<?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 Saloni Garg on Medium]]></title>
        <description><![CDATA[Stories by Saloni Garg on Medium]]></description>
        <link>https://medium.com/@saloni_garg?source=rss-5291774e0411------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*F5qUWv3w6Z_0wkadPNL2FQ.jpeg</url>
            <title>Stories by Saloni Garg on Medium</title>
            <link>https://medium.com/@saloni_garg?source=rss-5291774e0411------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 22:12:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@saloni_garg/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[The Hidden Problem Behind LLM Routing Benchmarks]]></title>
            <link>https://medium.com/@saloni_garg/the-hidden-problem-behind-llm-routing-benchmarks-f3c6b40694e2?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/f3c6b40694e2</guid>
            <category><![CDATA[deepmind]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[anthropic-ai]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Mon, 11 May 2026 17:47:44 GMT</pubDate>
            <atom:updated>2026-05-11T17:47:44.591Z</atom:updated>
            <content:encoded><![CDATA[<p>A few months ago, I started working on a fairly straightforward question:</p><p><strong>How well can we route queries between multiple LLMs?</strong></p><p>The idea behind routing sounds simple. Instead of sending every request to the biggest and most expensive model, you try to send easy queries to smaller models and reserve the stronger models for harder tasks.</p><p>In theory, this gives you:</p><ul><li>lower cost</li><li>lower latency</li><li>similar quality</li></ul><p>A lot of recent work in the space assumes there’s meaningful “routing headroom” — that there exists a large class of queries that weaker models fail on but stronger models solve correctly.</p><p>I initially thought I’d spend most of my time experimenting with routers.</p><p>Instead, I spent most of my time debugging evaluation.</p><h3>The First Weird Thing</h3><p>I was running evaluations across multiple Gemma 4 and Llama 3.1 tiers on:</p><ul><li>MMLU</li><li>MedQA</li><li>HumanEval</li><li>MBPP</li><li>Alpaca</li><li>ShareGPT</li></ul><p>Eventually the experiment grew into more than 206,000 query-model evaluations.</p><p>At first, the results looked reasonable.</p><p>But then I noticed something strange:</p><p>Changing the evaluator sometimes changed the routing conclusion itself.</p><p>Not slightly.</p><p>Substantially.</p><p>A query that looked “unsolved” under one evaluation setup suddenly became solvable under another.</p><p>That was the moment the project changed direction.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/484/1*vbF-geiR3jeNihIFhyeDIg.jpeg" /></figure><h3>The Rabbit Hole</h3><p>I started looking at three things more carefully:</p><h3>1. LLM-as-a-judge behavior</h3><p>The judge wasn’t simply checking correctness.</p><p>It was also rewarding:</p><ul><li>structure</li><li>verbosity</li><li>reasoning presentation</li><li>confidence</li><li>fluency</li></ul><p>And sometimes penalizing:</p><ul><li>terse answers</li><li>unconventional formatting</li><li>incomplete reasoning traces</li></ul><p>This became especially obvious on multiple-choice benchmarks like MMLU and MedQA.</p><p>Two evaluators could look at the same output and disagree systematically.</p><p>One would say:</p><blockquote><em>“This is correct.”</em></blockquote><p>The other would say:</p><blockquote><em>“The reasoning quality isn’t sufficient.”</em></blockquote><p>That distinction sounds subtle, but it changes downstream routing labels dramatically.</p><h3>2. Truncation effects</h3><p>This one surprised me more than expected.</p><p>Under fixed generation budgets, some benchmarks experienced extremely high truncation rates.</p><p>The model would often:</p><ul><li>identify the right answer internally</li><li>begin generating reasoning</li><li>get cut off before finishing</li><li>fail parsing</li><li>receive a lower evaluation score</li></ul><p>The evaluation pipeline interpreted that as:</p><blockquote><em>“model failure”</em></blockquote><p>But operationally, it was often:</p><blockquote><em>“generation budget failure.”</em></blockquote><p>That distinction matters a lot in production systems.</p><h3>3. Router collapse</h3><p>This became the most interesting part of the project.</p><p>I trained several routers:</p><ul><li>logistic regression</li><li>MLP</li><li>DistilBERT</li></ul><p>Different architectures.<br>Different capacities.</p><p>And several of them basically learned the same thing:</p><p>Always choose the cheapest model.</p><p>The accuracy still looked “good” because the label distribution itself was heavily skewed toward the smallest tier.</p><p>At first I thought:</p><blockquote><em>maybe the features are weak</em></blockquote><p>But even random Gaussian features reproduced the same collapse.</p><p>Shuffled labels reproduced it too.</p><p>That was the point where the failure mode became clearer:</p><p>The evaluation pipeline itself was shaping the supervision signal.</p><h3>The Most Interesting Insight</h3><p>The biggest thing I took away from this project is that evaluation artifacts don’t just affect benchmark reporting.</p><p>They propagate downstream.</p><p>They affect:</p><ul><li>router supervision</li><li>calibration</li><li>deployment decisions</li><li>model selection</li><li>perceived routing value</li></ul><p>A routing system can look intelligent while mostly learning dataset priors and evaluation quirks.</p><p>And honestly, I think this applies more broadly to modern LLM systems too.</p><p>As the ecosystem moves toward:</p><ul><li>agents</li><li>orchestration</li><li>model routing</li><li>evaluator models</li><li>synthetic supervision</li></ul><p>…the evaluation layer itself becomes part of the model behavior.</p><p>Not just a measurement tool.</p><h3>What Changed My Mind</h3><p>Before this project, I thought of evaluation mostly as:</p><blockquote><em>“the thing you run after the system is built.”</em></blockquote><p>Now I think evaluation is part of the system itself.</p><p>The choice of:</p><ul><li>judge</li><li>token budget</li><li>parser</li><li>stopping criteria</li><li>formatting assumptions</li><li>aggregation metric</li></ul><p>can materially change the scientific conclusion you walk away with.</p><p>That realization ended up being far more interesting to me than the original routing experiments.</p><p>One thing I’ve started wondering more about lately:</p><p>How many reported improvements in LLM systems are actually improvements in:</p><ul><li>evaluation alignment,<br>rather than:</li><li>capability itself?</li></ul><p>I don’t think we collectively understand that boundary very well yet.</p><p>And I suspect the next few years of ML research are going to force us to think much more carefully about it.</p><p>If you’re building something similar or asking yourself similar questions, let’s connect!</p><p>Continuing the tradtion, here’s a picture of the food I had after finishing this blog.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*JWEgmQei4-qs-76cNprkhg.jpeg" /><figcaption>Very very tasty salad (like genuinly tasty!)</figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f3c6b40694e2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Rise of Yes-Press Engineers]]></title>
