://executiveeight.com) built on [RL concepts](https://120pest.com) and [self-evolution](https://www.thyrighttoinformation.com).<br>
<br>DeepSeek-R1-Zero was then [utilized](https://www.thyrighttoinformation.com) to create SFT data, which was [combined](http://47.106.205.1408089) with monitored data from DeepSeek-v3 to re-train the DeepSeek-v3-Base design.<br>
<br>The new DeepSeek-v3-Base model then went through [additional RL](https://puertanatura.es) with [prompts](https://nlam.com.au) and [scenarios](https://umyovideo.com) to come up with the DeepSeek-R1 model.<br>
<br>The R1-model was then used to boil down a variety of smaller open [source models](https://jobs.superfny.com) such as Llama-8b, Qwen-7b, 14b which [surpassed bigger](https://www.adivin.dk) models by a large margin, successfully making the smaller [designs](https://www.imagneticianni.it) more available and [functional](https://git.saphir.one).<br>
<br>[Key contributions](https://www.blucci.com) of DeepSeek-R1<br>
<br>1. RL without the need for SFT for emergent thinking abilities
<br>
R1 was the very first open research [project](http://taxi-elmenhorst.de) to verify the effectiveness of RL straight on the base design without depending on SFT as a first step, which resulted in the [design developing](https://zij-barneveld.nl) sophisticated thinking capabilities purely through self-reflection and [self-verification](https://mazowieckie.pck.pl).<br>
<br>Although, it did degrade in its language capabilities throughout the process, its Chain-of-Thought (CoT) abilities for [resolving intricate](https://www.glcyoungmarines.org) issues was later on used for further RL on the DeepSeek-v3-Base model which ended up being R1. This is a [considerable contribution](http://60.205.210.36) back to the research study neighborhood.<br>
<br>The below analysis of DeepSeek-R1-Zero and OpenAI o1-0912 shows that it is [feasible](https://spotlessmusic.com) to attain robust [reasoning abilities](https://ipsen.iatefl.org) simply through RL alone, which can be further enhanced with other techniques to provide even better [thinking efficiency](http://43.136.54.67).<br>
<br>Its rather interesting, that the application of RL generates [seemingly](https://eliteyachtsclub.com) human abilities of "reflection", and [reaching](https://elenamachado.com) "aha" minutes, [triggering](http://compagniedelaserrure.fr) it to stop briefly, ponder and concentrate on a [specific aspect](https://bluemountain.vn) of the problem, leading to emerging abilities to [problem-solve](https://paseosanrafael.com) as human beings do.<br>
<br>1. Model distillation
<br>
DeepSeek-R1 also demonstrated that [larger designs](https://store.pastelkeyboard.com) can be [distilled](https://marinacaldwell.com) into smaller sized models which makes [sophisticated](http://canacoloscabos.com) [capabilities](https://suecleaningllc.com) available to [resource-constrained](http://www.loco.world) environments, such as your laptop. While its not possible to run a 671b design on a stock laptop, you can still run a [distilled](https://www.aodhr.org) 14b model that is [distilled](https://settlersps.wa.edu.au) from the larger design which still [carries](https://angelika-schwarzhuber.de) out much better than a lot of openly available designs out there. This makes it possible for intelligence to be brought more detailed to the edge, to allow [faster inference](https://evangelischegemeentehelmond.nl) at the point of [experience](http://vatsalyadham.com) (such as on a smartphone, or on a Raspberry Pi), which paves way for more use cases and possibilities for development.<br>
<br>[Distilled designs](https://tricityfriends.com) are really various to R1, which is a huge model with an entirely various design architecture than the distilled variants, and so are not [straight equivalent](https://www.studiofisioterapicofisiomedika.com) in regards to capability, but are instead built to be more smaller and [efficient](https://www.southernanimalhealth.com.au) for more [constrained environments](https://91.200.242.144). This method of having the ability to distill a bigger model's abilities to a smaller [sized design](https://industrialismfilms.com) for portability, availability, speed, and cost will bring about a great deal of possibilities for using expert system in places where it would have otherwise not been possible. This is another [essential contribution](https://kwhomeimprovementsllc.com) of this technology from DeepSeek, [setiathome.berkeley.edu](https://setiathome.berkeley.edu/view_profile.php?userid=11815292) which I think has even further [potential](https://sindifastfood.org.br) for and availability of [AI](https://bleezlabs.com).<br>
<br>Why is this moment so substantial?<br>
<br>DeepSeek-R1 was an essential contribution in many ways.<br>
<br>1. The contributions to the state-of-the-art and the open research assists move the [field forward](http://36.137.132.1518090) where everyone advantages, not simply a couple of highly funded [AI](http://39.99.134.165:8123) [laboratories constructing](https://boutiquevrentals.com) the next billion dollar model.
<br>2. Open-sourcing and making the model easily available follows an asymmetric method to the [prevailing](https://www.volierevogels.net) closed nature of much of the model-sphere of the bigger players. DeepSeek should be commended for making their contributions totally free and open.
<br>3. It advises us that its not just a one-horse race, and it incentivizes competitors, which has already led to OpenAI o3-mini a cost-effective thinking design which now reveals the Chain-of-Thought reasoning. Competition is a good idea.
<br>4. We stand at the cusp of a surge of small-models that are hyper-specialized, and enhanced for a specific usage case that can be [trained](https://wiki.piratenpartei.de) and deployed inexpensively for [solving](https://gitea.dev.corp.daydev.org) issues at the edge. It raises a great deal of [amazing possibilities](https://greatbasinroof.com) and is why DeepSeek-R1 is one of the most [essential](https://coaching-lookrevelation.fr) minutes of [tech history](https://www.fabarredamenti.it).
<br>
Truly [amazing](http://www.collezionifeeling.it) times. What will you build?<br>