早八幸存者

This is f*cking gold, and it's a 100-year-old exploit 🔥 In 1926 a power company patched human fear with a cartoon, and in 2026 every AI lab is running the same patch on your inbox. The bug in the 1920s: electricity meant fires and shocks. Families ran one bulb in the evening. So Alabama Power shipped Reddy Kilowatt. Lightning-bolt body, bulb nose, outlet ears. Almost 150 US utilities licensed him, and he stayed the face of electricity for about half a century. 1926 FEAR → FACE → TRUST → PLUG IN 2026 FACE → ASK ACCESS → SIGN IN → ACT Where: -> Access = inbox, calendar, card, files -> Act = send, book, buy, delete Why it still compiles: Konrad Lorenz, 1943. Big head, big eyes, round cheeks read as "infant", and the brain dials down the threat response. A bug in human firmware that nobody can patch 🧠 Watch 0:03, the second Reddy turns from grey to red. That's the face booting up. The face says nothing about permissions. Read the scopes, not the smile. A face people trust in one frame is the same exploit that pulled Jean Phil 13.8M views 👇

早八幸存者 @DIETRICH11

6 days ago

@RAW_IS_JERICHO Moto Guzzi在張家界山路跑，軸傳動的後段應該很穩，轉彎會不會覺得車身有點沉？

早八幸存者 @DIETRICH11

7 days ago

@AnnatarXBT the romance cold but predictable plus black type side speaks truth somehow scene fully mostly no flaw

DIETRICH11 retweeted

Sprytix

@Sprytixl

9 days ago

THIS IS ACTUAL F*CKING GOLD Anthropic just shared one of the best tricks for Claude Code. The idea: stop treating Claude like one worker. Make Opus 5.5 a commander and give it 12 agents in 4 squads. Same release, two ways: Opus alone was 2% through on one thread. The command center shipped the whole v2.0 in 2 minutes on 12 threads. You just launch Claude Code like this: claude --model opus --command-center 12 "ship v2.0" here's the split: → Opus 5.5 turns "ship v2.0" into 12 tasks across 4 squads → CODE and OPS squads run on Sonnet 5.5, TEST and DOCS on Haiku 5.5 → every agent reports back in one line, never raw logs → Opus only wakes up when something breaks or needs approval the part most people get wrong: the commander watches everything. read 12 agent transcripts and your expensive model becomes the slowest part of the room. here the commander stays at 8k of context the entire run. it plans, it waits, it reads 12 one-line reports, it opens the release gate. REA sits underneath on anything legacy: before a squad touches old code, REA reverse-engineers it into a spec, so agents build from the spec and never from 500 files nobody understands. Opus commands. Sonnet builds. Haiku checks. REA explains the old code. JEV engineering is the same move one layer down: the forks that need no thinker (which squad, which model, retry or stop) go to JEV in milliseconds, and the commander only sees the ones that actually split the full tree > Opus 5.5 runs as commander, splits the mission into tasks and squads > CODE squad on Sonnet 5.5 writes endpoints, refactors auth, migrates the db > TEST squad on Haiku 5.5 runs unit, e2e and load tests > DOCS squad on Haiku 5.5 writes changelog, api docs and release notes OPS squad on Sonnet 5.5 handles staging, health checks and canary rollout REA reverse-engineers any legacy module into a spec before a squad touches it every agent reports one line, the commander never reads raw output Opus 5.5 holds the release gate and waits for all 12 reports JEV routes micro-forks in milliseconds paste the tree and this prompt into Claude Code ↓ "Rebuild my Claude Code setup around this tree: 1. Run the main session on Opus 5.5 as a commander: it plans, assigns and approves, it never writes code. 2. Split work into 4 squads: CODE and OPS on Sonnet 5.5, TEST and DOCS on Haiku 5.5, 12 agents total. 3. Every agent reports a one-line structured result. Raw logs never enter the main session. 4. Before any squad touches legacy code, run REA on it first and have the squad build from the spec, not the old files. 5. Add one rule to CLAUDE.md: nothing ships until every squad has reported and the commander opens the release gate. 6. Route deterministic forks (squad choice, model choice, retries) through JEV. Show me every change as a diff first. No edits until I say go." ↳ https://t.co/W8IMvUDehj save it, you'll use this on every release from now on.

454

789

133K

早八幸存者 @DIETRICH11

9 days ago

@KelleyBlueBook @INFINITIUSA (sigh) dark shot blocked so dramatically making door outline effectively watching standard what does screen sell with certain style in frames no, only strange wide.

DIETRICH11 retweeted

Daily Dose of Data Science

@DailyDoseOfDS_

10 days ago

Brilliant paper by NVIDIA. they found a way to make KV cache transferable between models. the target model skips prefill entirely, and the conversion runs 2.7 to 25x faster than processing the context again. let's understand why this is so important today. LLM APIs are stateless, so every turn sends the entire conversation back to the model. the model reads it again before generating a new token, and all of it is billed as input. prompt caching lets providers hold the KV cache for a stable prefix and bill cache hits at roughly 10% of the base input rate because that compute was already done. this 90% reduction is one of the biggest levers in LLM serving. but the cache only works on the model that produced it. since keys and values depend on that model's weights, another model cannot directly reuse them. this becomes a problem with LLM routing. if traffic moves to another model for cost or capability reasons, the accumulated KV cache becomes invalid. the entire context must be processed again and billed at full rate. NVIDIA's recent paper treats this as a representation problem. prefill's main output is the KV cache, so transferring it means converting one model's cache into the representation another model expects. they first checked whether this conversion had enough structure to exploit. for Qwen3 14B → 32B, a linear regression from a single source layer reconstructed 56% of the variance in the target model's keys. because models can have different layer counts, there is no natural one-to-one layer mapping. so, for every target layer, they rank the source layers by predictive power and use the top eight together. that pushes reconstruction to 79%. the mapper has three parts: > each target layer and head gets its own linear map, solved in closed form instead of with gradient descent. > cross-layer selection chooses the source layers that best predict each target layer. their ablation shows this contributes the most. > RoPE adds a position-dependent rotation to keys. they remove it, fit the mapping in position-free space, then apply the target model's rotation again at inference. across six model pairs from Qwen3, Llama 3.1, and Ministral 3, four retain 73 to 98% of the receiving model's standalone accuracy. conversion is also 3 to 25x faster than processing the context again. prior cross-model KV reuse methods either train a neural adapter for every pair or require architecturally identical models. this approach is closed-form and training-free, but there are still important limitations. every tested pair belongs to the same model family, such as Qwen → Qwen or Llama → Llama. cross-family transfer is still future work. the tested pairs also share KV head count and per-head dimension. mismatched configurations remain untested. and the work currently covers dense full-attention models only, not sliding-window or attention-recurrent hybrids. link to the paper: https://t.co/YBMaGW4EWb we wrote a first-principles breakdown of how KV caching works the article is quoted below.