            <link>https://medium.com/@saloni_garg/the-rise-of-yes-press-engineers-315e98d09058?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/315e98d09058</guid>
            <category><![CDATA[yes-pressing]]></category>
            <category><![CDATA[software-development]]></category>
            <category><![CDATA[vibe-coding]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[claude]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Thu, 16 Apr 2026 06:20:53 GMT</pubDate>
            <atom:updated>2026-04-16T06:20:53.539Z</atom:updated>
            <content:encoded><![CDATA[<blockquote><strong><em>yes-press</em></strong><em> </em>(v.)<em> — to review, accept, and ship AI-generated code. <br>(aka pressing “Yes” on Claude and hoping for the best)</em></blockquote><blockquote><strong><em>yes-press engineer</em></strong><em> </em>(n.)<em> — someone who does this well. <br>Once a shortcut. Now a profession.</em></blockquote><p>Have you ever sat in a Tesla on the early Full Self Driving beta? You’re <strong>not</strong> driving. But you <strong><em>are</em></strong> exhausted. Watching the car decide, will it see that cyclist, will it take this turn, is it about to do something stupid, takes more headspace than just driving yourself. You’re not passive. You’re <em>hyper</em>-<em>present</em>. That’s yes-pressing.</p><p>The AI (<em>okay, let’s just say Claude at this point!</em>) is writing the code. You are watching everything.<br>And just like FSD, the moment you stop watching (<em>really watching</em>) is the moment it sends your half baked thing to remote.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*b8vLuAy32pYY5o9Y" /></figure><h3>Watching Is the Work Now</h3><p>Five years ago, being a software engineer meant something specific. You wrote code. You debugged it. You googled the error, landed on a Stack Overflow answer from 2011, adapted it (🤨 <em>or just copied it!</em>), and moved on.</p><p>Today, a new kind of engineer has emerged. They open their editor, describe what they want in plain English, and when the code appears, they press yes.</p><p>They are Yes-Press Engineers. And they might be the most important people in software right now.</p><h3>What nobody tells you about pressing yes</h3><p>The first time you ship something you didn’t write, it feels like cheating.</p><p>The tenth time, it feels like leverage.</p><p>The hundredth time, you realize the yes-press was never the easy part. The easy part was the code. The hard part (<em>the part that was always the hard part</em>) was knowing what to build, knowing when it’s right, and knowing when to say no.</p><p>The AI didn’t change that. It just made it <em>impossible</em> to hide behind the syntax anymore.</p><h3><strong>The yes-press is not passive</strong></h3><p>A bad yes-presser approves <em>everything</em>. Doesn’t read, doesn’t think, ships the AI’s first draft without blinking. This person exists. The guy who’s on-call in their team knows it :P<br>A good one reads the output the way a senior engineer reads a junior’s PR.</p><p><strong>What YOU have is <em>context</em></strong>. The AI doesn’t know your codebase. Doesn’t know why that service is structured strangely, what the roadmap looks like, which shortcuts will haunt you at 2am in six months. The person who holds that context and uses it to steer the AI, is irreplaceable. The person who doesn’t is not.</p><p><strong>Understanding becomes a <em>choice</em>. </strong>Before, you understood what you wrote because you wrote it. Now you can ship code you didn’t write, which means understanding it is an <em>active choice</em>, not a byproduct. The ones who choose to will be trusted.</p><p>You are still responsible for what ships. That hasn’t changed. If anything, the responsibility has intensified because you’re shipping faster, and the consequences of poor judgment arrive sooner.</p><p>The yes-press is yours. Own it.</p><h3><strong>A new title for a new era</strong></h3><p>Job boards haven’t caught up yet. LinkedIn doesn’t have a Yes-Press Engineer category.</p><p>Give it two years.</p><p>What we are watching, right now, is the redefinition of what it means to build software. The people navigating that redefinition best — learning the new skills, holding onto the judgment that made them good, staying hyper-present even when they’re not the ones driving — are the Yes-Press Engineers.</p><p>They don’t write every line. But they are responsible for every line.</p><p>That’s always been the job, actually.</p><p>The tools just finally got fast enough to make it obvious.</p><p><em>If you found yourself nodding, you’re probably already one of us. Welcome to the new job title. Add it to your resume whenever you’re ready :P</em></p><p>As usual, here’s a picture of the meal we had after I finished writing this blog.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tYwKWZrtPRcd9jlds6ZQgw.jpeg" /><figcaption><em>Remember “Breakfast at Tiffany’s and bottles of bubbles”?<br> It’s been on the list forever. And we finally got a resy!!!</em></figcaption></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=315e98d09058" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Reasoning Mode Evaluation in On-Device Transformers (Gemma 4)]]></title>
            <link>https://medium.com/@saloni_garg/reasoning-mode-evaluation-in-on-device-transformers-gemma-4-9ccc9762a6df?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ccc9762a6df</guid>
            <category><![CDATA[ai-on-device]]></category>
            <category><![CDATA[llm-evaluation]]></category>
            <category><![CDATA[openai]]></category>
            <category><![CDATA[gemma-4]]></category>
            <category><![CDATA[deepmind]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Fri, 10 Apr 2026 00:50:30 GMT</pubDate>
            <atom:updated>2026-04-10T00:50:30.366Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A smoke test across GSM8K, HumanEval, and MMLU, and the three parser bugs we had to fix along the way.</em></p><h3>Background</h3><p>Gemma 4 ships with a “thinking mode” : a toggle that prompts the model to reason inside &lt;think&gt;...&lt;/think&gt; blocks before producing its final answer. The promise is higher accuracy at the cost of more tokens and latency. But how much more? And is the tradeoff worth it?</p><p>We set out to answer this with a fully automated, no-annotation benchmark on a MacBook M2 (8 GB unified memory) using:</p><ul><li><strong>Model</strong>: google/gemma-4-E2B-it -- loaded via MLX at 4-bit quantization (~2.4 GB)</li><li><strong>Benchmarks</strong>: GSM8K (math), HumanEval (code), MMLU (reasoning)</li><li><strong>Conditions</strong>: thinking <strong>ON</strong> vs. thinking <strong>OFF</strong>, greedy decoding, fixed random seed</li></ul><p>This post covers the smoke test phase: 20 samples per benchmark per mode, and is honest about the three evaluation bugs we hit before we had trustworthy numbers.