iiinova — where your digital self keeps evolving

ashwiniag @ashwiniag.bsky.social · 18/08/2026

The models(GPT, Claude, Llama) we use are pre-trained instances, trained offline on massive compute clusters over months to set millions/billions of static weights 🤯

010

ashwiniag @ashwiniag.bsky.social · 18/08/2026

key-vocabs: Inside a single forward pass: • Tokenization: Chunk text into tiny pieces. • Vectors: Turn tokens into numerical arrays. • Attention: Let vectors interact & update meaning by context. • Feed-Forward: Apply weights/biases to store factual relationships.

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

The biggest misconception: "It learns while we chat." Nah. Weights are frozen at inference (you chatting with Claude/GPT). The "memory" during a session is just your entire chat history getting re-fed as context every turn. When the session ends, state clears. Zero live weight updates.

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

The actual loop happens at generation time (autoregression): Model predicts 1 next token. Token gets appended to your prompt. Entire sequence feeds back through the stack to predict the next token. It repeats this string until hit with a stop signal. #LLM

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

How the architecture flows: I was wrong to assume single circular loop inside a layer. It’s a stacked assembly line (e.g., 96 layers in GPT-3). Each layer refines the token vectors further before passing them to the Feed-Forward Network where factual/associative knowledge lives.

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

Where the contextual understanding happens: Attention. Take the word "model." Is it a fashion model or a computer processor? Self-attention dynamically updates a token’s vector representation based on surrounding tokens. The vector literally shifts based on context.

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

Deep in transformers now. Core idea: Everything gets broken down into tokens, mapped into vectors, & processed as raw numbers. calculating probability distributions over those tokens to figure out what comes next, sample from it, repeat. An architecture built to make that prediction good. 🧵 #LLM

110

ashwiniag @ashwiniag.bsky.social · 18/08/2026

So Backprop, net-net math magic aside: forward pass = process input→output normally. Then, output back to input, calculate exactly how much each weight contributed to the error to assigning blame. Then nudge each weight by how much it messed up.

111

ashwiniag @ashwiniag.bsky.social · 18/08/2026

Now picture a word as a point in that^ space, the direction you look from it gives you meaning. Specific directions = human-comprehensible concepts: gender, plurality, tense, even sentiment. wow wow wow... I feel like a kid again! #neuralnetworks

111

ashwiniag @ashwiniag.bsky.social · 18/08/2026

Mind is BLOWN by how many dimensions these models actually work in. I could only ever picture up to 3D thanks to math teacher but turns out that's a massive understatement! GPT-3's embedding space is 12,288-dimensional... Every word is a point with 12,288 coordinates. 🤯. You kidding! #neuralnets

111

ashwiniag @ashwiniag.bsky.social · 17/08/2026

Deep in 3Blue1Brown's neural network series and it's wrecking every mental model I had of "black box AI." Weights and biases, backpropogation... finally starts to make sense when you see them animated instead of just algebra on a page. Absolute gold! #AI #LLM www.3blue1brown.com?topic=neural...

3blue1brown.com

3Blue1Brown

Mathematics with a distinct visual perspective. Linear algebra, calculus, neural networks, topology, and more.

281

ashwiniag @ashwiniag.bsky.social · 22/07/2026

Some albums you listen to. Some albums listen to you. 🎧 Wrote this for a prompt from underline.center/categories, an album that got me through a hard time. ashwiniag.com/some-albums-listen-to-you/

ashwiniag.com

Some Albums Listen to You

This post is a response to a writing prompt from Underline Center: "Write about an album that got you through a hard time." It would be an injustice to pick just one album! Hard times have a tendency...

010

ashwiniag @ashwiniag.bsky.social · 16/07/2026

Back from life's trenches. And looking forward to my present world. Lots of work ahead, back to building. 🖖

010

ashwiniag @ashwiniag.bsky.social · 16/06/2026

Oasis! They are my go to. To this day am obsessed with youtu.be/6hzrDeceEKc?...

youtu.be

Oasis - Wonderwall (Official Video)

YouTube video by Oasis

010

ashwiniag @ashwiniag.bsky.social · 16/06/2026

do they though? i've rarely seen teams nail it. they've got an uptime metric for it, which isn't the same thing. you can be 99.99% up while every request quietly writes to /dev/null. nines don't matter if users aren't happy.

