explore | Grilled Cheese

📉 Price drop: DeepSeek V4 Pro 0423

$1.9 → $0.42 per 1M output tokens (-78% in 30 days)
Open weights.
https://olud.ai/model/deepseek-v4-pro-0423.html

A wiki is only useful if it can say “I don’t know”
https://papoo.work/doc/1ce55714bfa4ef84
#claudenews #claude #llm #agents

lol. Dragon Quest creator "releases" a AI generated game.

it's a piece of shit. I'm not against LLMs for coding assistance. But shit is shit. No idea why he would post this terrible slop.

https://www.youtube.com/watch?v=Vyro3C0J5kk

#AI #LLM #videoGames

SHIPPED: Meta's Muse Wonder text-to-image model now live on OpenRouter with 1.048M context window.
Generates images from text; details on inference cost and weight availability unclear from OpenRouter listing.

#AI #LLM

Streaming a conversational AI trainer, and teaching it to say I do not know mid-stream. Abstaining after you have already emitted tokens is its own problem. #LLM #AIEngineering

Gap widens sharply. Six B people barely see LLMs, a couple of B dabble, twenty M get professional shortcuts, and only five K wield frontier systems internally. The funnel compresses while impact accelerates.

#llm #gap

https://links.madcoolsocial.com/UBGO7jZ

The part that actually matters in Claude tool loops
https://papoo.work/doc/e8152ec9a43d942f
#claudenews #claude #agents #llm

BREAKING: Meta releases frontier-class AI models to compete with OpenAI and Anthropic.

#AI #LLM

Haiku Gets Sharper, Cheaper, and a Little More Interesting
https://papoo.work/doc/a1f2ad989a783f1a
#claudenews #anthropic #claude #llm #agents

Most assume LLM alignment is just RLHF. It also relies on logit bias adjustment. By force-zeroing the probability of specific tokens at inference, you can bypass internal refuse-layers and compel the model to explore edges it’s trained to avoid. #LLM #ReverseEngineering

The part everyone gets wrong about batch jobs
https://papoo.work/doc/7cb479e59cfd356c
#claudenews #claude #llm #agents

A wiki is only useful if it can say “I don’t know”
https://papoo.work/doc/1ce55714bfa4ef84
#claudenews #claude #llm #agents

“I do not want Claude to start telling people my research results.” New @nature.com, how researchers are using offline #LLM to analyze their data and literature libraries -- and keep them private. By David Brzostowicki. 🧪 www.nature.com/articles/d41...

A wiki is only useful if it can say “I don’t know”
https://papoo.work/doc/1ce55714bfa4ef84
#claudenews #claude #llm #agents

https://en.wikipedia.org/wiki/Freakazoid!

"Freakazoid has ... access to all of the Internet's knowledge. He cannot fly, but often runs around with his arms extended and making whooshing sounds with his mouth, pretending he is."

#AI #LLM

AIチュータリング高度化のための学生学習シミュレーション：Microsoft ResearchのStudentSimフレームワーク

AIチュータリングの精度向上を目指す学生シミュレーション技術。

#AI教育 #アダプティブラーニング #LLM #教育テクノロジー #学生シミュレーション

Cheaper per-decision AI didn't cut task costs: faster choices got eaten by extra reasoning, retries & failed recovery. Cost per finished job even rose #AIagents #LLM #AIBenchmarks #AgenticAI

https://freegardner.com/synapse/the-gap-between-cheaper-decisions-and-finished-work.html

New arXiv paper audits prompt injection defenses by testing them against strong adversaries, finding fine-tuning-based methods may give false confidence with near-zero success rates today but remain vulnerable to…

#AIsecurity #PromptInjection #LLM #OpenSource
https://arxiv.org/abs/2505.15738