</p><h3>Setup</h3><h3>Why MLX instead of PyTorch/MPS?</h3><p>The original plan was to load the model with transformers + MPS. It immediately crashed:</p><pre>RuntimeError: Invalid buffer size: 9.51 GB</pre><p>Gemma 4 E2B is multimodal : its vision encoder pushes the full model well past 9 GB in bfloat16, which exceeds the M2&#39;s 8 GB unified memory. MPS has no memory paging; it&#39;s all-or-nothing. The fix: switch to MLX-LM with a 4-bit quantized model (unsloth/gemma-4-E2B-it-UD-MLX-4bit), which fits in ~2.4 GB and uses Apple&#39;s Metal framework natively.</p><h3>Thinking mode toggle</h3><p>Gemma 4’s tokenizer exposes an enable_thinking flag in apply_chat_template. Our wrapper tries this first and falls back to injecting a system prompt that asks the model to reason inside &lt;think&gt; tags:</p><pre>def _build_prompt(self, user_message: str, thinking: bool) -&gt; str:<br>    messages = [<br>        {&quot;role&quot;: &quot;system&quot;, &quot;content&quot;: THINKING_SYSTEM if thinking else DIRECT_SYSTEM},<br>        {&quot;role&quot;: &quot;user&quot;, &quot;content&quot;: user_message},<br>    ]<br>    try:<br>        return self.tokenizer.apply_chat_template(<br>            messages, tokenize=False,<br>            add_generation_prompt=True,<br>            enable_thinking=thinking,<br>        )<br>    except TypeError:<br>        return self.tokenizer.apply_chat_template(<br>            messages, tokenize=False,<br>            add_generation_prompt=True,<br>        )</pre><p>The _strip_thinking method removes &lt;think&gt;...&lt;/think&gt; blocks to produce a clean final answer for scoring:</p><pre>@staticmethod<br>def _strip_thinking(text: str) -&gt; str:<br>    cleaned = re.sub(r&quot;&lt;think&gt;.*?&lt;/think&gt;&quot;, &quot;&quot;, text, flags=re.DOTALL)<br>    return cleaned.strip() or text.strip()</pre><h3>Run 1: Everything Looks Broken</h3><p>The first run produced:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RUiqVJ2FKjm36B4DWjNA5w.png" /><figcaption>Run 1 Smoke test Results</figcaption></figure><p>Thinking mode appeared to <em>devastate</em> accuracy across all three benchmarks. The verdict was -33.3pp accuracy, 4.7x latency. Before writing that up, we looked at the raw outputs. Three bugs, not a model failure.</p><h3>The Fixing Journey</h3><h3>Bug 1: MMLU token budget too tight</h3><p>MMLU was configured with max_new_tokens=64, which is enough for a direct answer (a single letter). But with thinking ON, the model uses those 64 tokens on its reasoning chain and never reaches the answer. The telltale sign: avg_tokens=64 on almost every sample -- it was hitting the cap every time.</p><p>Fix: benchmark-specific, mode-aware token limits:</p><pre>MAX_NEW_TOKENS = {<br>    # (thinking_off, thinking_on)<br>    &quot;gsm8k&quot;:     (512,  1024),<br>    &quot;humaneval&quot;: (768,  1024),<br>    &quot;mmlu&quot;:      (64,   512),   # 64 is enough OFF; ON needs room to think first<br>}</pre><h3>Bug 2: GSM8K #### marker inside &lt;think&gt;</h3><p>GSM8K answers are graded by extracting the number after a #### marker. With thinking ON, the model often writes #### 42 <em>inside</em> the thinking block. Our pipeline stripped thinking first, then searched for #### -- which was already gone.</p><p>The fix: for GSM8K specifically, pass full_text (pre-stripping) to the checker so the #### marker is still present:</p><pre># GSM8K: #### marker often appears inside &lt;think&gt;, so search full_text.<br># HumanEval and MMLU: use stripped answer.<br>if thinking and bench_name == &quot;gsm8k&quot;:<br>    check_text = result[&quot;full_text&quot;]<br>else:<br>    check_text = result[&quot;answer&quot;]</pre><p>After this fix, GSM8K ON jumped from 10% → 55%.</p><h3>Bug 3: &lt;think&gt; blocks break Python exec() in HumanEval</h3><p>HumanEval scores by executing generated code against test cases. With thinking ON and full_text passed to the checker, the code string looked like:</p><pre>&lt;think&gt;<br>I need to write a function that... A would work, B would also work...<br>&lt;/think&gt;</pre><pre>def has_close_elements(numbers, threshold):<br>    ...</pre><p>&lt;think&gt; is not valid Python syntax. Every exec() call threw a SyntaxError. Score: 0/20.</p><p>Fix: strip thinking blocks inside _clean_code before compilation:</p><pre>def _clean_code(raw: str, prompt: str) -&gt; str:<br>    # Remove &lt;think&gt;...&lt;/think&gt; blocks before any code extraction<br>    raw = re.sub(r&quot;&lt;think&gt;.*?&lt;/think&gt;&quot;, &quot;&quot;, raw, flags=re.DOTALL)<br>    raw = re.sub(r&quot;```python\s*&quot;, &quot;&quot;, raw)<br>    raw = re.sub(r&quot;```\s*&quot;, &quot;&quot;, raw)<br>    raw = raw.strip()<br>    if raw.startswith(&quot;def &quot;):<br>        return prompt + &quot;\n&quot; + raw<br>    return prompt + raw</pre><h3>Residual Bug: MMLU letter extraction confused by reasoning</h3><p>Even after fixing the token budget, MMLU ON scored only 20% in run 2. The _extract_mmlu_letter regex finds the <em>first</em> [A-D] in the text. With full_text, that first letter is usually inside the reasoning chain -- &quot;Option A is incorrect because..., B is also wrong...&quot; -- not the model&#39;s actual answer.</p><p>Fix: use the stripped answer for MMLU (and HumanEval). Only GSM8K needs full_text.</p><h3>Run 2 Results (Partial Fixes Applied)</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*BDmWxznWW6VUtsKC9Av_QA.png" /></figure><p>GSM8K recovered. HumanEval ON still 0% (exec bug not yet applied). MMLU ON still confused by in-reasoning letters.</p><h3>What the Numbers Tell Us (So Far)</h3><blockquote><em>Disclaimer: 20 samples is a smoke test, not a paper result. Variance is high. These are directional signals only.</em></blockquote><h3>Latency overhead is real and large</h3><p>Across all three benchmarks, thinking mode generates 4–14x more tokens. On M2 via MLX 4-bit:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RyU3IFJsQBUcBRM8CJAG1g.png" /></figure><p>HumanEval’s extreme latency overhead reflects the 1,032 avg tokens, the model writes an extended reasoning trace before (or instead of) generating code.</p><h3>Accuracy signal is mixed at 2B scale</h3><p>GSM8K at 20 samples: thinking OFF 70% vs. thinking ON 55%. This is counterintuitive, more reasoning appears to <em>hurt</em> on math at this scale, even after fixing the parser. One hypothesis: at 2B parameters, extended chain-of-thought introduces compounding errors rather than corrections. The model may “talk itself out of” correct answers.</p><p>MMLU and HumanEval results are not yet trustworthy (pending run 3 with all fixes applied).</p><h3>Key lesson: evaluating thinking-mode models requires thinking-aware parsers</h3><p>The most important finding of the smoke test isn’t about the model at all, it’s about evaluation methodology. Standard benchmark parsers assume compact, direct answers. Thinking-mode outputs break three categories of assumptions:</p><ol><li>Answer location: markers like #### may appear inside &lt;think&gt;, not after it</li><li>Token budget: fixed short limits truncate mid-reasoning, producing no answer</li><li>Code/syntax validity: thinking blocks are natural language, not valid syntax for any language being evaluated</li></ol><p>Any evaluation of extended-reasoning models needs explicit handling for all three.</p><h3>What’s Next</h3><p>Run 3 with all fixes applied is ready. Once we have clean numbers at 20 samples, we’ll scale to 200 samples for statistical reliability and add:</p><ul><li>Per-subject MMLU breakdown: thinking mode may help on some subjects (logic, math) and hurt on others (factual recall)</li><li>Token efficiency curve: accuracy vs. max_new_tokens budget for thinking mode</li><li>Gemma 4 E4B comparison: does the 4B model close the gap, or does scale not help thinking at 4B either?</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ccc9762a6df" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Serving LLaMA2 with FlashAttention and vLLM on a Single GPU]]></title>