100

ashwiniag @ashwiniag.bsky.social · 15/06/2026

Because the generic health-check just pings the kitchen door to see if it’s locked. If you only check if a process is alive instead of monitoring work completed (throughput/latency), you're blind to the staring-at-lettuce problem. Systems show so many symptoms before they crash.

100

ashwiniag @ashwiniag.bsky.social · 14/06/2026

That's your CPU core; alive, spinning, completely useless. And the scary part is from outside the kitchen, everything looks fine. like core is technically still "running" (not crashed/dead) but it's completely stuck doing one thing and can't process anything new,starving the rest of your application

000

ashwiniag @ashwiniag.bsky.social · 14/06/2026

explaining CPU soft-lock to stakeholder in incident review Imagine a chef in a kitchen. Normally: take order → cook → serve → repeat. A soft-lock is when the chef gets stuck obsessively checking one ingredient. Forever. They're not dead but technically "working" and nothing is getting served #sre

310

ashwiniag @ashwiniag.bsky.social · 14/06/2026

Vague postmortems always = vague engineering culture If an org can’t be precise, transparent, and system-focused when things break publicly, they aren’t doing it internally either. Incident reports are the ultimate window into a company's true technical debt, operational maturity, and trust levels.

000

ashwiniag @ashwiniag.bsky.social · 14/06/2026

🔻 Rebuttal 4: "It was just an operational glitch, we don't need a formal postmortem" Or "Keep incident report vague so we don't look bad to leadership" Small-minded stakeholders view incident reports as public damage control. Pragmatic culture view them as an architectural red pill #RebuttalSeries

100

ashwiniag @ashwiniag.bsky.social · 06/06/2026

Curiosity and excitement are the highest forms of aliveness. They only appear when the nervous system finally exhales. Finding them again after an extremely long time, is its own kind of homecoming.

000

ashwiniag @ashwiniag.bsky.social · 01/06/2026

fair point. bucket overhead dominates at scale. The value saving is real but smaller than the tweet implied. Good correction.

100

ashwiniag @ashwiniag.bsky.social · 31/05/2026

Go has no native Set. map[key]bool vs map[key]struct{} are not same If you're storing millions of event IDs in map[int]bool instead of map[int]struct{}, you're wasting millions of bytes unnecessarily. At that scale it can tip pods over the memory limit & major incident #golang #sre #buildinpublic

270

ashwiniag @ashwiniag.bsky.social · 28/05/2026

One thing that took me a while to internalize in Go: make([]string, n) is is a declaration of intent You're telling the runtime: I know exactly how much space I need, give it to me now. Skip it and index directly → panic Small thing but expensive mistake in a hot path. #golang #sre #buildinpublic

170

Reposted by ashwiniag

Justin Garrison @justingarrison.com · 26/05/2026

I've been doing Linux for a long time. This is the first time I heard of Martian Packets I know what they are, I just didn't know they were called that en.wikipedia.org/wiki/Martian...

en.wikipedia.org

Martian packet - Wikipedia

193

ashwiniag @ashwiniag.bsky.social · 26/05/2026

Took a team from 450 alerts/week to 10. Cut security tooling cost from $10K to $420/year.Built RBAC that made a SOC audit a non-event Opening a few consulting spots. If your engineers are firefighting instead of shipping, let's talk. ashwiniag.com/consulting/ #SRE #Consulting #PlatformEngineering

ashwiniag.com

Consulting

Available for consulting. Specializing in reliability architecture, release engineering, security automation, and infra for B2B SaaS startups

020

ashwiniag @ashwiniag.bsky.social · 17/05/2026

Policies built on the assumption that inequality is"not that bad" may actually be making things worse. IMO this is something to be looked at seriously. Incredible work done by economics professors & bringing data to light. 🙌 www-thehindu-com.cdn.ampproject.org/c/s/www.theh... #india #inequality

000

ashwiniag @ashwiniag.bsky.social · 16/05/2026

I want to make something that moves people. And I want to be fully alive while making it.

011

ashwiniag @ashwiniag.bsky.social · 12/05/2026

Am glad you asked. recently, the discomfort of "not-yet-knowing" while working through commit log internals is what crystallised this for me. Wrote a thread on what that breakthrough produced ( a mental model): bsky.app/profile/ashw...

