20 | engineer
Joined May 2024
- Tweets 689
- Following 199
- Followers 638
- Likes 822
118 Photos and videos
Pinned Tweet
I built a GPT from scratch in PyTorch, trained on Kabir ke dohe. 🧘
Tokeniser, self-attention, multi-head attention, transformer blocks, inference, all hand-written, no transformer libraries.
Wrote up the full journey in blog below along with Github :
1
1
71
I built a GPT from scratch in PyTorch, trained on Kabir ke dohe. 🧘
Tokeniser, self-attention, multi-head attention, transformer blocks, inference, all hand-written, no transformer libraries.
Wrote up the full journey in blog below along with Github :
1
1
71
Blog : blog.anuragmaurya.com/blog/g…
Github : github.com/justanuragmaurya/…
1
41
What if ISRO used AI to redesign their frontend?
I rebuilt it in one prompt using Cursor fal ai MCP.
The workflow:
→ describe the vibe
→ fal ai generates assets on the fly
→ Cursor codes the whole thing.
@isro @cursor_ai @fal
1
3
128
Models Used :
@cursor_ai 's composer2.5, (for coding)
@fal /seedance2.0(bytedance for video)
@fal /nano-banana-2(google - for image assests)
127
Anurag Maurya retweeted
May 18
I have a confession to make...
I accidentally built world's best subtitle API.
Free @fal API credits for first 350 users who RT comment "sub".
88
67
142
35,755
Anurag Maurya retweeted
modi ji built his portfolio in jquery and you guys are fighting whether to use next or vue.
just ship fast , framework / techstack is not important.
1
4
147
Anurag Maurya retweeted
Apr 19
hey anthropic
if you really wanna flex your models
clone the entire adobe creative suite
i dare you
277
564
19,835
498,551
Anurag Maurya retweeted
Apr 20
Kimi K2.6 Launched: Open-Source Competing with Frontier Models
- Agentic coding king: #1 on SWE-Bench Pro (58.6), beating GPT-5.4 xhigh (57.7), Gemini 3.1 Pro (54.2), Claude Opus 4.6 (53.4)
- Open-source SOTA: Moonshot AI released weights code on HuggingFace first open model competing at frontier level in complex agentic tasks
- Extreme long-horizon: handles 4,000 tool calls over 12-hour runs, manages full-stack repos, DevOps, performance tuning across 100 files
- Massive agent swarms: supports up to 300 parallel sub-agents for large-scale autonomous execution (upgrade from K2.5)
- Frontend mastery: strong in motion-heavy UI WebGL shaders, Three.js, GSAP, Framer Motion
- Multimodal prowess: 54.0 on HLE (with tools), 93.2 on MathVision, strong beyond just coding
- Where it lacks: behind Muse Spark in visual factuality (SimpleVQA), slightly behind GPT-5.4 & Gemini 3.1 Pro in top-tier reasoning
- Cost advantage: local runs → no API cost, better privacy, more control
Open-Source is now seriously competing not replacing frontier models yet, but closing the gap fast
5
3
36
3,466
Apr 20
-> Kimi k2.6 vs Opus 4.7
-> Both still making the same mistake as GPT-1,2
-> imagine these models taking decision in the Defence sector.
50
-> you can find out the plots of movies and series by using NLP and some basic maths.
-> i used Word2Vec to create embedding of the content in "Game of Thrones" books.
-> found out the main plot of the series. (IFYKYK)
2
66
Anurag Maurya retweeted
Mar 31
Anthropic CEO, wasn't lying when he said
“I have engineers within anthropic who don’t write any code, they just let Claude write the code and they edit it and look it over”
he gave us live examples and respect for the transparency 🫡
Mar 31
Claude code source code has been leaked via a map file in their npm registry!
Code: pub-aea8527898604c1bbb12468b…
138
347
4,570
644,237
Mar 31
what a day !
-> using a custom version of claude code for FREE using smartest and latest chinese llm model.
thanks to openrouter( for free llm ) and who ever was responsible for leaking the source code.
2
8
230
Mar 31
on my way to add chinese uncensored model support in this.
Mar 31
say hello to free-code
claude code source code fully recompiled, telemetry stripped, security guardrails prompts stripped, all working experimental features enabled
including ultraplan mode - a new async agentic planning mode where claude starts a multi-agent research session in the browser lasting between 10 and 30 minutes
since i know there are gonna be DMCA strikes on this i've uploaded it on the blockchain on IPFS
1
2
355
Mar 31
we're getting more access to open knowledge than ever : )
Mar 31
Security has completely gone off the rails! And most might not even realize it, because if you look at just the last month, the number of serious incidents across the AI dev ecosystem is honestly insane.
Axios, a package with over 100M weekly downloads, was allegedly hijacked and used to ship a RAT. LiteLLM, which sits deep inside many modern AI stacks, was backdoored on PyPI with a multi-stage credential harvester across ~97M monthly installs.
Railway ended up leaking authenticated user data due to a CDN misconfiguration. There’s now a disclosure around command injection in OpenAI Codex via branch names, which should make anyone building agentic coding systems pause for a second.
And it doesn’t stop there because Mercor AI allegedly had ~939GB of source code and ~4TB of total data exfiltrated! There are claims around Copilot injecting ads into 1.5M GitHub PRs without developer consent, and even the Claude Code source code has been leaked!
If you zoom out instead of looking at these as isolated incidents, the pattern becomes very obvious. Core infrastructure packages are getting compromised, supply chains are being poisoned, CI/CD pipelines are being used as entry points, and even the systems that are supposed to guarantee trust are failing.
We have effectively built an ecosystem where installing a dependency is equivalent to running arbitrary code, dependencies are just layers of transitive blind trust, and all of this is happening inside environments that are loaded with high-value secrets like cloud credentials, API keys, and production access, and then we casually stack hundreds of these dependencies without questioning it.
The attack surface is no longer your application or your infrastructure, it’s your dependency graph, and that’s what makes this so dangerous because it scales insanely well for attackers. You don’t need to break into systems anymore, you just compromise one package, one maintainer, or one pipeline, and the blast radius takes care of itself.
Focus on security!
1
45
Mar 31
imagine if they accidentally leak there opus4.6's models weights , that would be diabolical.
Mar 31
Claude code source code has been leaked via a map file in their npm registry!
Code: pub-aea8527898604c1bbb12468b…
1
68
Mar 31
thank you opus 4.6
Mar 31
Claude code source code has been leaked via a map file in their npm registry!
Code: pub-aea8527898604c1bbb12468b…
1
73
Anurag Maurya retweeted
Mar 31
Claude code source code has been leaked via a map file in their npm registry!
Code: pub-aea8527898604c1bbb12468b…
3,335
7,533
48,506
35,676,226