Nilay

317 posts

Nilay banner
Nilay

Nilay

@localm_tuts

AI Researcher | 2x Native AI Startups | Follow for tutorials, podcasts, and hacks to keep your skills sharp. For Hindi follow @localm_hi

London | San Jose Katılım Aralık 2025
25 Takip Edilen15 Takipçiler
Nilay
Nilay@localm_tuts·
@HemantDotDev vscode - never support whol lock the door.
English
1
0
1
52
Hemant
Hemant@HemantDotDev·
Be honest, which IDE is the best in this AI era ?
Hemant tweet media
English
33
2
33
2.1K
Nilay
Nilay@localm_tuts·
@mr_r0b0t @xyster @MemoryReboot_ @intel @NVIDIAAI tobe honest, tokenmaxxing has no meaning unlese prefixed with `local` - localtokenmaxx I am at 260M tokens, and only $8 (premium - external) ;) 3xGB10 (one cluster) – various models + 1xGB10 (Qwen 27B or 35B A3B)
Nilay tweet media
English
0
0
2
60
Nilay
Nilay@localm_tuts·
@mr_r0b0t @Tono_Ken3 Yes, I assume for the size of minimax and supporting reasonable context for multiturn (with a 20% token increase turn by turn to support up to 4-6 turns with 64K context), you are looking at 396GB VRAM for 4-6 parallel streams.
English
0
0
2
66
mr-r0b0t
mr-r0b0t@mr_r0b0t·
@Tono_Ken3 You need 256GB for this, I was nearly OOM 😭
English
3
0
10
775
mr-r0b0t
mr-r0b0t@mr_r0b0t·
Are you considering a @NVIDIAAI DGX Spark or GB10? Looking for tips and best practices for the one you have? Want to show off your new projects? Join the newly formed DGX (GB10) User Group! I'll be there and happy to help as best I'm able ♥️ x.com/i/chat/group_j…
English
19
5
70
7.7K
ÆON FORGE ✨
ÆON FORGE ✨@SpaceTimeViking·
@NVIDIAAI is cooking up something “Ultra” and this could be their big break. The post training model has so much potential distilling the weights and data down to the purest form possible. Isolating the signal and removing the noise. Scaling that up could be a big deal.
NVIDIA AI@NVIDIAAI

@TheAhmadOsman 👀 "Ultra" ⏳️

English
1
1
9
440
Nilay
Nilay@localm_tuts·
@thegenioo if you have one - great, if you haven't that's not blockers. key is multiturn accuracy. one can have image agent.
English
0
0
0
11
Nilay
Nilay@localm_tuts·
FYI. this is experimental for dgx sparks only - optimised with modelopt from BF16 image, and post quant-optimised. I haven't benchmark to validate loss (but my hunch is it is lossless) - as NVFP4 was < 0.02 it use llamacpp's nvfp4 underlaying... huggingface.co/nilayparikh/Qw…
English
0
0
0
19
Nilay
Nilay@localm_tuts·
I am so impressed with llamacpp and especially when combined mtp and ngram-mod - qwen 3.6 27b - fly up to 30 toks/s @ 192K context size, on single node DGX spark. With far better accuracy. FYI. I am running my own experimental model - details in thread
Nilay tweet media
English
1
0
1
79
Nilay
Nilay@localm_tuts·
@KaiXCreator everyother. only nonserious actors use that
English
0
0
0
12
Kaito
Kaito@KaiXCreator·
Can you name a programming language better than JavaScript ?
English
159
2
107
14K
Nilay
Nilay@localm_tuts·
I get this question often: Why are you capturing all your in/out via proxy even when you are a lone person using this? Lone or not lone, it doesn't matter; an agentic conversation with an LLM, along with user inputs, is a gold mine. Next 2 years are all about RLHF.
English
1
0
2
22
Nilay
Nilay@localm_tuts·
@ash_twtz They have no value, instead they need to keep mastering interface. Yes they should build and invest into mid and small models, as it will increase demand for their prosumer based.
English
0
0
0
80
Mr Ash
Mr Ash@ash_twtz·
What’s stopping NVIDIA from creating the world’s best AI model?
Mr Ash tweet media
English
154
6
165
11.4K
Nilay
Nilay@localm_tuts·
@gregpr07 Can it meet my acceptance criteria? Yes than rest is noise as legally can buy a service.
English
0
0
0
4K
Gregor Zunic
Gregor Zunic@gregpr07·
I think I know why deepseek is so good
Gregor Zunic tweet media
English
272
153
5.9K
441.5K
Nilay
Nilay@localm_tuts·
@ylecun Because there will be midterms in few months !
English
0
0
0
36
Nilay
Nilay@localm_tuts·
Most beautiful image, when comes to agentic solutions :D
Nilay tweet media
English
0
0
1
12
Nilay
Nilay@localm_tuts·
@wieslawsoltes do you know `kill switch` - gh copilot has turned on
English
0
0
0
79
Wiesław Šoltés
Wiesław Šoltés@wieslawsoltes·
Gemini 3.5 Flash in GitHub Copilot has 14x multiplier basically on par with Claude Opus 4.7 🤡
Wiesław Šoltés tweet media
English
6
0
12
1.2K
Nilay
Nilay@localm_tuts·
When coding agents do retrospective at end of week, what went well, what was off. Now updating agents.md and agent definition. that's my claw moment, when they are in discussion what skill will improve.
English
1
0
1
21

Keşfet

@HemantDotDev @mr_r0b0t @xyster @MemoryReboot_ @intel @NVIDIAAI @Tono_Ken3 @SpaceTimeViking