Andrei 𒅴𒂠

chief goon officer @anara

[Andrei 𒅴𒂠 @andreidro]

Reality and love are almost contradictory for me. It's funny, every one of my exs are now married. Men go out with me, we break up and then they get married.

Later they call me to thank me for teaching them what love is, that I taught them to care and respect women.
- Celine

[Andrei 𒅴𒂠 @andreidro]

@PiticVic drop

[Andrei 𒅴𒂠 @andreidro]

trying to get --enable-dp-attention too, was crashing with deepseek and EAGLE

[Andrei 𒅴𒂠 @andreidro]

started self-hosting models, my Arrakis, my Dune

deepseek-v4-flash on modal 8xB200 180 TPS

[Andrei 𒅴𒂠 @andreidro]

using sglang

python -m sglang.launch_server
--host 0.0.0.0
--port 8000
--model-path deepseek-ai/DeepSeek-V4-Flash
--tp-size 4
--dp-size 2
--config /root/config.yaml
--enable-mixed-chunk

yaml config (config_deepseek_v4.yaml):
host: 0.0.0.0
log-level: info

tool-call-parser: deepseekv4
reasoning-parser: deepseek-v4
trust-remote-code: true

# Memory
mem-fraction-static: 0.92
context-length: 1048576 # 1M
chunked-prefill-size: 8192

# MoE
moe-runner-backend: flashinfer_mxfp4

# Observability
enable-metrics: true
enable-cache-report: true

# Batching — 8x B200
max-running-requests: 64
cuda-graph-max-bs: 128

# EAGLE 3/1/4 speculative decoding
speculative-algorithm: EAGLE
speculative-num-steps: 3
speculative-eagle-topk: 1
speculative-num-draft-tokens: 4

# Scheduling
schedule-policy: lpm
num-continuous-decode-steps: 4

# Kernel fusions
enable-mixed-chunk: true
enable-fused-qk-norm-rope: true
disable-flashinfer-autotune: true

env vars (in modal container):
SGLANG_ENABLE_SPEC_V2=1
SGLANG_ENABLE_THINKING=1
SGLANG_JIT_DEEPGEMM_PRECOMPILE=0
HF_HOME=/cache/huggingface
HF_XET_HIGH_PERFORMANCE=1
HF_HUB_ENABLE_HF_TRANSFER=1

[Andrei 𒅴𒂠 @andreidro]

benchmaxxed

[Logan Kilpatrick @OfficialLoganK]

Gemini 3.5 Flash has made huge progress from 3.1 Pro on GDPval, Flash is competing at the frontier, post training going strong :)

[Andrei 𒅴𒂠 @andreidro]

3.1 pro is utterly horrible for any agentic use case and everyone knows it

[News from Google @NewsFromGoogle]

New stat from @vercel's AI Gateway in @BusinessInsider: Gemini 3 Flash is leading across AI models in token usage as of April. 🚀 See more stats on how developers are using our models → https://t.co/0xbulg4X2S

📊via @BusinessInsider

[Andrei 𒅴𒂠 @andreidro]

@sebisemeniuc top

[Andrei 𒅴𒂠 @andreidro]

anara mogging - last 7 days bro

[dax @thdxr]

our teams last 7 days of spend
damn gpt5.5

[Andrei 𒅴𒂠 @andreidro]

@sebisemeniuc what models

[Octavian Cristea @octavicristea]

I made a game where Sam Altman calls your pitch "pre-seed energy" and Peter Thiel offers you a term sheet

AI rewrites the whole story every play, pulled from real SF news and stories

built in couple of days for #ElevenHacks @elevenlabs and @zeddotdev hack #6

[Andrei 𒅴𒂠 @andreidro]

@BogdanGruiaIvan we goon now

[Sebastian Semeniuc @sebisemeniuc]

I really liked wispr flow but didn’t feel like paying $12/month for it, so i built a local-only open source version

it’s a native macos dictation app that runs whisper fully on-device and types directly into any app

if you use dictation, this is a simple alternative that just works, fully local

you can download the dmg and use it right away

[Andrei 𒅴𒂠 @andreidro]

new tattoo frfr

[Andrei 𒅴𒂠 @andreidro]

@octavicristea i like the colour in last one, but is hard to read, would try a mesh gradient over the image on the text side

[Andrei 𒅴𒂠 @andreidro]

@octavicristea top

[Octavian Cristea @octavicristea]

i built a tool that lets AI agents explore the web like a human would

give it a URL and watch it go. it scrapes the page, previews every link, picks where to go next, and the whole thing plays out as a live graph

open source. here's what it looks like:

[Andrei 𒅴𒂠 @andreidro]

@octavicristea clawing on phone fr

[Andrei 𒅴𒂠 @andreidro]

Claude vs OpenAI

[Andrei 𒅴𒂠 @andreidro]

please uwu

https://t.co/DQ6GQl3Ede

[Andrei 𒅴𒂠 @andreidro]

an idea for sending custom approval data in AI SDK