Ramith Padaki

ML engineer by day. Only tweet on blue moons.

[Ramith Padaki @ramith41p]

"load bearing"
"provenance"

[Ramith Padaki @ramith41p]

This character `—` (em-dash) is the most common AI tell in Feb 2025. Interesting phenomena where a few years ago it was exclusive to grammar purists, and now it's a ~95% accuracy classifier for AI use.

[Ramith Padaki @ramith41p]

New indicator that's bothering me because of the extent it's everywhere online now:

Not x but y

*This cadence is not only everywhere, it's a dead giveaway of model output(s)"

[Ramith Padaki @ramith41p]

I didn't think I'd like coding-agent kanban, but I do. It's like a bunch of post-its that ask me how they can go from left to right. Abstractions over abstractions over abstractions over...

[Alec Radford @AlecRad]

By the way - I think a valid (if extreme) take on GPT-2 is "lol you need 10,000x the data, 1 billion parameters, and a supercomputer to get current DL models to generalize to Penn Treebank."

[Ramith Padaki @ramith41p]

An incessant need to pass coding benchmarks on first-pass-based metrics and prompts filled with "EXCEPTIONS MEAN CRITICAL BUSINESS FAILURE" has led us down this path of models that could use a break 😔

[Andrej Karpathy @karpathy]

I don't know what labs are doing to these poor LLMs during RL but they are mortally terrified of exceptions, in any infinitesimally likely case. Exceptions are a normal part of life and healthy dev process. Sign my LLM welfare petition for improved rewards in cases of exceptions.

[Ramith Padaki @ramith41p]

👀

[Cursor @cursor_ai]

Sonnet 4.5 is now available in Cursor!

We're excited to hear what you think.

[Jack Morris @jxmnop]

first i thought scaling laws originated in OpenAI (2020)

then i thought they came from Baidu (2017)

now i am enlightened:
Scaling Laws were first explored at Bell Labs (1993)

[Ramith Padaki @ramith41p]

Not sure who took over the dspy account, but the sass is🔥

[DSPy @DSPyOSS]

unbelievable, someone should build this — a whole language for writing software that learns natural-language rules and examples from execution traces

[Ramith Padaki @ramith41p]

Bm25 will save us all haha

I'm biased on the subject, but this is true imo.. Access to knowledge will be a concern for as long as models don't have perfect memory, and I don't see that on the horizon

[Omar Khattab @lateinteraction]

AI/ML are to AI software what algorithms are to systems (e.g., databases).

One can study fundamental algorithms and data structures (sorting, trees, ...).

Or study database systems, which compose algorithms into a whole new space: joins, transactions, query optimization, etc.

[Ramith Padaki @ramith41p]

Why am I bad at using R1?

My experience with sonnet is 10x R1 at all times and all tasks. 🤔

[Ramith Padaki @ramith41p]

Surely I'm better than Claude at it

Probably not 🥲

[Ramith Padaki @ramith41p]

Suddenly feel like playing Pokémon

[jacob @jsnnsa]

Ah! I see the issue now. Let me just fix that
Ah! I see the issue now. Let me just fix that
Ah! I see the issue now. Let me just fix that
Ah! I see the issue now. Let me just fix that
Ah! I see the issue now. Let me just fix that

[Ramith Padaki @ramith41p]

Man, it's fast! To the point where it's jarring to see a response that quick

[Cerebras @cerebras]

le chat 🤝 le wafer

[Ramith Padaki @ramith41p]

@nuwandavek 😭

[Ramith Padaki @ramith41p]

Why oh why 😭

[Dirk Nowitzki @swish41]

😳

[Ramith Padaki @ramith41p]

@nuwandavek Scary good

[Ramith Padaki @ramith41p]

@SparSamp Lol. My point was partly that I don't get to do as much cool stuff as I want/wish to anymore

[Ramith Padaki @ramith41p]

Living life in the green bubbles

[James Cowling @jamesacowling]

2. Your job is the strategy stuff and the dirty work stuff. All the cool stuff in the middle is for everyone else. You're not too senior to carry a pager or respond to outages, this keeps you in touch with how things are going.