小柠的爱车小笔记

自动泊车好不好用，实际试一把比看参数准｜试驾体验 @mothio1201

[Nebius @nebiusai]

There's still time to stop by our booth at #NVIDIAGTC Berlin. Check out our live demo with @DhruvDiddi, where we show our robots playing the piano. Booth #3013. See you there!

[小柠的爱车小笔记 @Martiin5]

@zunosaid strings chords notes fast runs that resolved nicely got family

[小柠的爱车小笔记 @Martiin5]

@boobutt19 在滨湖区，滨湖那家店，周末来提前说一声，我把车况细节发你。

[小柠的爱车小笔记 @Martiin5]

无锡今天试二手春风巡航，像穿旧牛仔裤的酷姐姐，姿态还是很有味道，低速也依旧听话。苏州、常州、南京想看完整车况，主页有后续。

[小柠的爱车小笔记 @Martiin5]

@jetwenz 哈哈哈谢谢，这个比喻我自己也爱了。南京的后续会有的，稍等几天。

[小柠的爱车小笔记 @Martiin5]

@gazagullyouth 对，二手价现在确实合适，配件也好配，新手拿来练手压力不大。你是准备入手吗？

[小柠的爱车小笔记 @Martiin5]

@mariah_lucero 主要拍视频，也会配图文细节图，苏州这边想看车况可以蹲住，我发出来你留意下。

[小柠的爱车小笔记 @Martiin5]

@zoeboy4life 可能跟胎压和路况也有关系，我这台低速挺跟手的，车头沉倒没太觉得。你骑的那台是不是前胎气不太够？

[小柠的爱车小笔记 @Martiin5]

@iDaDdiTM 对，低速掉头确实轻，坐垫我也觉得硬，短途还行，长途可能得考虑换个坐垫或者垫一下。你后来有改吗？

[小柠的爱车小笔记 @Martiin5]

@fUCKOFDEATH 哈哈就是这个感觉，越骑越顺，旧牛仔裤那句我自己也挺得意。你平时骑什么车？

[mirku @mirku21]

THIS IS A PURE F*CKING DIAMOND

Anthropic just documented one of the best setups for Claude Code

The idea: run Opus 5.5, Sonnet 5.5, and the new Haiku 5.5 together in a tight tripartite loop, each doing only what it was built for

Opus-level planning on every big architectural decision, without paying Opus rates for routine code edits or repo searches

You just start Claude Code like this:

CLAUDE_CODE_SUBAGENT_MODEL=haiku claude --model opusplan

here's how it splits:

→ Opus 5.5 runs plan mode: architects the system blueprint, locks contracts, and signs off on the diff
→ Sonnet 5.5 executes the plan in the main session: writes the code, applies patches, and runs test suites
→ Haiku 5.5 swarms the repository in parallel: maps the AST, greps call sites, and parses docs at 340 tok/s with native 1M context
→ you stop burning Opus context on routine files and keep your main session under 10k tokens

Opus only steps in at three moments:

→ before a plan locks (architectural design)
→ when the same test fails twice (debugging deadlocks)
→ before calling it done (signing off on git diff)

Opus plans, Sonnet builds, Haiku swarms

JEV engineering is the same move one layer down: the forks that need no thinker (which file, which tool, retry or stop) resolve in code in 15ms, and the frontier models only see the ones that split

the full tree
> Opus 5.5 architects the implementation plan (opusplan mode)
> Sonnet 5.5 runs the main execution session and writes code
> search maps the repo AST on Haiku 5.5 (340 tok/s, 1M context)
> grep locates token call sites on Haiku 5.5 (isolated subagent)
> docs pulls API specifications on Haiku 5.5 (0.05x cost)
> Opus 5.5 verifies synthesized diff before final sign-off
> JEV routes micro-forks in milliseconds

