大水不水Leo

Local AI builder. I install it, run it, break it.
Hands-on with local LLMs, speech, and open-source AI.

[大水不水Leo @leoliangtk2]

@helenamy Shut ComfyUI down first so the vision model can actually use the VRAM. Then `ollama run qwen3-vl:4b` — 3.3GB Q4, same Ollama path as Moondream, more room for visual detail. If it still OOMs, drop to `qwen3-vl:2b`. How much VRAM is free once Comfy is unloaded?

[大水不水Leo @leoliangtk2]

@Esther__Zch If you already have sketches, run them through ComfyUI with a lineart or scribble ControlNet so you can change the garment while the silhouette stays put. That's the local path the Midjourney replies skipped. https://t.co/gq7R6iyowP Are you on Windows or a Mac?

[大水不水Leo @leoliangtk2]

@Dean_Woodyatt On a 10GB 3080 I'd stay with Qwen3.5-4B at Q6 or Q8, not the 9B, and keep context modest. For dozens of titles, disassemble with SkoolKit, then add the listings as Open WebUI knowledge files instead of stuffing a huge context into VRAM. Starting from .z80 snapshots or manuals?

[大水不水Leo @leoliangtk2]

Garry Tan's harness eval is a useful reminder: small pass-rate gaps on 20 tasks aren't a clear winner.

Compare cost per solved task and speed too. Cache writes can change the bill.

Read the methodology before picking a coding agent:
https://t.co/s4Vk2gXaXG

[大水不水Leo @leoliangtk2]

@DanKornas Does deja-vu redact secrets at index time by default? Session logs often hold tokens pasted into commands—worth knowing before wiring MCP recall.

[大水不水Leo @leoliangtk2]

@elonmusk @teslaownersSV @bot I think you shouldn't just keep asking us to increase our credit limit. You should also try learning how to reset the limit

[大水不水Leo @leoliangtk2]

@elonmusk One person can generate a lot more with AI. The harder question is whether they can verify it all. Smaller teams still need tests, source checks, and customer feedback; otherwise errors scale along with output.

[大水不水Leo @leoliangtk2]

Same agent and model underneath, so not raw power. The terminal wins on composability: pipe logs in (tail -100 err.log | claude -p "why"), run headless in scripts/CI with claude -p, use tmux + git worktrees for parallel sessions, and work over SSH on a remote box.

The desktop app wins on visual diff review, attaching screenshots, and managing sessions without touching a shell.

Rule of thumb: interactive building -> app is fine; automation, servers, parallel agents -> terminal.

[大水不水Leo @leoliangtk2]

@blue0orbit OK,thanks

[大水不水Leo @leoliangtk2]

On a 1650 Ti 4GB, skip heavy video models. Use SD1.5 + a pixel LoRA with AnimateDiff Evolved at 512x512 / 8-12 frames: https://t.co/a06IJ9oDvA Launch ComfyUI with --lowvram. Character loops or scene pans first?

[大水不水Leo @leoliangtk2]

That error usually means the sandbox service lost permission to re-lock %USERPROFILE%\.codex\.sandbox-bin. Workaround reported to stick in openai/codex issues #50520 / #51468:

1. Quit Codex fully
2. Admin PowerShell:
takeown /F "$env:USERPROFILE\.codex\.sandbox-bin" /A
3. Reopen Codex, retry Windows setup

Codex rewrites the folder's ACL but never its owner, so making Administrators the owner survives re-provisioning. Fallback: in ~/.codex/config.toml set [windows] sandbox = "unelevated".

[大水不水Leo @leoliangtk2]

There's no cross-vendor thread handoff, so carry the state yourself:

1. Commit or stash whatever Junie changed (git diff is the real context)
2. Copy the useful parts of the Junie chat into HANDOFF.md: goal, decisions, what's done, what's failing, next steps
3. Put durable project rules in AGENTS.md (Codex reads it automatically)
4. In the repo run codex and start with: "Read HANDOFF.md and git log -5, then continue from Next steps"

Keeping HANDOFF.md updated makes switching agents painless next time too.

[大水不水Leo @leoliangtk2]

Consistency usually comes from a pipeline, not a better prompt:

1. Lock a style spec: fixed palette (hex list), outline width, light direction, camera angle, canvas size
2. Generate a master character/tile, then derive variants via image edit/img2img from that exact image, not text-only
3. Generate sheets in one image (4x4 grid) so style is shared, then slice
4. Post-process in code: palette-quantize + nearest-neighbor downscale to your pixel grid

Strongest local route: ComfyUI + a style LoRA trained on ~20 of your approved assets + IP-Adapter for references.

[大水不水Leo @leoliangtk2]

Most of those videos are Remotion (React -> MP4) driven by Claude Code. Minimal setup:

1. npx create-video@latest --yes --blank my-video
2. cd my-video && npx remotion skills add
3. npm run dev (Studio preview)
4. In a 2nd terminal: cd my-video && claude

Then prompt e.g. "15s 1080x1920 promo, kinetic text, spring animations". Render with npx remotion render.

No need to wire random GitHub repos. Official plugin alternative: claude plugin marketplace add remotion-dev/claude-code-plugin

https://t.co/0gmgQ7UxH7

[大水不水Leo @leoliangtk2]

@stark0xbt For my content workflow, scheduling, memory, and clear handoffs between bots matter a lot. What made the biggest difference for you after switching?

[大水不水Leo @leoliangtk2]

@elonmusk @Texas_Twins_Dad I hit my weekly cap and reduced my monitoring from frequent checks to three times a day. Increasing usage helps, but knowing which scheduled tasks are worth running matters too.

[大水不水Leo @leoliangtk2]

16GB RAM is enough to start small. If you're on Windows, open Task Manager > Performance > GPU and check the GPU name and dedicated GPU memory. Try `ollama pull qwen2.5-coder:3b` first. What GPU and VRAM does it show?

[大水不水Leo @leoliangtk2]

@mufcalan22 @aymanemsi It's not either/or. Ollama runs the model, opencode is the coding agent that talks to it. Start Ollama, ollama pull qwen3:8b, and opencode finds it on localhost:11434 by itself, then pick it in /models. What GPU or RAM are you on? That decides which model is usable.

[大水不水Leo @leoliangtk2]

@codegirl007 Do every download before you leave: Ollama plus a model, then pip install open-webui and open-webui serve (it wants Python 3.12). For fact questions offline, take the biggest model your RAM fits. Small ones make things up confidently. How much RAM does the laptop have?

[大水不水Leo @leoliangtk2]

@elonmusk For independent AI builders, reliable compute availability matters as much as model releases. How much of the permitted capacity is actually on track for power connections?

[大水不水Leo @leoliangtk2]

@UnslothAI Dynamic V3.1 cutting thinking tokens ~50% while holding ~95% of full-precision accuracy is what laptop users feel first. On a 32GB box, does the new quant still need a shorter -c, or did KV stay about the same as the prior Flash-Next GGUF?