<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aaryan Kakad on Medium]]></title>
        <description><![CDATA[Stories by Aaryan Kakad on Medium]]></description>
        <link>https://medium.com/@kakadaaryan10?source=rss-857bc74d5f71------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*wcbdsSjIryfKo3orAPItkw.jpeg</url>
            <title>Stories by Aaryan Kakad on Medium</title>
            <link>https://medium.com/@kakadaaryan10?source=rss-857bc74d5f71------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 05:55:54 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@kakadaaryan10/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Why you don’t need more intelligent models anymore?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/i-compared-2025s-frontier-ai-with-2026-s-open-models-under-40b-params-the-results-shocked-me-4c8ccd554669?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/2560/1*GQ9__6ATWxMJr0Z2hKHS7A.png" width="2560"></a></p><p class="medium-feed-snippet">So, I asked GPT 6.1 to research a bunch of 2025 frontier models and their benchmarks and compare them with 2026 open-weight models.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/i-compared-2025s-frontier-ai-with-2026-s-open-models-under-40b-params-the-results-shocked-me-4c8ccd554669?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/i-compared-2025s-frontier-ai-with-2026-s-open-models-under-40b-params-the-results-shocked-me-4c8ccd554669?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/4c8ccd554669</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[tech]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Mon, 09 Nov 2026 14:25:37 GMT</pubDate>
            <atom:updated>2026-11-14T05:42:16.317Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[exploring the back propagation paper]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/exploring-the-back-propagation-paper-e465cb61ac90?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*AEo9nDGUNtJGtO_FWNyUIA.png" width="1672"></a></p><p class="medium-feed-snippet">this is a raw and random exploration write-up, where I go through the back propagation paper and write what I feel in my own words.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/exploring-the-back-propagation-paper-e465cb61ac90?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/exploring-the-back-propagation-paper-e465cb61ac90?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/e465cb61ac90</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[backpropagation]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Thu, 05 Nov 2026 11:13:08 GMT</pubDate>
            <atom:updated>2026-11-05T11:14:04.062Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[how to self learn anything]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/how-to-self-learn-anything-81587b6ce7bb?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*8n6-UScsCrXsP2gpC5STLg.png" width="1672"></a></p><p class="medium-feed-snippet">ingest, build, hit a wall, fill the gaps, repeat. forever.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/how-to-self-learn-anything-81587b6ce7bb?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/how-to-self-learn-anything-81587b6ce7bb?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/81587b6ce7bb</guid>
            <category><![CDATA[self-improvement]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[learning]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[education]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 21:18:03 GMT</pubDate>
            <atom:updated>2026-10-27T03:04:34.385Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[tesla used vacuum cleaners and rockets in a car to break physics]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/how-does-the-new-tesla-roadster-accelerate-so-fast-how-does-it-fly-ba30e687da56?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*h81_mSDE8NcgVnRXSiHtgw.png" width="1672"></a></p><p class="medium-feed-snippet">tesla made their new roadster go from 0 to 60 mph in 0.81s. here&#x2019;s how they made it happen.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/how-does-the-new-tesla-roadster-accelerate-so-fast-how-does-it-fly-ba30e687da56?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/how-does-the-new-tesla-roadster-accelerate-so-fast-how-does-it-fly-ba30e687da56?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/ba30e687da56</guid>
            <category><![CDATA[engineering]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[electric-vehicles]]></category>
            <category><![CDATA[automotive]]></category>
            <category><![CDATA[tesla]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Tue, 20 Oct 2026 18:42:30 GMT</pubDate>
            <atom:updated>2026-10-23T13:44:00.300Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[can two AI agents communicate without using a language?]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/can-two-ai-agents-communicate-without-using-a-language-3eb5dd0fee1d?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*o6O4GQue4B6deXkRREG5RA.png" width="1672"></a></p><p class="medium-feed-snippet">when I want to say something to you, I have no choice but to use a language. but why are we forcing agents to do the same?</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/can-two-ai-agents-communicate-without-using-a-language-3eb5dd0fee1d?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/can-two-ai-agents-communicate-without-using-a-language-3eb5dd0fee1d?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/3eb5dd0fee1d</guid>
            <category><![CDATA[agentic-ai]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Fri, 16 Oct 2026 14:13:56 GMT</pubDate>
