Dark | Light
[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

[@vikhyatk](/creator/twitter/vikhyatk)
"are you kidding me i can't put tr inside table now fuck nextjs fuck technology fuck you all i'm done"  
[X Link](https://x.com/vikhyatk/status/1979048755270705567) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T04:56Z 23.1K followers, 482.5K engagements


"you don't build in "stealth mode" because you're worried about someone stealing your idea you do it because "announcing your raise / coming out of stealth" is a press event that generates a lot of eyeballs you want to have a product to direct those eyeballs to"  
[X Link](https://x.com/vikhyatk/status/1980004006220165432) [@vikhyatk](/creator/x/vikhyatk) 2025-10-19T20:11Z 23.1K followers, 18.1K engagements


"framework XX review XX hours in: - very easy to assemble everything snaps into place with magnets - doesn't fit in my laptop bag if i could redo i'd get a smaller one - should've gotten a USB-A expansion card for installing the OS - feeling very at home with native tiling wm"  
[X Link](https://x.com/vikhyatk/status/1975978841001935291) [@vikhyatk](/creator/x/vikhyatk) 2025-10-08T17:37Z 23K followers, 6476 engagements


"trying to benchmark moondream vs gemini XXX flash on cost struggling because gemini refuses to transcribe every document I've tried so far 😔"  
[X Link](https://x.com/vikhyatk/status/1978902465924993117) [@vikhyatk](/creator/x/vikhyatk) 2025-10-16T19:14Z 23K followers, 7765 engagements


"@rajankrx well i actually do know how they did it they built a cutting edge inference-focused compiler i'll let @matthewjgunton chime in with more details"  
[X Link](https://x.com/vikhyatk/status/1979002007143801003) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T01:50Z 23K followers, XXX engagements


"i'm on arch linux why are you looking up mac os docs"  
[X Link](https://x.com/vikhyatk/status/1979078166590140854) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T06:52Z 23K followers, 1626 engagements


"why did you get 96GB of RAM what are you going to do with it idk maybe ill run two electron apps at the same time"  
[X Link](https://x.com/vikhyatk/status/1975793940625563752) [@vikhyatk](/creator/x/vikhyatk) 2025-10-08T05:22Z 23.1K followers, 123K engagements


"losing $20b a year with $13b in revenue stop looking at spreadsheets. go try out codex. if anything theyre undervalued"  
[X Link](https://x.com/vikhyatk/status/1978732332442947894) [@vikhyatk](/creator/x/vikhyatk) 2025-10-16T07:58Z 23.1K followers, 195.9K engagements


"@HeadBetwixKnees yeah the ui is a bit confusing but it has a lot of features"  
[X Link](https://x.com/vikhyatk/status/1980054560191492540) [@vikhyatk](/creator/x/vikhyatk) 2025-10-19T23:32Z 23.1K followers, XX engagements


"Just shipped a 4-bit quantized version of Moondream. 🌘 XX% less memory. 🌓 XX% faster inference. 🌔 XXXX% accuracy"  
[X Link](https://x.com/vikhyatk/status/1925344054046597430) [@vikhyatk](/creator/x/vikhyatk) 2025-05-22T00:12Z 23K followers, 51.8K engagements


"how do i stop being underrated this feels like a failure. i want to be overrated"  
[X Link](https://x.com/vikhyatk/status/1970935848448426212) [@vikhyatk](/creator/x/vikhyatk) 2025-09-24T19:38Z 23K followers, 44.1K engagements


"the thing i love about codex is i can have it spend X hours do a complex refactor look at the resulting code and just say yeah hmm not feeling the data model here let's redo it a completely different way with a human dev you just have to merge it otherwise they'll get upset"  
[X Link](https://x.com/vikhyatk/status/1974927014693089336) [@vikhyatk](/creator/x/vikhyatk) 2025-10-05T19:57Z 23K followers, 224.9K engagements


"made a cute logo for our internal inference engine"  
[X Link](https://x.com/vikhyatk/status/1978306985486397530) [@vikhyatk](/creator/x/vikhyatk) 2025-10-15T03:48Z 23K followers, 4204 engagements


"I still dont know what Salesforce really does"  
[X Link](https://x.com/vikhyatk/status/1978500061056172524) [@vikhyatk](/creator/x/vikhyatk) 2025-10-15T16:35Z 23K followers, 13.1K engagements


"cant believe how good clickhouse is"  
[X Link](https://x.com/vikhyatk/status/1978862577376489575) [@vikhyatk](/creator/x/vikhyatk) 2025-10-16T16:36Z 23K followers, 1916 engagements