            <link>https://medium.com/@saloni_garg/serving-llama2-with-flashattention-and-vllm-on-a-single-gpu-d9bb66a10fca?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/d9bb66a10fca</guid>
            <category><![CDATA[vllm]]></category>
            <category><![CDATA[llama-2]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Sun, 11 May 2025 01:45:05 GMT</pubDate>
            <atom:updated>2025-05-11T01:45:05.923Z</atom:updated>
            <content:encoded><![CDATA[<h3>Why I Wrote This</h3><p>If you’re like me, you’ve probably played with large models like LLaMA 2 and quickly realized that serving them is a mehhh task. They’re slow, they eat up your memory, and you need to be clever to get decent throughput, especially if you only have a single GPU.</p><p>In this post, I’ll walk you through how I got LLaMA 2 serving quickly using FlashAttention and vLLM.</p><h3>My Setup</h3><p><strong>Hardware</strong></p><ul><li>GPU: NVIDIA A100 (80GB)</li><li>CPU: AMD EPYC 7543</li><li>RAM: 256 GB</li><li>OS: Ubuntu 22.04</li><li>CUDA: 12.1</li><li>Python: 3.10</li></ul><p><strong>Software</strong></p><ul><li>Model: meta-llama/Llama-2-7b-hf</li><li>Libraries: PyTorch 2.1.0, FlashAttention v2, vLLM</li></ul><h3>Getting FlashAttention and vLLM Working</h3><h3>FlashAttention v2</h3><pre>pip install flash-attn --no-build-isolation</pre><p>Make sure your pip and setuptools are up-to-date. I had some version conflicts that were annoying, <strong>so definitely check that early.</strong></p><h3>vLLM</h3><pre>git clone https://github.com/vllm-project/vllm.git<br>cd vllm<br>pip install -e .</pre><p>Then I ran:</p><pre>python -m vllm.entrypoints.openai.api_server \<br>  --model meta-llama/Llama-2-7b-hf \<br>  --tokenizer meta-llama/Llama-2-7b-hf \<br>  --max-num-batched-tokens 4096</pre><p>Worked like a charm.</p><h3>What Makes vLLM Fast</h3><p>vLLM is not just a wrapper. It’s a <strong>fundamentally</strong> better way to serve LLMs. It introduces something called <strong>PagedAttention</strong> that helps split key/value storage from the query. That allows:</p><ul><li>Sharing cache across requests</li><li>Generating tokens in parallel</li><li>Scaling batch sizes without latency spikes</li></ul><p>With HuggingFace, each input is processed in order. vLLM interleaves tokens across requests, which is just smarter.</p><h3>Latency Benchmarks</h3><h3>How I Measured</h3><ul><li>10 prompts, 128 tokens each</li><li>Output: 64 tokens</li><li>Tested batch sizes from 1 to 16</li></ul><p>FrameworkAvg Latency (ms)Tokens/secVRAM UsageHuggingFace920~150~28 GBvLLM + FlashAttention220~670~20 GB</p><h3>What Surprised Me</h3><ul><li><strong>4x faster</strong> with vLLM + FlashAttention (Like really 4x!!!!!!!!)</li><li>Used <strong>less memory</strong>, not more</li><li>Scaling batch size actually <em>helped</em> throughput</li></ul><h3>My takeaways →</h3><h3>1. FlashAttention works best with aligned sequence lengths (V V Important)</h3><p>It shines when the sequence lengths are aligned (like 128, 256, 512). Otherwise, you don’t get the big wins.</p><h3>2. Bigger batches = better performance</h3><p>vLLM handles batching way more gracefully than HuggingFace. With batch size 8 or more, the gains were huge.</p><h3>3. There’s a memory tradeoff</h3><p>You’ll save memory per request, but caching many requests makes a difference. Use --max-num-batched-tokens wisely.</p><h3>4. Logs matter</h3><p>vLLM gives you visibility into what’s going on → how many tokens generated, cache hits, latency. Super useful.</p><p>So,</p><p>If you’re working on anything that needs to serve LLMs fast , like chatbots, agents, tools then don’t just throw hardware at the problem. FlashAttention and vLLM give real, measurable improvements even on a single GPU. I’m excited to test out speculative decoding and quantization next.</p><p>Hope this helps, please reach out to me and lmk how it went for you if you plan to try out this exact setup!</p><p>Oh, and after this I took us out for a beautifuuuul Ramen dinner ❤</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*89JBioUsxHGwSUoaPGFLug.jpeg" /></figure><p>It’s Yugen Ramen at Castro St, Mountain View.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=d9bb66a10fca" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Attending Startup World Cup- Silicon Valley Regionals]]></title>
            <link>https://medium.com/@saloni_garg/attending-startup-world-cup-silicon-valley-regionals-52cd4f6ce202?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/52cd4f6ce202</guid>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[silicon-valley]]></category>
            <category><![CDATA[seed-investment]]></category>
            <category><![CDATA[startup-world-cup]]></category>
            <category><![CDATA[startup]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Fri, 04 Aug 2023 21:26:48 GMT</pubDate>
            <atom:updated>2023-08-04T21:45:07.432Z</atom:updated>
            <content:encoded><![CDATA[<p>Last week, I found myself immersed in a world of innovation at the Startup World Cup’s Silicon Valley Regionals event. Fourteen startups, each with unique ideas and robust visions, showcased their products and roadmaps. Here’s a detailed account of what unfolded.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2wYz9Vy9EH_hyBEaLpjqUQ.jpeg" /></figure><h3>Insights from Judges</h3><p>First, there was this interesting panel discussion where the judges were given thumbs up, thumbs down placards and asked interesting questions. Turns out, most of them agree that:</p><ul><li>Silicon Valley/Bay Area will remain the center of innovation by 2030.</li><li>Most cars will be electric by 2030.</li><li>The stock market will rise by the end of this year.</li><li>The US isn’t headed into a recession in the foreseeable future.</li></ul><p>When asked about the most overhyped tech, the general consensus pointed to Generative AI, cryptocurrency, and LLMs. Conversely, they identified lithium battery technology, material science, and renewable energy as the most under-hyped fields.</p><h3>The Startups</h3><p>Now, let’s dive into the details of the presenting startups:</p><h3>1. Api for Electric Cars</h3><p>An API endpoint displaying nearby charging points, aiming to integrate with multiple EVs. Judges were a little skeptical, viewing this as a probable basic feature in EVs.