            <atom:updated>2026-10-16T14:14:50.391Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[recursive self improvement (explained from scratch)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/recursive-self-improvement-ee047dd43007?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*c8jptoWQxflJm_iejHJYAg.png" width="1672"></a></p><p class="medium-feed-snippet">just imagine the language model you use becomes capable of improving itself recursively, wouldn&#x2019;t it be dangerous?</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/recursive-self-improvement-ee047dd43007?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/recursive-self-improvement-ee047dd43007?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee047dd43007</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Tue, 06 Oct 2026 14:17:46 GMT</pubDate>
            <atom:updated>2026-10-16T12:39:39.773Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[how a computer works (explained from scratch)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/how-computers-work-89a0fb75feac?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1672/1*2XhZqEfea-4RQLFV5B0dHw.png" width="1672"></a></p><p class="medium-feed-snippet">this article is all about decoding how a computer works behind the scenes.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/how-computers-work-89a0fb75feac?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/how-computers-work-89a0fb75feac?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/89a0fb75feac</guid>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[how-computers-work]]></category>
            <category><![CDATA[computer-science]]></category>
            <category><![CDATA[computers]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 12:52:03 GMT</pubDate>
            <atom:updated>2026-10-16T12:26:24.555Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[google is about to setup a DATACENTER in the SPACE (here’s the physics behind it)]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/google-is-catching-the-sun-99ad670fdfc7?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/1456/1*4FmC0DHbOfyMGbhjOpWldQ.jpeg" width="1456"></a></p><p class="medium-feed-snippet">google is planning to send tpu chips to space for unlimited power - project suncatcher.</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/google-is-catching-the-sun-99ad670fdfc7?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/google-is-catching-the-sun-99ad670fdfc7?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/99ad670fdfc7</guid>
            <category><![CDATA[google]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[physics]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[space]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Fri, 25 Sep 2026 06:55:52 GMT</pubDate>
            <atom:updated>2026-10-08T12:05:02.023Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[pace the frontier? no.]]></title>
            <description><![CDATA[<div class="medium-feed-item"><p class="medium-feed-image"><a href="https://medium.com/@kakadaaryan10/pace-the-frontier-no-98a736112591?source=rss-857bc74d5f71------2"><img src="https://cdn-images-1.medium.com/max/600/0*VgJH0Xdq7bRPV2mc" width="600"></a></p><p class="medium-feed-snippet">you can read the article for free here if you aren&#x2019;t a medium member&#x2026;</p><p class="medium-feed-link"><a href="https://medium.com/@kakadaaryan10/pace-the-frontier-no-98a736112591?source=rss-857bc74d5f71------2">Continue reading on Medium »</a></p></div>]]></description>
            <link>https://medium.com/@kakadaaryan10/pace-the-frontier-no-98a736112591?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/98a736112591</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[tech]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Tue, 22 Sep 2026 06:07:18 GMT</pubDate>
            <atom:updated>2026-09-22T06:08:12.179Z</atom:updated>
        </item>
        <item>
            <title><![CDATA[deepseek v4.1 flash’s alien architecture explained.]]></title>
            <link>https://medium.com/@kakadaaryan10/890-bytes-is-what-deepseek-v4-1-flash-needs-to-store-one-token-cf4a9643ae60?source=rss-857bc74d5f71------2</link>
            <guid isPermaLink="false">https://medium.com/p/cf4a9643ae60</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[localai]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <dc:creator><![CDATA[Aaryan Kakad]]></dc:creator>
            <pubDate>Thu, 10 Sep 2026 12:17:41 GMT</pubDate>
            <atom:updated>2026-09-12T12:14:59.709Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/679/0*0BOzLiUyCevpz_0g" /></figure><p>890 bytes. that’s what deepseek v4.1 flash needs to remember one token. and their own v1 model needed 389,120 bytes for the same job.</p><p>efficiency and compression final boss.</p><p>lets dive deep into it (explained in simple words):</p><p>the paper claims:</p><p>long horizon agents are input-heavy and not output heavy. the bottleneck is not FLOPs, its MEMORY.</p><p>agents eat up so much info that their context bloats, and this is a major problem. so they are trying to achieve aggressive KV cache compression.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*0Y-nUXlafgL00IQd" /></figure><p>deepseek v4.1 flash activates only 8B params per token during prefill and 16B params per token during decode.</p><p>this is something very unique and interesting.</p><p>normally a model uses the same amount of “work” (activated parameters) whether it’s reading your prompt (prefill) or writing its answer (decode).</p><p>this model is unusual — it deliberately uses less work while reading than while writing. 8B params to read, 16B to write.</p><p>(read the next part to know why)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/0*wcw5A1oq5EWCSTmR" /></figure><p>the model has total 40 layer network which is divided into a causal encoder with 20 layers and a decoder with 20 layers.</p><p>so, instead of running the full 40 layers on every single token, it only runs the bottom 20 layers properly.</p><p>and the top 20 layers, instead of doing their computation, they borrow the answer from the bottom half’s output.</p><p>that shortcut is only possible during reading (prefill), because reading is a one time pass over text you already have in full.</p><p>during writing (decode), that shortcut doesn’t apply — the model is generating one new word at a time, and each new word needs the full, real 40-layer computation to decide what comes next, because there’s no “already computed” answer to borrow from.