Swarit

DSA + Shipping open source AI-Infra as a beginner (currently @LiteLLM) | Learning in Public . Papers , PRs , failures.

[Swarit @SwaritS20]

@nitesh_singh5 Interested

[Swarit @SwaritS20]

Every senior engineer has an idea they've wanted to exist for years and never had a free weekend to build.

I have the one resource you don't: time.

Final-year student, currently shipping PRs to LiteLLM. Reply with the thing you never got around to building. I'll build it in public — commits, failures, the whole thing — and credit you on every post.

Worst case you get to say "told you it wouldn't work."

Best case it exists.

[Swarit @SwaritS20]

you and Clem might be pointing at the same axis from different sides — his sample-size scaling and your per-sample intelligence feel like scaling-law vs measurement framings of one thing.

the part I keep getting stuck on when I try to explain this in threads: none of the current benchmarks can measure it without confounding scale. is there work on this you'd point at, or is that still the gap?

[Swarit @SwaritS20]

the "same registers in the same order" finding is the part I want to see unpacked. is there a stable ordering to which registers degrade first — colloquial before formal, formal before literary, something else — or is "same order" just "same set, damaged in whatever sequence each run"?

asking because if there's a genuine ordering, that's a fingerprint of how narrow-distribution training damages models, which is a bigger result than the specific TamilLM case.

[Swarit @SwaritS20]

your 95% accurate AI pipeline is not a 95% accurate product.

0.9 × 0.9 × 0.9 = 0.729

every "let me add one more layer" is compounding failure. your app hits 95 in dev and 70 in prod and nobody in the room can tell you which step is bleeding, because nobody in the room wrote it — Claude did.

95 is the ceiling of what you can ship without understanding. everything past it is theory.

[Swarit @SwaritS20]

@andrewho03 how do you find the "basic but valuable" gaps before a lab tells you?
feels like the hard part isn't building the dataset, it's knowing which unglamorous thing every 3rd user hits. is that coming from lab conversations, model eval traces, or something else

[Swarit @SwaritS20]

@ShalsX the "study more before I build" trap is so specific.
mine took a different shape — I'd solve a DSA problem, feel productive, and use that feeling as an excuse to not open the project repo for another day. green square as a decoy.
what's the repo btw, curious what you picked

[Swarit @SwaritS20]

in 15 minutes? roll back the last deploy. it's almost always the last deploy. if it isn't, you now have 10 minutes and a cleaner search space.
everyone in these replies is answering the system design textbook version of this question. the on-call version has a different first move.

[Swarit @SwaritS20]

@auto_grad_ on non-verifiable tasks "best checkpoint" is doing all the work in this sentence.
best by reward model or best by you re-reading the outputs and going yeah that one

[Swarit @SwaritS20]

@YuDai_Tsai matches expire if you don't respond within the context window

[Swarit @SwaritS20]

@yashwanthsai29 Been thinking about this from the other side — a lot of "AI agent" demos I see fail for exactly this reason.
The agent isn't the hard part. The eval set is. And you can't build an eval set for a process nobody wrote down. Tracking is the actual moat.

[Swarit @SwaritS20]

Every senior engineer has an idea they've wanted to exist for years and never had a free weekend to build.

I have the one resource you don't: time.

Final-year student, currently shipping PRs to LiteLLM. Reply with the thing you never got around to building. I'll build it in public — commits, failures, the whole thing — and credit you on every post.

Worst case you get to say "told you it wouldn't work."

Best case it exists.

[Swarit @SwaritS20]

feeling of completing a project you worked days n nights off is much greater than being in love.

[Swarit @SwaritS20]

Good one !
Open weights aren't the risk - they're the safeguard.

[Jensen Huang @JensenHuang]

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter.

AI will transform every industry, power every company, and be built by every country.

Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty.

The world needs both frontier closed models and frontier open models.

https://t.co/AUKzoQ5Ikb

[Swarit @SwaritS20]

Curious how the Tool Retriever decides relevance before the LLM even sees the query intent — is it embedding-similarity search over tool descriptions, a smaller classifier model, or rule-based routing?
That retrieval step feels like it could become its own bottleneck at 1000+ tools.

[Swarit @SwaritS20]

On week 3 - before you have a real client's data , how are people actually constructing a credible golden dataset for evals ? Hand-labelling synthetic exception cases yourselves , or bootstrapping off existing open-source agent benchmarks and swapping in real examples once you land a client?

[Andrej Karpathy @karpathy]

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

[Dorsa @dorsa_rohani]

We made Claude Code multiplayer!

Use it with other people in the same terminal, and connect your Claudes so they can all talk to each other.