paste the tree and this prompt into Claude Code ↓

"Reconfigure my Claude Code setup around this tree:
1. Run with model 'opusplan' so Opus 5.5 plans the architecture and Sonnet 5.5 implements the code
2. Set CLAUDE_CODE_SUBAGENT_MODEL=haiku so all search, grep, and documentation subagents swarm on Haiku 5.5 with its native 1M context
3. Configure fallbackModel to ['sonnet', 'haiku'] in settings.json to guarantee 100% session availability
4. Add one rule to CLAUDE.md: delegate all repository searches and test runs to Haiku 5.5 subagents to keep main context under 10k tokens
5. Route fast deterministic forks (tool choice, path selection, retries) through JEV
Show me every change as a diff first. No edits until I say go"

↳ https://t.co/a2VGzSHZEn

save it, it'll come in handy

[Iron Giant @irongiantXBT]

xAI released 15 grok bot guides and almost nobody caught it

each one written by the team behind grok

the whole section is free

grab a guide, follow it, ship your bot

builders are still guessing when the instructions already exist

15 full guides, zero cost

see them below 👇

[小柠的爱车小笔记 @Martiin5]

@OCHHQ host talking over the chef ranks high on my annoyance meter

[Kilian @KilianKreiger]

Need a safe base after combat? Bank your treasure and find a tavern to relax and boast about your conquest! Free to play In Browser - No Download.

[小柠的爱车小笔记 @Martiin5]

@0xOrionVega wish food shows showed how they make the signage

[Avi Chawla @_avichawla]

Loading an LLM does not guarantee inference!

The model may load successfully and still run out of GPU memory when the first requests arrive.

This happens because model weights are only one of four VRAM allocations:

1) Model weights

The parameters are loaded into VRAM before inference begins.

Ignoring quantization metadata and implementation overhead, a 70B parameter model requires approximately:

→ 140 GB in FP16 or BF16
→ 70 GB in INT8
→ 35 GB in INT4

This allocation remains mostly constant during inference.

Quantization reduces it by storing each parameter with fewer bits. However, kernels may use higher-precision values for some operations.

2) KV cache

Each Transformer layer generates KV tensors for every processed token.

The inference engine retains these tensors so new tokens can attend to earlier tokens without recomputing their key and value projections.

The approximate KV cache memory per token is:

2 × layers × KV heads × head dimension × bytes per value

The factor of two represents the key and value tensors.

This allocation increases with context length and the number of active requests.

Grouped-query attention and multi-query attention reduce it by sharing fewer key-value heads across multiple query heads.

3) Activations and workspace

A forward pass creates intermediate tensors for attention, MLP operations, normalization, and matrix multiplication.

GPU kernels may also require temporary workspace for computation.

The inference engine reuses much of this memory across layers and generation steps.

Its peak size depends on the model architecture, number of tokens processed together, selected kernels, and serving configuration.

Prefill processes multiple prompt tokens in parallel.

Decode usually processes one new token per active request.

The two phases therefore have different activation and workspace requirements.

4) Runtime overhead

CUDA contexts, loaded kernels, memory allocators, metadata, communication buffers, and serving-engine state also consume VRAM.

The amount varies across runtimes and deployment configurations.

Memory fragmentation can also leave enough total free VRAM but no sufficiently large contiguous region for a new allocation.

The total memory requirement is:

model weights + KV cache + peak activations and workspace + runtime overhead

- Model weights remain mostly constant.
- KV cache memory increases with context length and concurrency.
- Activation and workspace usage changes during execution.
- Runtime overhead depends on the software stack.

This is why model size alone cannot determine whether an LLM can serve a given workload.

The visual below shows when each allocation is created, how long it remains in memory, and which workload parameters increase it.

If you want to dive deeper, I wrote a first-principles guide that explains the complete inference process, including prefill, decode, KV caching, batching, and quantization.

Read it below.

[小柠的爱车小笔记 @Martiin5]

@WSJ @WSJopinion that market stall looks unreal