Tom Mathews (🧑‍💻+🐍+🤖🧠)

ML/AI Developer & Advocate, a passionate photographer, an avid reader and tech enthusiast...
- SDE 🧑‍💻
- Python 🐍
- Machine Learning 🤖🧠

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

"If there are really infinite possibilities in infinite universes, then every possibility have a chance to be an outcome, but it is not necessary that every possibility might become an outcome." ~ "The infinite probabilities of the infinite possibilities" #RandomThoughts 🤔🤔🤔

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

I built archex: local code context for AI agents. It indexes a repo and returns a ranked, token-budgeted bundle of symbols, imports, graph neighbors, and provenance instead of making Claude/Cursor grep through files.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

Install:

uv tool install archex
archex doctor

Repo: https://t.co/NUCKzJUcwz

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

It seems, @ZEE5India is doing load testing on their system during the group stage match. You can see the great quality of #vibecoded integration across the app.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

@saurabh_dhall As of now, it has been multiple times more than just 2 minutes since the complete crash.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

It was bound to happen. @Zee5 literally dropped the ball during #BRAvsMOR match #WorldCup2026

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

@trq212 Doesn't this potentially open up a much larger surface for injection attacks? Also, the benefits of moving to HTML come with a heavy tax on token usage. I don't think we can or should afford that high tax.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

@sydchess @karpathy Retrieval for both the user and the agent made trivial. 11x token reduction without losing retrieval quality is huge for building personalized agents that don't break the bank on context windows.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

A noble heart is no substitute for a clear mind; purpose requires both the courage to act and the wisdom to weigh the cost.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

@rubenhassid "the era of prompt engineering is over" followed by an 8-step prompt framework is peak irony. good context in, good output out — you don't need a ceremony for it.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

⭐ https://t.co/KIHtpknSn1

MIT licensed. PRs welcome.

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

Most ML engineers can't implement attention from scratch.

Not because it's hard. Because they've never had to.

`https://t.co/ZcUWyWeoem()` is the most dangerous line in machine learning.

Thread 🧵

[Tom Mathews (🧑‍💻+🐍+🤖🧠) @tom_mathews]

→ Junior devs: fastest path from `https://t.co/ZcUWyWeoem()` to actually understanding what's underneath
→ Senior engineers: the one-file reference you wish existed
→ Interview prep: being able to implement these from scratch puts you in the top 1%