Aditya 🌌

AI Engineer at Singapore's largest metro operator

[OpenAI @OpenAI]

Introducing dots, powered by GPT-6 Astra.

Remarkably capable, always-on agents built to handle everything.

[Aditya 🌌 @adtygan]

https://t.co/TZwAqWxZ8G

This article means that an S3-based LLM inference engine is very much possible once memory hierarchy and scheduling algos are sorted out

[Aditya 🌌 @adtygan]

@Drewch @tobi Makes a lot more sense now, thanks!

[Aditya 🌌 @adtygan]

@brianchew congrats brian!

[Aditya 🌌 @adtygan]

@AbhishekPaul_11 @craftian_keskin So kino

[Aditya 🌌 @adtygan]

@dmcgco @AbhishekPaul_11 gotta check this out

[Ashwarya Maratha @AshwaryaMaratha]

Introducing @PhyseraAI, an applied research and product lab working on the next layer of useful intelligence. We are focused on efficient AI systems, high-fidelity behavioral simulations, and turning model understanding into reliable outcomes.

[Aditya 🌌 @adtygan]

I'm working on long context compression research. I have access to GPU computer cluster, and use Windsurf as my IDE. I think having access to the Max coding plan will help me 10x my experiments. The DeepWiki and Codemaps have been a game changer. Will be sure to cite the support in paper 😅

[Teppo Felin @teppofelin]

Data-first thinking is a trap

Here's the full introductory chapter https://t.co/dWwdZhvP32

[himanshu @himanshustwts]

if the recipe for good RL + reasoning is no longer locked inside 5 frontier labs, then a new class of entity might become a sophisticated buyer of data, envs and ai infra. we can call it as the forward-deployed vertical AI startup / neolab.

HOW?

> take a team of 5-10 strong MLEs who understand the post-training stack.
> pick a large, high-value white-collar vertical (insurance. legals, financial, compliance, healthcare, funds etc).
> build an ai-native product that automates meaningful chunks of that workflow.
> use the recipe which is already public (GRPO/RLVR + domain-specific environments + rubric-based rewards from domain experts) to train agents that are genuinely good at the vertical task.
> W!

[sky @notsky]

hey i’m still looking for a job for when i graduate around may

have experience in backend and full stack roles in the past, dm if ure hiring in singapore or know someone who is

retweets are appreciated 💙🙏

[cozy @vec0zy]

@nodcodes i did this to mintlify too
https://t.co/V8soZPIDfG

[Aditya 🌌 @adtygan]

lol this was fun

[itch.io @itchio]

1,000,000 shrimp: incremental shrimp boating

https://t.co/RAjf1Dab3M

[Aditya 🌌 @adtygan]

@k7agar Is it really that good?

[Aditya 🌌 @adtygan]

@ChinmayKak Congrats!

[Aditya 🌌 @adtygan]

Presented our work at @RealAAAI today. First proper conference paper for me, so this was special. Will be attending tomorrow and the day after as well.

Lmk if anyone's coming!

[Aditya 🌌 @adtygan]

@ivanleomk Mirrors the thoughts shared by a few others on my feed, including an article by the founder of Browser Use (https://t.co/gM8FfVF8Qe)

[Gregor Zunic @gregpr07]

https://t.co/ovNlDea7BS

[Aditya 🌌 @adtygan]

@danieltsk Okay this is a cool idea. Never thought of using LLMs this way. Quite useful!

[Aditya 🌌 @adtygan]

3/ The tool I made aims to teach these skills, opting for the simplest trick when applicable at any given game state.

The prototype is pretty rudimentary at this stage. I'll push more updates soon!

[Aditya 🌌 @adtygan]

Okay so I play LinkedIn games quite often and hate getting smoked by my friends in Mini-Sudoku. So I made a Mini-Sudoku coach with the free @AmpCode credits I been getting. And my game's been improving slowly. Will try to ship it to public soon. More details below 👇🏻

[Aditya 🌌 @adtygan]

Gonna try this today. Never used Amp before, and this got me excited

[Aditya 🌌 @adtygan]

2/ The techniques I found online can be grouped in levels as:

Level 1: Hidden singles (scanning), Naked singles (last remaining)
Level 2: Box/Line reduction, Naked pairs/triples
Level 3: Hidden pairs/triples
Level 4: XY-chains, Coloring, AIC
Level 5: Forcing chains