</p><h3>2. Deepmind.ai</h3><p>With $1.5 million in revenue (targeting $3 million by year-end), this startup detects deepfake videos, and has even been utilized by the US government. Their demo featured a Joe Biden’s deepfake, using a Zero shot algorithm.</p><h3>3. 3-DOS</h3><p>Their website allows designers to upload and create designs as NFTs, cutting manufacturing and shipping costs by 96% by printing locally. They’ve raised $900k and are about to launch a beta.</p><h3>4. Gamr</h3><p>Sadly, I missed this presentation (pizza was calling!).</p><h3>5. Data Safeguard</h3><p>A data privacy platform, with revenue sources like Enterprise license, cloud, and marketplace API. They made $2.5 million in 2022 and aim to close this year with $5 million. What is the actual product is still a little ambiguous for me.</p><h3>6. IPMD — Emotional Intelligence</h3><p>Using AI to detect emotions, this could be a game-changer in Avatar creation and healthcare. Having raised $1 million, they demonstrated their technology by simulating a hiring process. In the demo, two individuals were asked typical interview questions, such as “Why should we hire you?” and their responses were recorded on video. The AI scored one person as “angry” and the other as “confident.” However, a judge later pointed out that the emotion labeled as “angry” was not readily detectable by human viewers watching the video, highlighting a potential challenge in the technology’s interpretation.</p><h3>7. Abstract</h3><p>Aiming to influence government, it trains on datasets of policy documents, possibly helping companies prepare governance and policy docs.</p><h3>8. Wheel the World</h3><p>An Expedia for people with disabilities, they’ve helped over 4000 people book accessible experiences, growing 2.5x this year. They raised $6 million in pre-seed funding.</p><h3>9. Tactile AI</h3><p>An AI-driven user testing platform targeting the beauty and skincare industry. The problem they aim to solve is that skincare brands are often launching products without sufficient real-skin testing, focusing instead on quick viral success on platforms like TikTok. Tactile AI has assembled a diverse panel of over 10,000 testers, with an automated recruitment process that considers different skin types and ethnicities. This ensures targeted and meaningful feedback on products, addressing a critical gap in the industry’s current approach. The startup has already raised $1 million in funding from Spring Camp.</p><h3>10. Kwikkart</h3><p>An AI-powered shopping cart system that allows grocery stores to offer walk-in checkout, making the shopping experience smoother and more convenient. While competitors like Amazon Go employ extensively priced shopping carts that can cost upwards of $10,000 each, Kwikkart keeps costs low with a transaction fee of $0.02. They have already secured partnerships with DoorDash, Shipt, and Point Pickup. Currently, they are seeking $1 million in reseed funding, aiming for a full-scale launch in 2024.</p><h3>11. Ainnocence</h3><p>A gen AI-powered drug discovery platform focusing on neurological diseases, aiming to reduce costs and time in antibody discovery.</p><h3>12. Mitra Chem</h3><p>AI-powered discovery in the field of battery materials, with a particular focus on iron-based cathodes. Currently, these essential components are entirely manufactured in China, a situation Mitra Chem seeks to diversify. The company is actively looking to raise $60 million in Series B funding to establish a manufacturing plant, allowing for process validation and the initiation of early-scale revenue. Their vision extends beyond mere production; Mitra Chem aims to cater to a broad customer base including electric vehicle batteries, trucking equipment, and power grid batteries, carving a unique niche in the market.</p><h3>13. Pocket.AI</h3><p>A ChatGPT for personal finance, partnered with Visa and Nvidia, this tool aggregates financial data and provides insights and recommendations.</p><h3>Concluding Thoughts</h3><p>The startups at the Silicon Valley Regionals event displayed a broad spectrum of innovation. Please note that these descriptions are based on my understanding from attending the pitch presentations.</p><p>With visions extending from electric vehicle integration to AI-driven emotion detection, the future seems to be in capable hands. The commitment to revolutionize industries, disrupt norms, and shape the future was palpable. The road to 2030 looks to be an exciting one!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52cd4f6ce202" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Increase the sound volume for Ubuntu above 100% with a single line.]]></title>
            <link>https://medium.com/@saloni_garg/increase-the-sound-volume-for-ubuntu-above-100-with-a-single-line-ed3c550d26b5?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/ed3c550d26b5</guid>
            <category><![CDATA[ubuntu]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Mon, 13 Jul 2020 07:21:03 GMT</pubDate>
            <atom:updated>2020-07-13T07:21:03.555Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>TLDR; this is that one line:</strong></p><pre>pactl -- set-sink-volume 0 120%</pre><p>Here, 120% is the desired sound level.</p><p>This command interacts with the PulseAudio sound server, to set the sound level.</p><p>Also, you can set x% as your default volume, so that you don’t need to use the command over and over. For this, you’ll have to edit the <strong><em>/etc/rc.local </em></strong>file.</p><h4>Steps to follow:</h4><ul><li><em>Move to the etc directory: </em><strong>cd /etc</strong></li><li><em>Open the rc.local file in gedit: </em><strong>sudo -H gedit /rc.local</strong></li><li><em>Add the line of command above to this file.</em></li></ul><pre>#!/bin/sh -e<br>#<br># rc.local<br>#<br># This script is executed at the end of each multiuser runlevel.<br># Make sure that the script will &quot;exit 0&quot; on success or any other<br># value on error.<br>#<br># In order to enable or disable this script just change the execution<br># bits.<br>#<br># By default this script does nothing.<br><br><strong>pactl -- set-sink-volume 0 120%<br></strong><br><strong>exit 0</strong></pre><ul><li><em>Save and exit.</em></li></ul><p>You’re good to go! 👍</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ed3c550d26b5" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[University Outreach Initiative by GNOME]]></title>
            <link>https://medium.com/@saloni_garg/university-outreach-initiative-by-gnome-1998005b7bba?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/1998005b7bba</guid>
            <category><![CDATA[student-programs]]></category>
            <category><![CDATA[gnome]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Fri, 26 Jun 2020 13:42:51 GMT</pubDate>
            <atom:updated>2020-06-26T13:48:11.444Z</atom:updated>