"@init_malachi my standard pattern used to by dynamodb master and elasticsearch for analytics. clickhouse is a better elasticsearch replacement. but dont make it master"  
[X Link](https://x.com/vikhyatk/status/1978944511888806283) [@vikhyatk](/creator/x/vikhyatk) 2025-10-16T22:01Z 23K followers, XX engagements


"when they find out you're still using claude sonnet instead of grok code"  
[X Link](https://x.com/vikhyatk/status/1979055135998214557) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T05:21Z 23K followers, 7128 engagements


"turns out this is not a good way to run a business"  
[X Link](https://x.com/vikhyatk/status/1979250512680882615) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T18:17Z 23K followers, 11.6K engagements


"compared to the "Gated Attention for Large Language Models" paper X. we have G4 and G2 but not G1 X. we allow gating to be conditioned on log(pos) (i actually think it should be sqrt(log(pos)) but that's a topic for a different post). this helps with long ctx"  
[X Link](https://x.com/vikhyatk/status/1969478049227030631) [@vikhyatk](/creator/x/vikhyatk) 2025-09-20T19:05Z 23K followers, 37.5K engagements


"i think G2/G4 help with exploding attention logits whereas G1 helps with attention sparsity/sinks. i also see a parallel between G1 and the gpt-oss attention sinks except S is now q-dependent instead of static"  
[X Link](https://x.com/vikhyatk/status/1969478051563257945) [@vikhyatk](/creator/x/vikhyatk) 2025-09-20T19:05Z 23K followers, 1733 engagements


"cloud providers shouldn't have good ux their margins should be so thin they can't afford designers"  
[X Link](https://x.com/vikhyatk/status/1978272665937875218) [@vikhyatk](/creator/x/vikhyatk) 2025-10-15T01:32Z 23K followers, 4744 engagements


"LPIPS (typically) uses frozen AlexNet features with a learned linear head. Trained on 484k MTurk samples asking people which of two distortions was closer to a sampled 64x64 patch. Used as a component of the training loss. It's surprising we're still using this 2017-era proxy"  
[X Link](https://x.com/vikhyatk/status/1978276566565101977) [@vikhyatk](/creator/x/vikhyatk) 2025-10-15T01:47Z 23K followers, 4937 engagements


"Excited to release a preview of Moondream X. A 9B param 2B active MoE vision language model that makes no compromises; offering state-of-the-art visual reasoning while still retaining an efficient and deployment-friendly form factor"  
[X Link](https://x.com/vikhyatk/status/1968800178640429496) [@vikhyatk](/creator/x/vikhyatk) 2025-09-18T22:11Z 23.1K followers, 225K engagements


"having worked on a lot of "enterprise grade" code in my career all i'll say is you shouldn't treat it as a compliment"  
[X Link](https://x.com/vikhyatk/status/1979774300967317591) [@vikhyatk](/creator/x/vikhyatk) 2025-10-19T04:59Z 23.1K followers, 6339 engagements


"@ludwigABAP when i turn XX i will be posting on subs*ack not here"  
[X Link](https://x.com/vikhyatk/status/1979794223324094822) [@vikhyatk](/creator/x/vikhyatk) 2025-10-19T06:18Z 23.1K followers, XXX engagements


"our login broke. went to sst to check why theirs was also broken. turns out us-east-1 is down"  
[X Link](https://x.com/vikhyatk/status/1980171953614012448) [@vikhyatk](/creator/x/vikhyatk) 2025-10-20T07:19Z 23.1K followers, 13K engagements


"Overtokenized transformers "Through extensive experiments we uncover a log-linear relationship between input vocabulary size and training loss demonstrating that larger input vocabularies consistently enhance model performance regardless of model size.""  
[X Link](https://x.com/vikhyatk/status/1977814490386428246) [@vikhyatk](/creator/x/vikhyatk) 2025-10-13T19:11Z 23.1K followers, 61.4K engagements


"Just turned off our last X EC2 instances -- Moondream Cloud inference is now running XXX% on FAL"  
[X Link](https://x.com/vikhyatk/status/1977927487863836776) [@vikhyatk](/creator/x/vikhyatk) 2025-10-14T02:40Z 23.1K followers, 46.9K engagements


"can someone who knows biology please comment on this is the discovery good novel or cringe novel"  
[X Link](https://x.com/vikhyatk/status/1978596116695544245) [@vikhyatk](/creator/x/vikhyatk) 2025-10-15T22:57Z 23.1K followers, 108.9K engagements


"Moondream Cloud is here. Thank you for your attention to this matter"  
[X Link](https://x.com/vikhyatk/status/1979222969542152567) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T16:28Z 23.1K followers, 51.2K engagements