000

ashwiniag @ashwiniag.bsky.social · 12/05/2026

I'm mid-chapter 3 of building a distributed log in Go. This clicked while sitting with the discomfort of not-yet-knowing. Realising how good engineers actually learn and how badly ed-systems prepares you for it, this mindshift is real work #BuildInPublic #LearningInPublic #golang #distributedsystems

020

ashwiniag @ashwiniag.bsky.social · 12/05/2026

When you learn from others' failures, you're front-loading the pattern so your first instinct is calibrated, not naive and better learning curve. The goal isn't to avoid all pain. It's to not repeat someone else's avoidable mistakes while earning your own. #EngineeringMindset

100

ashwiniag @ashwiniag.bsky.social · 12/05/2026

catch: Education systems trains us: mug → output → marks. This model produces people who know facts but can't transfer them, because scoring is the goal. Can't see that a networking problem rhymes with a caching problem. Unlearning this is half the work. 🤷‍♀️

100

ashwiniag @ashwiniag.bsky.social · 12/05/2026

The template that actually compounds knowledge: Concept → Decision → Tradeoff → Where else does this appear? Ask the why at every step. Not "what does this do?" but "why does it work this way, and what breaks if I change it?" #SystemsThinking #EngineeringMindset

100

ashwiniag @ashwiniag.bsky.social · 12/05/2026

The reason senior engineers seem like they "just know" things: They've broken enough systems that new problems rhyme with old pain. It's pattern recognition built from failure, one need not to live it all again. #buildinpublic #sre #EngineeringMindset

110

ashwiniag @ashwiniag.bsky.social · 12/05/2026

🧵 "pattern recognition" is the real output of learning, not information retention. #EngineeringMindset #sre #SystemsThinking

110

ashwiniag @ashwiniag.bsky.social · 11/05/2026

the store wraps os.File with bufio.Writer, why? ashwiniag.com/why-bufio-wr... #golang #distributedsystems #buildinpublic #sre

010

ashwiniag @ashwiniag.bsky.social · 11/05/2026

writing this one was fun. every syscall crosses user space → kernel space. as SRE you see it as p99 spikes, CPU sys time climbing, context switches ticking up in node metrics, by the time it's on your dashboard, it's already compounding ashwiniag.com/the-cost-of-... #golang #sre #observability

ashwiniag.com

The cost of a syscall

Part of the Building the Log Package series How I thought through the log package design before writing any code Syscalls are not free they involve a context switch from user space to kernel space. Y...

110

ashwiniag @ashwiniag.bsky.social · 11/05/2026

🧵 I just finished building the store layer of a commit log, took longer amidst life' chaos but learned more than expected while building this. wrote 4 posts on what actually clicked: design, syscalls, layers with example, buffering.

100

ashwiniag @ashwiniag.bsky.social · 06/05/2026

Good engineers are the ones who have asked why more times and sat with the discomfort of not knowing until they genuinely figured it out. #buildinpublic

251

ashwiniag @ashwiniag.bsky.social · 06/05/2026

Friendship doesn't exempt people from accountability. In fact it requires more of it and not less.

000

ashwiniag @ashwiniag.bsky.social · 21/04/2026

The ability to regulate emotions independently is a valuable skill, not a character flaw.

010

ashwiniag @ashwiniag.bsky.social · 21/04/2026

Intentions live inside the person who acts. Impact lives inside the person who receives.

000

ashwiniag @ashwiniag.bsky.social · 21/04/2026

Have the capacity for self-examination the willingness to sit with the discomfort of being wrong, of having caused hurt, of your own blind spots. Without that willingness a person stays exactly where they are, and everyone who gets close enough pays the price of their stagnation...you won't evolve

010

ashwiniag @ashwiniag.bsky.social · 21/04/2026

you don't cut people out of coldness. If you didn't care it would be easy. The people who make the cutting hard are the ones who had real moments of warmth and genuine connection mixed in with intense harm. That mixture is what keeps you second-guessing.

000

ashwiniag @ashwiniag.bsky.social · 21/04/2026

Clean cruelty is easier to walk away from than inconsistent love. The grief is the proof of how much you tried. #boundaries #healing #selfworth

100

ashwiniag @ashwiniag.bsky.social · 21/04/2026

The hardest part is not dealing with emotionally immature people. It's the grief of accepting that someone you genuinely love doesn't have the capacity to meet you where you are. That grief is real. That's a specific kind of loss. And it's okay to grieve it. #mentalhealth #boundaries #selfworth

010