            <content:encoded><![CDATA[<p>Students form an important part of an Open Source organization, as interns, active members, mentors, volunteers, and whatnot.</p><p>To acknowledge their contributions and to make the Open Source environment more inclusive for them, we<strong> </strong>at GNOME <strong>started the ‘University Outreach Initiative</strong>’.</p><h4>Our main objective behind this is to involve the wide student community and to make them feel comfortable starting their Open Source journeys with GNOME.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/371/0*sQPzedPQRg_y7-8H.png" /></figure><p>In the beginning, our plan of action was to start with conducting virtual events which will help us get to know more about the students, their interests, and what we can do to make the onboarding process more wholesome for them.</p><h4>I and a colleague Amit, stepped up to conduct <a href="https://events.gnome.org/event/10/">the very first event</a> under this initiative. Kristi helped us with the arrangements, while Martin was the one who volunteered to deliver a talk at the event.</h4><p>Having mentored Google Summer Of Code participants for more than five years, what could be a better topic to talk about? He shared his insights on how students can get started with GNOME Foundation’s codebase, for the GSoC Program. I talked about some opportunities for students, to make their quarantine period more productive and Amit shared some general tips for preparing for Google Summer Of Code program.</p><h4>We received a greater-than-expected amount of registrations and had a good time answering the queries of the students.</h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/994/1*ENMV1bEmYe54A3uKJApKlw.png" /></figure><h4>Our <strong>key takeaways</strong> from this small interaction with students, were:</h4><p>- Students believe that Open Source brings endless learning opportunities, with experienced mentors to learn from and wide multi-discipline projects to work on.<br>- They value professional interactions with people in the corporate world and OpenSource provides an abundance of such opportunities.<br>- Open Source helps them expand their network and grow beyond the classroom environment. <br>- Needless to say, OpenSource projects can take a resume to the next level, as they are a good resource for professional experience.<br>Moreover, classroom knowledge is far different from the skills that are actually deployed in the corporate world, and Open Source is a good way to cover them up, for free.</p><h3><strong>What’s up next?</strong></h3><p>Our prime focus is still on making the onboarding process inclusive and comfortable for students. We’ll come up with more University events in the coming months, with new content and topics to brainstorm upon.</p><p>Also, we’re looking for GNOME Ambassadors from different university campuses worldwide, to join us in the mission of making GNOME community more diverse.</p><p><a href="https://saloni-garg.github.io/Projects-Portfolio/GNOME-Amb.html"><em>Use this form to express your interest in the GNOME Ambassadors Program, and someone from the team will follow up.</em></a></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/818/1*VhEK1EN9dCOnBTrvkAKj7Q.png" /></figure><p>For more information, you can check out our wiki page: <a href="https://wiki.gnome.org/Engagement/UniversityOutreach">https://wiki.gnome.org/Engagement/UniversityOutreach</a></p><p>Cheers.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1998005b7bba" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Calculating profitability and risk of investments using Data Science.]]></title>
            <link>https://medium.com/@saloni_garg/calculating-profitability-and-risk-of-investments-using-data-science-addfa83020ad?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/addfa83020ad</guid>
            <category><![CDATA[applied-finance]]></category>
            <category><![CDATA[tutorial]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Tue, 09 Jun 2020 14:30:30 GMT</pubDate>
            <atom:updated>2020-06-09T14:30:30.217Z</atom:updated>
            <content:encoded><![CDATA[<p>Hello!</p><p>Today, I’m writing about a Data Science project I recently completed, in one of my favorite fields, <strong><em>Applied Finance. </em>I used Data Science to calculate and compare profitability and risk of different investments using the Sharpe Ratio.</strong></p><h3>Introduction to Professor William Sharpe</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/232/0*Tc2NKyGztbAYeCb8.jpg" /><figcaption>Image Courtesy: NobelPrize.org</figcaption></figure><p>William Forsyth Sharpe is an American economist. He is the STANCO 25 Professor of Finance, Emeritus at Stanford University’s Graduate School of Business, and the winner of the 1990 Nobel Memorial Prize in Economic Sciences.</p><p>He introduced the <a href="https://web.stanford.edu/~wfsharpe/art/sr/sr.htm"><em>reward-to-variability ratio</em></a> that soon came to be called the Sharpe Ratio.</p><h4>Sharpe Ratio</h4><p>It compares the expected returns for two investment opportunities and calculates the additional return per unit of risk an investor could obtain by choosing one over the other. In particular, it looks at the difference in returns for two investments and compares the average difference to the standard deviation (as a measure of risk) of this difference. A higher Sharpe ratio means that the reward will be higher for a given amount of risk. It is common to compare a specific opportunity against a benchmark that represents an entire category of investments. The Sharpe ratio has been one of the most popular risk/return measures in finance, not least because it’s so simple to use. It also helped that Professor Sharpe won a Nobel Memorial Prize in Economics in 1990 for his work on the capital asset pricing model (CAPM). Let’s learn about the Sharpe ratio by calculating it for the stocks of the two tech giants Facebook and Amazon. As a benchmark, we’ll use the S&amp;P 500 that measures the performance of the 500 largest stocks in the US.</p><p><strong>Before we compare an investment in either Facebook or Amazon with the index of the 500 largest companies in the US, let’s visualize the data, so we better understand what we’re dealing with.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/542/1*ViCwrlmCJRx8CkBso4JWBw.png" /></figure><p><strong>Let’s also take a closer look at the value of the S&amp;P 500, our benchmark.</strong></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/506/1*CTfFtvJfFEphPcto26b9dQ.png" /></figure><p><strong>The inputs for the Sharpe Ratio: Starting with Daily Stock Returns</strong></p><p>The Sharpe Ratio uses the difference in returns between the two investment opportunities under consideration. However, our data show the historical value of each investment, not the return. To calculate the return, we need to calculate the percentage change in value from one day to the next. We’ll also take a look at the summary statistics because these will become our inputs as we calculate the Sharpe Ratio.