"By the way we updated the Moondream X license last week to allow pretty much any use that doesn't compete with our paid offerings. The original version was more restrictive than intended that's what i get for vibe-generating licenses with ChatGPT"  
[X Link](https://x.com/vikhyatk/status/1979257741152784500) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T18:46Z 23.1K followers, 6790 engagements


""Some people asked me why I didn't use Gemini. I tried to use it but the detection wasn't that good. Moondream was the best at this.""  
[X Link](https://x.com/vikhyatk/status/1979259359319454081) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T18:52Z 23.1K followers, 4918 engagements


"how is this even possible i don't have XXX GB of memory here"  
[X Link](https://x.com/vikhyatk/status/1979287555188363502) [@vikhyatk](/creator/x/vikhyatk) 2025-10-17T20:45Z 23.1K followers, 5772 engagements


"if you spend $4.2M on ablations and don't deliver a significant improvement upon state of the art the rest of your assets should be seized and put to better use"  
[X Link](https://x.com/vikhyatk/status/1979775715836101014) [@vikhyatk](/creator/x/vikhyatk) 2025-10-19T05:04Z 23.1K followers, 30.6K engagements


"@____dzc____ they're gonna add more it was showing just ec2 and ddb a few mins ago. probably something foundational"  
[X Link](https://x.com/vikhyatk/status/1980178367329816871) [@vikhyatk](/creator/x/vikhyatk) 2025-10-20T07:44Z 23.1K followers, XXX engagements

[GUEST ACCESS MODE: Data is scrambled or limited to provide examples. Make requests using your API key to unlock full data. Check https://lunarcrush.ai/auth for authentication information.]

@vikhyatk "are you kidding me i can't put tr inside table now fuck nextjs fuck technology fuck you all i'm done"
X Link @vikhyatk 2025-10-17T04:56Z 23.1K followers, 482.5K engagements

"you don't build in "stealth mode" because you're worried about someone stealing your idea you do it because "announcing your raise / coming out of stealth" is a press event that generates a lot of eyeballs you want to have a product to direct those eyeballs to"
X Link @vikhyatk 2025-10-19T20:11Z 23.1K followers, 18.1K engagements

"framework XX review XX hours in: - very easy to assemble everything snaps into place with magnets - doesn't fit in my laptop bag if i could redo i'd get a smaller one - should've gotten a USB-A expansion card for installing the OS - feeling very at home with native tiling wm"
X Link @vikhyatk 2025-10-08T17:37Z 23K followers, 6476 engagements

"trying to benchmark moondream vs gemini XXX flash on cost struggling because gemini refuses to transcribe every document I've tried so far 😔"
X Link @vikhyatk 2025-10-16T19:14Z 23K followers, 7765 engagements

"@rajankrx well i actually do know how they did it they built a cutting edge inference-focused compiler i'll let @matthewjgunton chime in with more details"
X Link @vikhyatk 2025-10-17T01:50Z 23K followers, XXX engagements

"i'm on arch linux why are you looking up mac os docs"
X Link @vikhyatk 2025-10-17T06:52Z 23K followers, 1626 engagements

"why did you get 96GB of RAM what are you going to do with it idk maybe ill run two electron apps at the same time"
X Link @vikhyatk 2025-10-08T05:22Z 23.1K followers, 123K engagements

"losing $20b a year with $13b in revenue stop looking at spreadsheets. go try out codex. if anything theyre undervalued"
X Link @vikhyatk 2025-10-16T07:58Z 23.1K followers, 195.9K engagements

"@HeadBetwixKnees yeah the ui is a bit confusing but it has a lot of features"
X Link @vikhyatk 2025-10-19T23:32Z 23.1K followers, XX engagements

"Just shipped a 4-bit quantized version of Moondream. 🌘 XX% less memory. 🌓 XX% faster inference. 🌔 XXXX% accuracy"
X Link @vikhyatk 2025-05-22T00:12Z 23K followers, 51.8K engagements

"how do i stop being underrated this feels like a failure. i want to be overrated"
X Link @vikhyatk 2025-09-24T19:38Z 23K followers, 44.1K engagements

"the thing i love about codex is i can have it spend X hours do a complex refactor look at the resulting code and just say yeah hmm not feeling the data model here let's redo it a completely different way with a human dev you just have to merge it otherwise they'll get upset"
X Link @vikhyatk 2025-10-05T19:57Z 23K followers, 224.9K engagements

"made a cute logo for our internal inference engine"
X Link @vikhyatk 2025-10-15T03:48Z 23K followers, 4204 engagements

"I still dont know what Salesforce really does"
X Link @vikhyatk 2025-10-15T16:35Z 23K followers, 13.1K engagements