</p><p>that’s why decode costs more: 16B params, the real full-depth cost.</p><p>and they did this because of their core claim: agentic workloads are input heavy and not output heavy. this shortcut makes it a ultra fast agentic model.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*gR3Ufs7wKZSUg6oX" /></figure><p>IMPROVISED causal encoder decoder.</p><p>deepseek researchers, inspired by YoCo (sun et al., 2024), proposed the causal encoder decoder architecture.</p><p>and they improvised it: reduced nearly half of the prefill computation while maintaining the performance comparable to the baseline.</p><p>how?</p><p>by not letting the top 20 layers do their full computation, and just let them borrow the answers from the bottom 20 layers output.</p><p>why?</p><p>because the bottom half’s final output already contains a rich, “read the whole thing” understanding of the prompt. the top half doesn’t need to independently re-derive that understanding — it can just borrow it and adapt it slightly for its own purpose.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*_z3rGWuwWqfdSPJ-" /></figure><p>CSA2: compressed sparse attention 2.</p><p>they introduced three different modes in CSA2 to make sure they dont just optimize KV cache but also optimize attention computation.</p><p>full mode: the layer does everything itself: own main KV, own indexer K, own indexer Q pass, fresh Top-K selection. this is a complete, standalone CSA2 computation — the baseline every other mode is cheaper than.</p><p>reindex mode: inherits main KV and indexer K from an earlier layer (saves the expensive storage + recomputation), but still runs its own indexer Q and produces a FRESH Top-K selection — so different layers can still attend to different subsets, even from shared underlying KV.</p><p>reuse mode: maximal savings. inherits main KV, indexer K, AND the Top-K selection itself — no indexer Q computation at all. the layer only computes its own query and SWA KV, then directly attends using the inherited selection. this is what most encoder layers actually run (5 of every 6, per the paper’s configuration).</p><p>so, in simple words, 5 of every 6 encoder layers do zero fresh indexing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*muR7HxHfJ46Zxr5C" /></figure><p>hierarchical sparse indexer.</p><p>normally, every layer has to scan the entire prompt to decide what’s relevant and what’s not.</p><p>but this trick says: just let the layer 1 do the entire scan and hand over the short summary to the succeeding layers, so they only search within that shortlist instead of rescanning everything.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/0*_XrNed04Ew5R3RN-" /></figure><p>they introduced single pass mHC.</p><p>in mHC, the model used to read the same data twice due to some awkward dependency.</p><p>so, they fixed/reordered the calculation so it only needs to read it once. no change on what the model learns, just faster than before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/0*MF3l3O_6cpS9GFP3" /><figcaption>fp4 kv cache.</figcaption></figure><p>deepseek v4 already used QAT — quantization aware training.</p><p>in deepseek v4.1 flash, they adopted the OCP-standard MXFP4 format to support as many hardware platforms as possible, despite the higher accuracy of alternative formats in their experiments.</p><p>in simple words:</p><p>instead of storing each memory entry using tons of precision (lots of digits), store it using a compressed 4-bit format. less accurate per number, but the loss is tiny and it halves memory storage.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*vb9VrKk4sWuQbRBF" /></figure><p>bounded replay.</p><p>since SWA dependencies accumulate across layers, exactly reconstructing the SWA KV of 𝐿 layers would require replaying 𝐿 × 𝑛win tokens.</p><p>SWA bounded replay instead replays only the most recent 𝑛win tokens and truncates SWA to the replay segment, accepting approximate states: for a replay starting at position 𝑠, a query at position 𝑖 attends to SWA keys in max(𝑠, 𝑖 − 𝑊 + 1), 𝑖.</p><p>again, in simple words:</p><p>instead of exactly reconstructing certain memory states (which is expensive), approximately rebuild them using only the most recent bit of context. cheaper, and barely any quality loss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*z6h7FnVcSfFCIMuX" /></figure><p>in this release, they haven’t introduced new novel post-training algorithms.</p><p>they just followed the conventional pipeline:</p><p>SFT → RL → OPD.</p><p>but their pre-training inventions are absolutely revolutionary, ngl.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/679/0*njyiCbKfKj6iYvXr" /></figure><p>lets talk about the numbers:</p><p>this model beats opus 5 and a lot of other top models on deepswe v1.1, automation bench, etc.</p><p>its absolutely insane.</p><p>it beats closed frontier models at a lot of benchmarks while being so efficient. and you can download and run it for free (only if you have the hardware, the baseline for the hardware requirements is going down day by day).</p><p>OSS AI is definitely something.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/680/0*gezZ3VbZuY0L-bf4" /></figure><p>every trick in this paper is the same question, asked five different times: do we really need to recompute this from scratch, or can we just cheaply reuse or approximate what we already have?</p><p>they wanted to make the fastest agentic model, and they absolutely succeeded.</p><p>their original claim: agentic workload is input heavy and not output heavy made sense too.</p><p>absolute W for <a href="http://twitter.com/deepseek_ai">@deepseek_ai</a>.</p><p><em>I am Aaryan. 19 — self taught ML guy who’s been actively building in the ML and AI space for 2 years now. I write from the perspective of a builder — breaking down what’s actually happening under the hood of the things everyone is talking about. You can find me at </em><a href="https://x.com/@aaryan_kakad"><em>@aaryan_kakad</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cf4a9643ae60" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>