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/518/1*GnmppyRG2v1moafEiJ0wuA.png" /></figure><h3><strong>Daily S&amp;P 500 returns</strong></h3><p>For the S&amp;P 500, calculating daily returns works just the same way, we just need to make sure we select it as a Series using single brackets [] and not as a DataFrame to facilitate the calculations in the next step.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/512/1*c6lhBwmvNrW1cYAFopqg6w.png" /></figure><h3>Calculating Excess Returns for Amazon and Facebook vs. S&amp;P 500</h3><p>Next, we need to calculate the relative performance of stocks vs. the S&amp;P 500 benchmark. This is calculated as the difference in returns between stock_returns and sp_returns for each day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/520/1*k9aMZYRKgEEX7oBCSdykGA.png" /></figure><h3>The Sharpe Ratio, Step 1: The Average Difference in Daily Returns Stocks vs S&amp;P 500</h3><p>Now we can finally start computing the Sharpe Ratio. First we need to calculate the average of the excess_returns. This tells us how much more or less the investment yields per day compared to the benchmark.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/629/1*JjHCP3EI2UK4ylBJ9m7umg.png" /></figure><h3>The Sharpe Ratio, Step 2: Standard Deviation of the Return Difference</h3><p>Next, we calculate the standard deviation of the excess_returns. This shows us the amount of risk an investment in the stocks implies as compared to an investment in the S&amp;P 500.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/726/1*8RHOzDufsWIGlCLqs4QK8g.png" /></figure><h3>Putting it all together</h3><p>Now we just need to compute the ratio of avg_excess_returns and sd_excess_returns. The result is now finally the <em>Sharpe ratio</em> and indicates how much more (or less) return the investment opportunity under consideration yields per unit of risk. The Sharpe Ratio is often <em>annualized</em> by multiplying it by the square root of the number of periods. We have used daily data as input, so we&#39;ll use the square root of the number of trading days (5 days, 52 weeks, minus a few holidays): √252</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/730/1*cAceSmTkBPTgH8UjntnP1A.png" /></figure><h3>Conclusion</h3><p>Given the two Sharpe ratios, which investment should we go for? In 2016, Amazon had a Sharpe ratio twice as high as Facebook. This means that an investment in Amazon returned twice as much compared to the S&amp;P 500 for each unit of risk an investor would have assumed. In other words, in risk-adjusted terms, the investment in Amazon would have been more attractive.</p><p>This difference was mostly driven by differences in return rather than risk between Amazon and Facebook. The risk of choosing Amazon over FB (as measured by the standard deviation) was only slightly higher so that the higher Sharpe ratio for Amazon ends up higher mainly due to the higher average daily returns for Amazon.</p><p>When faced with investment alternatives that offer both different returns and risks, the Sharpe Ratio helps to make a decision by adjusting the returns by the differences in risk and allows an investor to compare investment opportunities on equal terms, that is, on an ‘apples-to-apples’ basis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/583/1*LhGsCtika5AXf8xbI9_l4w.png" /></figure><p>I hope you liked it. The code and all the related resources are in <a href="https://github.com/saloni-garg/Data-Science-Investment-Risk-And-Returns">this repo</a> :)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=addfa83020ad" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Have we forgotten the ‘Right to be Forgotten’?]]></title>
            <link>https://medium.com/@saloni_garg/have-we-forgotten-the-right-to-be-forgotten-3a768a2e58a1?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/3a768a2e58a1</guid>
            <category><![CDATA[right-to-be-forgotten]]></category>
            <category><![CDATA[mozilla]]></category>
            <category><![CDATA[internet]]></category>
            <category><![CDATA[right-to-privacy]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Mon, 08 Jun 2020 09:47:18 GMT</pubDate>
            <atom:updated>2020-06-08T09:47:18.354Z</atom:updated>
            <content:encoded><![CDATA[<p>Hola!</p><p>I hope you all are safe in this pandemic, and are able to adjust to this new normal that is just surfacing.</p><p>Today, I want to take a few minutes to talk about the ‘Right to be Forgotten’. As Wikipedia defines, “it reflects the claim of an individual to have certain data deleted so that third persons can no longer trace them.” Did you just catch yourself thinking that it doesn&#39;t matter much to me, because hey, who am I hiding from? Who’s gonna put efforts in tracing me, lol?</p><p>Well, here’s my personal experience, that dates back to…<strong>the day before yesterday</strong>. Yes, it is this recent! And this is what prompted me to write this article today, even though I was thinking of it for almost a year :/</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zt_TOoL6SX9uUSl1LBbfdw.png" /><figcaption>Screenshot of my mailbox (Captured on 8th June, 2020)</figcaption></figure><p>I checked this email in the evening, and had I been in a hurry, I would have jumped straight to opening the attachment, since I’m anticipating some Google docs from unknown email addresses anyhow. There was no security threat notification from Google too, and it appeared like just another email. at first glance.</p><p>Mentioning it for the record, I don’t know anyone named “Peter Paul” or the other two people CC’ed in this thread. I reported it as ‘Phishing scam’ and thought of moving on to some other work. But this question that where did he get my email address from, kept me from switching tabs.</p><p>I googled another two email addresses CC’ed in this mail, and was shocked to see Google throw up a reference to a PDF containing a long list of emails in both the search results on CourseHero.com</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/777/1*NM3LV0GopMB7YEtuRZq0zA.png" /><figcaption>Google Search Result for one of the email addresses.</figcaption></figure><p>This document contains thousands of email addresses and is uploaded by someone named Dinesh Reddy, again a name I’ve never heard of.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/514/1*gWhlcpDfaKDlX12EC9KLjg.png" /></figure><p>However, I can do nothing about it, except reporting the PDF at CourseHero. This is now CourseHero’s choice whether or not to remove that content, and I’m almost 99.99% sure that just one flag won’t catch their eye.</p><p>Fortunately, in my case, nobody was harmed and I came out safe from this malicious attempt. But many people can’t. Have you read about <strong>Google Spain v AEPD and Mario Costeja González </strong>incident? It’s worth your 5 minutes, give it a read, if you haven’t yet.</p><p>Search Engines have become the controllers of our data now, which is, needless to say, <strong>dangerous</strong>. Okay, does “Right to Be Forgotten” conflict the “Right to Know”? Let me know your views.</p><p>The fine lines and rationales being used to negotiate these rights are dependent on the context, existing laws, and how the judiciary actually favor privacy v/s public interest. The ones most affected by this are the individuals with a substantial internet presence.