Iamaev Andrei

| LLM | coding | anime | traveling | PhD at machine learning | Game dev | Just want to chat with whole world without limits |

[Iamaev Andrei @rewin123]

@ItsmeAjayKV @UnslothAI Tested this on 2 x RTX 3090 on Xeon E5 267X. Got same ~11 tok/s. My CPU bandwidth and old DDR4 limit offload very hard. 😅

[Iamaev Andrei @rewin123]

Found today, that LLM Laguna S2.1 Dflash drafter model has same structure as DSpark drafter, but just without Markov head. I trained Markov head from scratch and only got +5% accept rate and around 1 tok/s speed up compare to original 🫠 So I found why Markov head was disabled.

[Iamaev Andrei @rewin123]

@ItsmeAjayKV 64 gb dd4 ram (quad channel) and 24 gb vram

[Iamaev Andrei @rewin123]

@ItsmeAjayKV @poolsideai 57 tokens prefil it's ouch

[Iamaev Andrei @rewin123]

This is what I want to do on Monday instead of work.

[Iamaev Andrei @rewin123]

This is the Sanxingdui Museum in China. The masks are up to 3,000 years old and belong to an extinct culture discovered only in the 1900s.

[Iamaev Andrei @rewin123]

-- Ancient china, why --🫠

[Iamaev Andrei @rewin123]

@AbhiCodes15 Sometimes two, may be 3. But only if second and third agents do very long task, so I can move away from controlling them. For example right now one agent do coding right now, another creating hermes profile for my wife

[Iamaev Andrei @rewin123]

Just tested the new Claude Science (Anthropic's AI research workbench) on writing a ML paper.

⚖️ Verdict: it genuinely speeds things up, with great built-in tooling for generating papers — but it's not without rough edges.

🧪What I did: measured how well two attention layers combine — Gated DeltaNet-2 (linear attention, NVIDIA) and CSA (compressed sparse attention, DeepSeek) — and wrote it up.

📊 Finding: CSA + Gated DeltaNet-2 matches pure-CSA accuracy with fewer ops — up to 1.7× prefill wallclock speedup (from the sparser combined attention, not from the tool). Paper (methods + benchmarks): https://t.co/pLEuCerxzy

What worked well:
✅ Ready-made skills for plots and LaTeX — output looks clean out of the box, minimal cleanup
✅ Full versioning of every result and every pdf/tex/any file — nothing gets lost
✅ Smooth SSH remote compute offloading (better than Claude Code here)
✅ Strong autoreview that actually catches bugs and spelling issues
✅ Familiar Claude interface with nice artifact browsing
What didn't:
❌ Pure web-chat workflow — no file tree, no easy file referencing. Without a local project folder you're working a bit blind, and iterative ML coding feels clunkier than a VS Code + Git setup
❌ Main focus is life sciences; using it for ML is off-label

So, ofc Claude Science main usage path is genetic and biology, but why no to try in ML. So for now for me, If I actually needed to ship an ML paper, I'd still lean toward VS Code + Git + Claude Code. The web interface's upsides aren't decisive, and the missing file tree makes you work a bit blindly. May be I will use Claude Science for final paper polishing.

Hopefully this small review will help someone . Share your experience.

[Iamaev Andrei @rewin123]

@Im_IrushiK My AI don't want think 🤔