saumit

Ex-10 year old | MS EE @ncstate | ex AI @DeepTekAI |
Only personal views on this acc |
data science

[saumit @5aumit]

making an effort to take music a little seriously now. started a page on Instagram. pls follow 🫠🤝

(ps - I will be releasing something new soon)

https://t.co/BwKiNbi7Ja

[Mayor Zohran Kwame Mamdani @NYCMayor]

.@NYCSanitation I'd like to report a sweep

[saumit @5aumit]

never gonna forget an author of Attention shouted out my undergrad uni lol

[Aidan Gomez @aidangomez]

The only MIT relevant to our generation of artificial intelligence is Manipal Institute of Technology.

[saumit @5aumit]

documenting my progress, bugs and learnings on LinkedIn.
also maintaining a diary.md file with all manually typed notes as I partially vibe code this.

Will also try sharing here!

[saumit @5aumit]

been a long time since I got my hands dirty with something new. i've been learning AI engineering with this project lately.

Floki is a CLI agent to understand MLflow logs.

https://t.co/UOdd3tJniD

[saumit @5aumit]

@vasuman one question. the application says that sponsorship requirement will not be used to make hiring decision, but you mentioned "must be eligible to work in the US". anything I'm misunderstanding?

[saumit @5aumit]

@kuppanoodle from rd 1, yes TOs. but I think defense needs to ramp up too against OKC, and solid playmaking from Bron and Smart since Luka's out.

[saumit @5aumit]

@kuppanoodle no stress either I'm just glad I get to see king play hard with the rest of the team. i didn't have them winning it all this year. at this point onwards, everything is a + for me. gonna be a great and fun series against OKC

[saumit @5aumit]

@kuppanoodle i agree

[Abraham @the_don_abe]

I’m on the side of twitter where animals are trolling themselves

[saumit @5aumit]

highly agree. before the AI boom, the only thing I did that ran for hours was model training. but because it was all on remote servers, I could start a run in the evening and hop on RDR2 while getting training updates on slack with callbacks. life was good.

[Kevin Naughton Jr. @KevinNaughtonJr]

one of the many reason why you should do all your development remotely over ssh

-open and close your devices and work continues in the cloud
-connect and disconnect freely on all your different machines (including phone) and pick up exactly where you left off
-increases your battery life since nothing runs locally
-never need to reconfigure machines since the only one that matters is in the cloud
-worry less about your stuff: if something gets lost/stolen/breaks you won't care (aside from cost) since you machines are just portals to your remote devbox
-if you're ever without wifi you can work locally if you sync your files across devices with syncthing and changes are synced when you're back online
-many other benefits of having a cloud machine like using tailscale with it as a VPN/exit node and much more

try this workflow for a couple months and i guarantee you'll never go back to developing software locally

[saumit @5aumit]

@bansalg_ Perhaps there needs to be an "understanding" stage for a dev in the SDLC now that agents write a lot. One where you spend some time (proportional to the amt of generated code) to not write any more, but simply break down what's present. slowing down may help ship better?

[saumit @5aumit]

I spend at least 5-6 mins writing my initial prompt to start off a project. I spend one session on just "ask" or "plan" mode, but never just set it off after that. I still generate one class/function at a time. I might ship 2x faster instead of 10x, but I won't be disconnected.

[saumit @5aumit]

I never got to the point where I run multiple agents unsupervised to write my code. I always want to look at the code, understand at least 50% and then run it to verify it works. I don't understand how shipping 10x more would make me or my software better. It'd just be "more".

[Rahul Jain @rahulj51]

Not fully by hand but have definitely changed my process to reduce unattended long horizon yolo tasks and switched to more deliberate, step by step authoring. Also no more parallel subagents.

The biggest culprit of slop, imo, is agent generated plans and specs. These should be treated as throw away artifacts that are constantly evolving as you write more code.

My current process limits any spec to 2-3 pages max. And I have tailored Claude responses to be always in simple english eli10) and not more than 3 sentences (sometimes a single paragraph). And only single-level bullet lists of max 5 items. Feels slower but works better for complex systems.

[saumit @5aumit]

its funny how im back to the grind 4 yrs later and everything I said, and my friends said in the replies and QTs here is relevant for me again. im repeating my job-search phase and re-learning everything from back then.
I'm glad all my friends contributed, it helps me today.

[saumit @5aumit]

Placements are gonna start for incoming 4th years in 1-2 months, so here's a thread of tips I got from my seniors and some things I learnt on my own. If you think any of your followers might benefit, do RT.

[Peter Steele @petersteele]

What the hell are yall doing that you need to spend so much on AI for?

I literally have 1 sub, paid through work, for 100 a year to GH Copilot. I literally go into each new billing cycle with less than 70% used. Right now, its end of April and I am not even at 50%.

I can't fathom multiple hundreds a month in AI subs and still needing more, like WTF are yall doing?

[Sherwood @shcallaway]

OVERRATED: running tons of agents in parallel; working on too many things at once; perpetual context-switching; opening lots of low-quality PRs that may never land.

UNDERRATED: using one or two agents at a time; focusing on the task in front of you; thinking deeply; finishing stuff; making your code works in prod.

[saumit @5aumit]

@debofr_ I feel you mate. happy for you :) it's a humbling experience, isn't it?

[skooks @skooookum]

“I { built | vibecoded }
a { website | dashboard }
that lets you { interact with | visualize } <some publicly available data>”

[saumit @5aumit]

@bansalg_ @bakkermichiel I've been wondering if, in the long run, the value of expedited development with AI will be more valuable than more diligent development at a controlled pace. While AI can do what junior devs can do today, the future "senior" devs come from the same pool, right?