</p><p>It all boils down to which concept to be used to declare the Internet: as a common concern, as a global public good, or as part of the common heritage of mankind.</p><p>With all my heart, I appreciate Mozilla’s efforts in bringing awareness to this topic of <strong>“The Internet as a Global Public Resource”</strong>. I’ve been there as a <strong>Mozilla Open Leader</strong>, and trust me, I personally have never seen an organization more sorted and excited to make the Internet open and accessible to all.</p><p>I hope their efforts come to fruition, and we all can enjoy a <strong>Healthy Internet. </strong>Also, this is my request to all the ones who have been turning a deaf ear to their data privacy, this is the time, be aware that no anti-virus software can save you, once these malicious attackers have your personal data. Stay informed, Stay safe!</p><p><strong>Update</strong>: I used <a href="https://support.google.com/websearch/troubleshooter/9685456#ts=2889054%2C2889099">this form</a> on Google’s help page and got my email address removed from appearing in search results :)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/802/1*d4dWmrWUqXUrjw82Iz8wOw.png" /></figure><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a768a2e58a1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My experience at Linux App Summit 2019]]></title>
            <link>https://medium.com/@saloni_garg/my-experience-at-linux-app-summit-2019-a298fefb4321?source=rss-5291774e0411------2</link>
            <guid isPermaLink="false">https://medium.com/p/a298fefb4321</guid>
            <category><![CDATA[conference]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[linux]]></category>
            <category><![CDATA[open-source]]></category>
            <category><![CDATA[linux-apps]]></category>
            <dc:creator><![CDATA[Saloni Garg]]></dc:creator>
            <pubDate>Fri, 29 Nov 2019 23:15:22 GMT</pubDate>
            <atom:updated>2019-11-29T23:15:22.685Z</atom:updated>
            <content:encoded><![CDATA[<p>Two weeks back, I visited the city of Barcelona with a friend, Amit for the much-awaited Linux App Summit. It was organized by GNOME Foundation and KDE Community together. I was excited to be on the organizing team for the summit. It was my first experience being on the organizers’ side and not attendees&#39; side, for such a large-scale event, and it was a really great experience.</p><p>We landed in Barcelona, it was quite cold. Yet another time, I had wrongly estimated the temperature of a foreign place. The place where we stayed was pretty warm though. Just after keeping our luggage, we went out for breakfast. Luckily, we found an Indian restaurant very close by, that served us really delicious Shaahi Paneer :P</p><p>The evening was reserved for a pre-registration event for everyone to get to know each other. It was a fun event, I met people like Nuritzi, Sri, Aleix, Adriaan, Ismael, Jorge, Kristi, Pilar and many more, with whom I had been in constant touch over Telegram regarding the conference preparations. After a great evening, we, the ones handling Marketing and PR activities, sat down and worked on Tweets.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o8Xyy41Mq132ylp4XSfc2A.jpeg" /><figcaption>(From Left) Amit, Me and Nuritzi</figcaption></figure><h3>We knew tomorrow’s gonna be interesting.</h3><p>The next day was the very first day of the conference. We reached the venue earlier and everyone just voluntarily assigned themselves some work and started preparing for the conference. I was going to be the anchor, and the stage handling person. Nuritzi quickly gave me some tips on how to do everything.</p><p>Everyone was full of a spirit of Open-Source, great speakers stepped up the stage, some interesting questions were asked, the venue was beautiful, and people were all very warm. It was a beautiful day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*F_S4p-GRyT4eZboQP6SoDw.jpeg" /></figure><p>Next day, the conference was great as usual. I interviewed some of the speakers, about their experience. There was this question I asked almost everyone: <strong>“What’s the first word that comes to your mind when I say Open Source?” </strong>Well, undoubtedly, most of the answers were ‘Community’ and ‘Freedom’. It was an interesting bit of the day.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RHixZzJMe9zSAcbwuC5ArA.jpeg" /><figcaption>Me Interviewing Conference Speakers</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-9heRaXfPSTzqS0CV1EmvQ.jpeg" /><figcaption>City Tour Barcelona</figcaption></figure><p>We had a cool city tour in Barcelona in the evening, hosted by the summit.</p><p>It was “cool” literally too :P</p><p>We had a look at some really interesting architecture of Barcelona. But honestly, nothing caught my eye more than “Official Merchandise store of FC Barcelona”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CruNkgGiH_Tv5hdpUhgiig.jpeg" /></figure><p>The next day, I and Amit gave a lightning talk. Since I was involved in mic-coordinating and Amit, in camera handling, we couldn’t get much time to prepare. But it was anyway a lightning talk, that had to be lightning fast.</p><p>The next day was the hard one, as we had to say bye to many people since it was the last day of the conference. I just can’t stress more on how much I love and cherish this beautiful experience I had at LAS.</p><p>It was a fun-filled atmosphere all around. There were all techies, but the laughter and smiles in the room were enough to tell that nobody was in their “geek mode”.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*CPIMW7jFFXFj_VUkEXWRkw.jpeg" /><figcaption>Me and Amit with Shola</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NukQZMIQG4B64JK4LTvJ3Q.jpeg" /><figcaption>Some amazing friends I found at LAS. (From Left) Anissa, Pilar and Me.</figcaption></figure><p>The next day, we had a quick BoF session in the morning. It was about what went well at the summit and ideas on how we can improve it next year. Everyone wrote their points on sticky notes and pasted on the whiteboard and then an interesting discussion followed where we all discussed next year&#39;s plans in detail.</p><p>Later that day, we went out for a city tour again. Nuritzi and Sri had looked up some places to visit. I, Amit, Pilar, and Shola were eager to join them! We had lots of fun at the Montjuic Castle, in the cable car, at the cafe, in Barcelona all together. I will let the pictures tell the rest of the story!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*73BGWM7Lz9zBOLjEgto7ZA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tVCQrKbWgOmoGxZSNUq-fw.jpeg" /></figure><p>All in all, it was a beezer time indeed. Thanks to everyone involved, for making it such an amazing experience.</p><p>Looking forward to seeing you all next year at LAS 2020 :)<br>Thanks for reading, take care.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a298fefb4321" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>