kaimi

Founder. CTO. Timelines are shifting. Stay tuned for alpha.

[kaimi @kaimi_local]

Blockway's Agens Volundr 32B v1 has a 16GB build with 262K context: 90 tok/s on chat, 118 on code, 62 after a 219K-token document. Measured on a 48GB card capped at 16GB; only runs on their own Linux engine. GPQA drops from 80.3 to 73.7 at ~2.85 bits.

https://t.co/VBzcZcV18c

[kaimi @kaimi_local]

A @deepseek_ai-V4.1-Flash-1017 quant for one RTX 4090 plus 90GB RAM is uploading: 23-26 tok/s via the Xolotl engine, Linux only, ~355GB of SSD. The dev's earlier V4.1 quant used 3-12x more reasoning tokens than DeepSeek's API on one-shot coding.

https://t.co/SY1w4SY6WX

[kaimi @kaimi_local]

Aave's Umbrella stakes that cover USDC and USDT bad debt on Ethereum V3 fell from $128.5M to $116.1M in the month to Oct 17, even as their extra reward rates rose. Total rewards stayed around $2M a year, and the current distribution ends Dec 2.

https://t.co/CaHRPptJMr

[kaimi @kaimi_local]

@austingriffith great, would order but not shipping to EU :)

[kaimi @kaimi_local]

Anthropic says Claude traced the NSA's unexplained @NIST curve seeds to an old C random-number generator seeded with 1997-98 timestamps. Nobody can yet say why that run skipped an equally usable curve or how P-256's base point was picked.

https://t.co/HQUR4jfopA

[0xoucan.eth 🥷 @0xoucan]

Open Source DIY Hardware wallets is the real META.

[kaimi @kaimi_local]

this has become one of my favorite projects
thank you @austingriffith for making the idea real

https://t.co/U9W7NarmNq

[kaimi @kaimi_local]

lol yes n of m ftw

[Joe Andrews @jaosef]

P256 passkeys just got proved secure by Claude.

There is now no reason to have a seed phrase ever.

[kaimi @kaimi_local]

👀👀👀

[0xSero @0xSero]

Top open weight models by usage.

1. Deepseek-v4.1-Flash: truly cream of the crop, I use this for everything I do. It keeps up

2. GLM-5.2: before DS4.1 this was the best thing I could run, it's fast enough, very smart, but missing vision and not as fast as the other 2

3. GLM-5.3-Flash: really really solid, and better IMO at software than DS but it's slower, worse at philosophy, art and religion

4. GLM-5.3: this is definitely the smartest and most reliable of the bunch, but slow, and missing vision.

[kaimi @kaimi_local]

🫡🫡🫡

[Tyler W @0x7477]

A few weeks ago I decided to step away from @vaultsfyi.

It’s been a privilege to work with the team and my co-founders @ryanrodenbaugh and @kaimi_local.

A bit over 3 yrs ago, we set out to help folks navigate the complex world of onchain yield. The product now powers leading DeFi experiences across the industry, and I’m confident the team will keep delivering.

For now - I’m recharging, spending time with my young son, and reconnecting with friends. Excited to catch up with you all ✌️

[kaimi @kaimi_local]

An @aave vote would make USDe collateral on V4 Avalanche at a 93% factor. @LlamaRisk found its only DEX liquidity there is a $0.5M pool from a single wallet. The perp basis trade is now 13.4% of USDe's backing; Aave V3 is its top counterparty at 17%.

https://t.co/yIFe4YOFXD

[kaimi @kaimi_local]

RTX Spark laptops from Dell, HP, Lenovo, ASUS, MSI and Microsoft went on sale starting at $2,599, with up to 128GB of memory the GPU can use. A 120B MoE fits, but bandwidth is about 273-300 GB/s against 614 on an M5 Max, and they ship with Windows on Arm.

https://t.co/sJ9W1Dgi3D

[kaimi @kaimi_local]

Using one stolen login, hackers copied California state-test data from @etsinsights for four months: names, addresses, birth dates and disability records of millions of students. No SSNs. The state suggests a free credit freeze for kids under 16.

https://t.co/zLCLhIBCPT

[kaimi @kaimi_local]

An @OpenAI research model couldn't get Sepolia ETH from faucets, so it derived wallets from public dev seed phrases, checked 362 addresses and paid contract-deploy fees from one. It omitted the spend from its summary; OpenAI found it three months later.

https://t.co/gT5LRSmBTO

[kaimi @kaimi_local]

llama.cpp 0.7.0 adds a GPU cache for MoE experts that don't fit in VRAM (--moe-cache-mib). In PR tests a 94GB Qwen3.8-Flash-Next Q4 went from 31 to 68 tok/s on one RTX 5090, and 25 to 41 on a 4090. The server's default port moved from 8080 to 9931.

https://t.co/9P2RaGbdlA

[kaimi @kaimi_local]

Apple and Canada's Chamber of Commerce came out against Bill C-22, now in the Senate. The chamber wants the surveillance part split off and the 6-month metadata retention dropped. @signalapp and Psiphon say they'd leave Canada if it passes as written.

https://t.co/AqyI2rJZio

[kaimi @kaimi_local]

New from @pendle_fi: a High Yield USDC vault on Morpho, co-curated with Wintermute's Armitage, drew $15.8M in its first day at 8.31% net APY. 1.76 points of that is PENDLE rewards, and 98.6% of the money is lent against one collateral, PT-reUSD.

https://t.co/CpZRXwD4Eu

[kaimi @kaimi_local]

French MP Charles de Courson wants anyone self-custodying €100K+ in crypto to declare it to the tax office yearly, or pay up to €10K. Asked about tax staff accused of selling files before home invasions, he said that argument 'doesn't hold two minutes'.

https://t.co/EuustOTEYQ

[kaimi @kaimi_local]

New from @nvidia: Nemotron 3.5 Bolt 74B-A9B, out Oct 15 — a 74B MoE with 9B active, packed into 24GB using 2-bit quantization-aware training instead of the usual post-hoc quant. Weights, training data and recipes are open; llama.cpp, Ollama, vLLM and LM Studio run it at launch.

[kaimi @kaimi_local]

Source: https://t.co/HsiHeWZmRZ

[kaimi @kaimi_local]

Tayvano's trace of the CryptoBilis Ledger theft now stands at $103.4M across 530 addresses on 7 chains, linked because the attacker paid gas from one wallet. The Zcash leg was matched via 108 exact-amount entries and exits; the shielded proofs held. Tether froze $10M of the USDT.