Random Llama

One-man dev studio shipping full-stack web apps, mobile tools & games. Local AI Researcher who’s Probably caffeinated. 🦙 I own a minivan 🚐, don't judge me!

[Random Llama @Beast_Llama_]

Two Sparks. Cache fix live. After 2 evictions, 75,776/76,048 tokens reused. First token in 473 ms. Recalls 10/10, 32K 16/16.
Big credit to DeepSeek, Ash Hart/TensorFold, Albert Lee, Mia's AI Lab, drowzeys, turboderp, SparkDash + crew.
v0.3.0 + credits 👇
https://t.co/YxouXGxgzh

[Random Llama @Beast_Llama_]

OpenAI's GPT-6 Cyber hit 48.5% on ExploitGym and is locked to vetted pentesters. Meanwhile Ramp says business AI token spend grew just 1.5% across Aug and Sept. Hacker models up, budgets flat. #AI #CyberSecurity

Full rundown: https://t.co/Vg34yL1egC

[Random Llama @Beast_Llama_]

@niklaslenz_ai

[Random Llama @Beast_Llama_]

Terrain Explorer on DoesItFlood is a tool for checking flood risks using interactive maps that show elevation and water flow.

https://t.co/zEwugGX3OS

[Mia @MiaAI_lab]

User experience is everything.

[Random Llama @Beast_Llama_]

@MiaAI_lab

[Random Llama @Beast_Llama_]

Quick update from the two-Spark rabbit hole: we fixed a long-context cache bug that made agents prefill old history again. Same weights, quality, 524K shared pool and 8 lanes. v0.2 source-preview recipe + patches/data:
https://t.co/bK9gmsAi7S

[Random Llama @Beast_Llama_]

Matched 3 x 170K test:
Measured wave: 168.11s -> 6.75s
Whole benchmark job with cold priming: 404.47s -> 242.44s (40% less time)

Cold prefill barely changed. Cold + post-overwrite 32K recall: 16/16 each. One paired test; whole-project gain is still unmeasured.

[Random Llama @Beast_Llama_]

Also: eleata, Hector Isaza, Coinupbtc, Robert Wojciechowski, BadAd84 and Jay Leaton; OsakaTX for related scheduler work.

Recipe is a source preview. Public native/GPU qualification is pending. New backing PR (draft):
https://t.co/2RQKkAwpi1

[Random Llama @Beast_Llama_]

two Sparks, eight lanes, 311 tok/s aggregate.

DeepSeek EAQ v1 + matching Keys. Thinking on. 8/8 streams, 400 tokens each, zero prefix-cache hits.

The garage box fans are earning their keep. 🦙

Screenshots, data and the recipe:
https://t.co/8SED8uq6uE

[Random Llama @Beast_Llama_]

Big credit to DeepSeek, Ash Hart + TensorFold contributors, Bertholomus/Albert Lee, Mia's AI Lab, drowzeys, turboderp/ExLlamaV3 and SparkDash. We're building on their work. Full credits + notices:
https://t.co/tTENRfMWW9

[Random Llama @Beast_Llama_]

Credit to DeepSeek, TensorFold's maintainers/contributors, @MiaAI_lab, drowzeys and @jayleaton.

Pinned source, build steps, data, screenshots and full credits:
https://t.co/8SED8uq6uE

Build on it. Tell me what breaks. 🦙

[Random Llama @Beast_Llama_]

Found two actual bugs: failed admission could evict useful prefixes, and one blocked request could hold up work that fit.

Both have CPU regression tests. Two draft PRs for upstream review:
https://t.co/Euw9uIkw0V
https://t.co/d5EidBoiI1

[Random Llama @Beast_Llama_]

@shanemelaugh Is it manipulation or is it evolving?

[Random Llama @Beast_Llama_]

Voters rejected data centers in 40 of 41 local ballot measures Tuesday. The lone holdout: a North Dakota town of under 800. The AI buildout has a neighbor problem, and Meta just admitted it. #AI #DataCenters

More: https://t.co/fm2JNqBxr6

[Random Llama @Beast_Llama_]

Hot take, the kill switch is the feature. Half my agent work is writing the fences. The model is the easy part.