Avichal Dwivedi

Pilani Junior | 3x Hackathon Winner

[Avichal Dwivedi @Avichal2205]

@mrtbrgl hello

[Avichal Dwivedi @Avichal2205]

read about NCM over the weekend and tried running an experiment on top of it. it didn't work, and i never even got to the reasoning part

NCM if you haven't seen it: https://t.co/DAxb9v2Ygm

the idea: NCM writes text rough first, then fills in detail level by level, and each level gets locked once it's written. i wanted to give it a scratch space to think before each lock

but to test that you first need a codec, something that turns an answer into NCM's codes and back into text. if it can't give back the exact same answer, any reasoning result on top is fake. so i set a bar: 1,990 of 2,000 new answers back perfectly

trained it on 50k countdown math solutions and tested on 2k it never saw. best was 94%, so 115 answers came back broken. training longer made it worse, 83% by the end. and 99.87% of characters were right, one wrong digit still breaks the whole answer

my guess is half the time in training i hid the detailed codes but still asked for the exact answer. the rough codes can't carry that, so it learned to guess. it got worse on training data too, so it's not overfitting. haven't tested this yet, next run is hiding off vs on

whole thing cost $0.41 of GPU. if you've trained codecs like this, would love to hear what you'd try next

[Avichal Dwivedi @Avichal2205]

read about NCM over the weekend and tried running an experiment on top of it. it didn't work, and i never even got to the reasoning part

NCM if you haven't seen it: https://t.co/DAxb9v2Ygm

the idea: NCM writes text rough first, then fills in detail level by level, and each level gets locked once it's written. i wanted to give it a scratch space to think before each lock

but to test that you first need a codec, something that turns an answer into NCM's codes and back into text. if it can't give back the exact same answer, any reasoning result on top is fake. so i set a bar: 1,990 of 2,000 new answers back perfectly

trained it on 50k countdown math solutions and tested on 2k it never saw. best was 94%, so 115 answers came back broken. training longer made it worse, 83% by the end. and 99.87% of characters were right, one wrong digit still breaks the whole answer

my guess is half the time in training i hid the detailed codes but still asked for the exact answer. the rough codes can't carry that, so it learned to guess. it got worse on training data too, so it's not overfitting. haven't tested this yet, next run is hiding off vs on

whole thing cost $0.41 of GPU

@KumbongHermann @ishanskhare really enjoyed the paper. did you run into anything like this with nested dropout when training the tokenizer? would love to know if i set something up wrong

[Hermann Kumbong @KumbongHermann]

🚀 Introducing Next-Concept Modeling (NCM): our new paradigm for fast, parallel, coarse-to-fine text generation.

Compared to strong diffusion language models (DLMs) at similar parameter size, NCM achieves:

⚡ Up to 31× faster generation on downstream tasks like summarization and machine translation with comparable or better quality.

🏎️ 10–21× faster unconditional generation across TinyStories, LM1B, and OpenWebText, with comparable or better generative perplexity and using 3–5× fewer training tokens.

🧩 Better performance on reasoning tasks like Sudoku, outperforming both AR and diffusion baselines with fewer sampling steps.

With @ishanskhare, @IddahMlauzi, @ekellbuch, @simran_s_arora, @Azaliamirh, and @HazyResearch.

More details in the thread 🧵👇

[Avichal Dwivedi @Avichal2205]

@willzero yeah one in two runs is basically a coin flip. curious if the subagent version was more stable or just slower

[Avichal Dwivedi @Avichal2205]

@NeuralBricolage same boat, ~50 followers here. starting to think the replies are the real building in public

[Avichal Dwivedi @Avichal2205]

@sughanthans1 conway's law but for agents. every team ships its own subagent and the orchestrator turns into the meeting nobody wanted

[Avichal Dwivedi @Avichal2205]

@Suryanshti777 yeah the model is the easy part now. most of my time goes into the boring bits like who's allowed to touch what and when

[Avichal Dwivedi @Avichal2205]

@qiweiiiy agents.md as an attack vector is so obvious in hindsight. reading every config file like it's trusted was always going to end like this

[Avichal Dwivedi @Avichal2205]

@Ch0pin what happens when two of them edit the same file at once. that was the first thing that broke for me when i put multiple agents on one codebase. had to add locks

[Avichal Dwivedi @Avichal2205]

@marclou 10 days is insane. was it the audience or did the product just click faster

[Avichal Dwivedi @Avichal2205]

@poteto @shaoruu been doing this with my own bots. one finds leads, one writes outreach, one watches x for people struggling with support. feels like having a tiny team

[Avichal Dwivedi @Avichal2205]

@TheSuranaverse open weights is the real news. going to try it as a coding agent next to claude code when it drops, will post how it does

[Avichal Dwivedi @Avichal2205]

@ibuildthecloud per agent permissions finally. i was hacking this together myself for running multiple codex workers

[Avichal Dwivedi @Avichal2205]

@thdxr if tokens get that cheap, running stuff on device starts looking really good

[Avichal Dwivedi @Avichal2205]

@HamelHusain same thing happened with mine, support agent ended up more useful than the coding one. how are you even evaluating marketing output

[Avichal Dwivedi @Avichal2205]

@sh_reya the swarm part breaks way before the metrics do. my agents kept overwriting each other's files till i added locks

[Avichal Dwivedi @Avichal2205]

@thdxr p50 looks great, p95 cooked by one tray. classic. you routing around it or just waiting for the swap?

[Avichal Dwivedi @Avichal2205]

@deedydas honestly feels like scheduling matters more than raw tokens/sec here. whoever packs the gpus better wins

[Avichal Dwivedi @Avichal2205]

Hey! I’m Avichal, a BITS Pilani student, ex-Expedia and Siemens, and a former founding AI engineer. My experience spans Node.js, TypeScript, React/Next.js, Python, AI agents, RAG and vector databases. I’ve also led teams to multiple global hackathon wins with ~$18k in prizes.

I built LatchGrid, an AI canvas workspace that brings websites, videos and documents together: https://t.co/HYGAvQY0bz
Demo: https://t.co/4m1XfRUORi
Portfolio: https://t.co/uqCEv4FyTs
Resume: https://t.co/6I8eJY9CjD