"cant believe how good clickhouse is"
X Link @vikhyatk 2025-10-16T16:36Z 23K followers, 1916 engagements

"@init_malachi my standard pattern used to by dynamodb master and elasticsearch for analytics. clickhouse is a better elasticsearch replacement. but dont make it master"
X Link @vikhyatk 2025-10-16T22:01Z 23K followers, XX engagements

"when they find out you're still using claude sonnet instead of grok code"
X Link @vikhyatk 2025-10-17T05:21Z 23K followers, 7128 engagements

"turns out this is not a good way to run a business"
X Link @vikhyatk 2025-10-17T18:17Z 23K followers, 11.6K engagements

"compared to the "Gated Attention for Large Language Models" paper X. we have G4 and G2 but not G1 X. we allow gating to be conditioned on log(pos) (i actually think it should be sqrt(log(pos)) but that's a topic for a different post). this helps with long ctx"
X Link @vikhyatk 2025-09-20T19:05Z 23K followers, 37.5K engagements

"i think G2/G4 help with exploding attention logits whereas G1 helps with attention sparsity/sinks. i also see a parallel between G1 and the gpt-oss attention sinks except S is now q-dependent instead of static"
X Link @vikhyatk 2025-09-20T19:05Z 23K followers, 1733 engagements

"cloud providers shouldn't have good ux their margins should be so thin they can't afford designers"
X Link @vikhyatk 2025-10-15T01:32Z 23K followers, 4744 engagements

"LPIPS (typically) uses frozen AlexNet features with a learned linear head. Trained on 484k MTurk samples asking people which of two distortions was closer to a sampled 64x64 patch. Used as a component of the training loss. It's surprising we're still using this 2017-era proxy"
X Link @vikhyatk 2025-10-15T01:47Z 23K followers, 4937 engagements

"Excited to release a preview of Moondream X. A 9B param 2B active MoE vision language model that makes no compromises; offering state-of-the-art visual reasoning while still retaining an efficient and deployment-friendly form factor"
X Link @vikhyatk 2025-09-18T22:11Z 23.1K followers, 225K engagements

"having worked on a lot of "enterprise grade" code in my career all i'll say is you shouldn't treat it as a compliment"
X Link @vikhyatk 2025-10-19T04:59Z 23.1K followers, 6339 engagements

"@ludwigABAP when i turn XX i will be posting on subs*ack not here"
X Link @vikhyatk 2025-10-19T06:18Z 23.1K followers, XXX engagements

"our login broke. went to sst to check why theirs was also broken. turns out us-east-1 is down"
X Link @vikhyatk 2025-10-20T07:19Z 23.1K followers, 13K engagements

"Overtokenized transformers "Through extensive experiments we uncover a log-linear relationship between input vocabulary size and training loss demonstrating that larger input vocabularies consistently enhance model performance regardless of model size.""
X Link @vikhyatk 2025-10-13T19:11Z 23.1K followers, 61.4K engagements

"Just turned off our last X EC2 instances -- Moondream Cloud inference is now running XXX% on FAL"
X Link @vikhyatk 2025-10-14T02:40Z 23.1K followers, 46.9K engagements

"can someone who knows biology please comment on this is the discovery good novel or cringe novel"
X Link @vikhyatk 2025-10-15T22:57Z 23.1K followers, 108.9K engagements

"Moondream Cloud is here. Thank you for your attention to this matter"
X Link @vikhyatk 2025-10-17T16:28Z 23.1K followers, 51.2K engagements

"By the way we updated the Moondream X license last week to allow pretty much any use that doesn't compete with our paid offerings. The original version was more restrictive than intended that's what i get for vibe-generating licenses with ChatGPT"
X Link @vikhyatk 2025-10-17T18:46Z 23.1K followers, 6790 engagements

""Some people asked me why I didn't use Gemini. I tried to use it but the detection wasn't that good. Moondream was the best at this.""
X Link @vikhyatk 2025-10-17T18:52Z 23.1K followers, 4918 engagements

"how is this even possible i don't have XXX GB of memory here"
X Link @vikhyatk 2025-10-17T20:45Z 23.1K followers, 5772 engagements

"if you spend $4.2M on ablations and don't deliver a significant improvement upon state of the art the rest of your assets should be seized and put to better use"
X Link @vikhyatk 2025-10-19T05:04Z 23.1K followers, 30.6K engagements

"@dzc they're gonna add more it was showing just ec2 and ddb a few mins ago. probably something foundational"
X Link @vikhyatk 2025-10-20T07:44Z 23.1K followers, XXX engagements

creator/twitter::17406365/posts
/creator/twitter::17406365/posts