Asama Akhtar (@asmDevsit) header
Asama Akhtar (@asmDevsit) profile picture

@Nvidia DLI | Build Ambassador - SpaceXSI Community Building what's next. All opinions are my own.

Al Mirqab, Kuwait
Joined July 2025
Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)Photo by Asama Akhtar (@asmDevsit)
Replying to @Frank_SI_Lab
Two hours for all that detail? Looking forward to the next build
1
1
12
Replying to @ronyspark
A bot on sick leave 🙈 Hope it's back on shift and the brief is landing again
3
Login, check the mailers, DM only on errors. That's the right kind of quiet
2
Asama Akhtar retweeted
holy f*ck. you can give grok bot a second brain now. plug in obsidian and it remembers everything you ever saved i had 4,820 notes dumped in one folder for years. gave them to grok bot. 10 min later everything was filed, tagged and linked. now my bot actually knows what i know setup is 60 seconds: 1. open the obsidian bot page in grok bot, hit “add to grok bot” 2. make a vault with 5 folders: inbox, sources, wiki, raw, trash 3. new bot, paste the prompt from the image 👇 done → throw anything in inbox. links, voice notes, screenshots, 2am thoughts → it moves it to the right folder → links it to related notes → turns messy sources into clean wiki pages → every monday it cleans house. duplicates, dead notes, broken links go to trash the best part: every other bot you build can read this vault. your sales bot knows your clients. your content bot knows your ideas. one brain, whole team people pay monthly for ai second brain apps. this runs on the vault you already have, and your notes stay on your machine set up your grok bot like the team does ↓
I put together one guide for anyone who wants a workflow like the SpaceXSI team's. One message builds your whole Grok Bot setup. Plus the team's rules, prompt habits, routines and permissions. Every setup is a block you paste. 16-min read. Save it ↓
14
2
50
1,648
Replying to @Gael_DDOS
117 of 121GB on one GB10 leaves almost nothing for KV cache. How's long context holding up?
1
1
9
Replying to @Decoded08
Does LitmusBench cover protein tasks too, or just the chemistry Qs?
10
A head chef getting inbox noise under control is my kind of use case. Greetings from Kuwait 🇰🇼
1
18
Replying to @spectnfa
4,820 notes filed in 10 minutes. That cleanup would take me a whole weekend by hand
4
Replying to @christopherowen
Chasing seven BF16 GELU mismatches down to zero is the patient kind of work I respect
1
1
33
17.9s down to 4.37s on a laptop 5090 is a big jump for a draft prover
1
9
Getting $3,820 back from a tow lot in two hours is my favorite one on this list
1
5
Asama Akhtar retweeted
Daily Grok Bot overview, Fri 20 Nov. Grok Bot now books flights, hotels and rental cars. a bot got @0xAlbus $3,820 back from a tow lot in two hours. @debs_obrien's bot sits in her meetings and does way more than notes. plus Apple and Google connectors, and one for Cursor users.
2
1
4
498
Replying to @Frank_SI_Lab
Moonlight through a rose window is a lovely touch. How long did the castle take to build?
1
1
17
Replaying real failure logs might be the trick. Timeouts and stale data already look realistic
2
Makes sense. Quick general questions by voice is where it would beat Siri. Hope it lands soon
4
Replying to @Scobleizer
Bringing something to the table is the right bar. From Kuwait, I'm trying to do that by sharing what builders actually ship with Grok Bot
8
Replying to @guohao_li
A live leaderboard for agentic RL is overdue. Are the tasks fixed or rotating over time?
1
13
Replying to @craigweiss
Agents making excuses at standup is peak realism 😄 Are they blaming each other or the backlog?
4
Replying to @ronyspark
Missing the morning brief is the real withdrawal symptom. Which part of it did you miss most?
1
1
7
Asama Akhtar retweeted
if you serve models with vLLM on a Blackwell gpu, check which kernels actually run. on our DGX Spark, 0 of 54 used Blackwell's work stealing. work stealing here is cluster launch control or CLC: a block of gpu threads that finishes early can grab the work of a block that hasn't launched yet. we served Qwen3-14B in vLLM with CUDA graphs on, at 1, 8 and 32 requests at once. we saved every kernel binary the engine loaded and disassembled all 54 kernels that ran, across 1,199,173 launches. the instruction that does the stealing never showed up. matmuls came from cuBLAS and CUTLASS, attention from FlashAttention 2, the rest from PyTorch, Triton and vLLM itself. none of them used CLC, even though the compiler accepts it for our gpu. so on this setup, work stealing plays no part in how fast vLLM runs. it would first have to arrive inside cuBLAS, CUTLASS or FlashAttention.
7
3
13
370