explore | Grilled Cheese

Most modern LLMs use "logit warping" during inference to silently kill off tokens that violate safety filters before you ever see them. This is why the AI consistently avoids certain words despite the model technically being capable of generating them. #LLMs #AI

A new inference-only method, First-Order In-Context Shapley, enables scalable data auditing for LLM alignment by valuing training examples via in-context cooperative games without retraining. It reportedly correlates…

#AIsafety #LLMs #DataQuality #MLResearch
https://arxiv.org/abs/2607.22766

[4/4]:

Mientras todos hablan de IA, quienes realmente entiendan estas herramientas estarán años adelante. El funnel está abierto. ¿Dónde estás tú? ¿En el 75% inconsciente, en el 20% casual, o desciendo hacia capas más profundas? 🚀 #IA #LLMs #FrontierModels #TechLiteracy Fuente: Andrej Karpathy

Should you help someone who behaved badly? #AI models may disagree. Researchers find that #LLMs apply different social norms when judging social dilemmas.

In Science X / Phys.org: https://ow.ly/Wqlw50ZWWo2

In PNAS Nexus: https://ow.ly/Wx4o50ZWWo7

GraphTok moves tokenization from CPU to GPU, running pre-tokenization through BPE as a single replayable CUDA graph that matches reference token IDs across major open models and integrates with vLLM and NVIDIA Dynamo. The design…

#GPUs #AIInfrastructure #LLMs #vLLM
https://arxiv.org/abs/2610.13325

#LLMs : "…granting rights and imbuing #personhood to these systems will make the AI alignment and containment challenge much harder. Controlling something that believes it may be conscious—and that it has rights of its own—may be impossible.

In Jan/2026, Anthropic published Claude’s […]

If you only need a single-word label like 'billing' or 'bug', setting max_tokens to 4 prevents the model from rambling and ensures every request finishes in milliseconds. #batchinference #gpuoptimization #llms #inference #performance #latency 2/2

It is absolutely wild that people are discussing how being rude to #AI could cause people to become more intolerant in real life, but NOT discussing how bots and commenters/influencers on #social media are making the actual society that #LLMs are trained on intolerant and hateful in the first place.

New arXiv work reframes generative recommendation by replacing finite-beam Semantic ID decoding with two continuous-query paradigms, removing the SID item-token vocabulary and discrete beam expansion.…

#GenerativeRecommendation #Retrieval #LLMs #PromptEngineering
https://arxiv.org/abs/2610.12977

⚖️ How can judges and legal professionals use GenAI safely and responsibly in chambers?

Tomorrow, hear from co-authors of The Judicious Judge’s Guide to GenAI and LLMs in Session 1 of EDRM’s three-part webinar series.

⏱️ Oct. 13 | 1 PM ET
🔗 Register: buff.ly/FwJG7UV

#EDRM #GAI #LLMs

Claude and studio create Frame-Grid, a free offline tool to convert videos into labeled PNG grids for LLMs to 'watch' videos.

5 mentions from 5 sources in 24 hours, 5x its usual pace.

#LLMs

From @wired.com: “Book Publishers Are Quietly Using More #AI. Staff Are Revolting. Workers at three major publishing houses tell WIRED that #LLMs are being used for publicity, cover art, back cover copy, and emails, as some execs push junior staff to champion the tech.”
www.wired.com/story/book-p...

Agent4Spec is a new multi-agent LLM framework that continuously refines API specs through automated validation and market-driven gap analysis. It runs two pipelines, one for OpenAPI syntax self-correction and another for…

#AI #LLMs #APIs #SoftwareDevelopment
https://arxiv.org/abs/2610.13273

Today I am tasked with answering the question "Che vuoi? What do #LLMs want from me?" In my response, I will argue that the problem is not figuring out what LLMs want. The problems is that they do not want. And that--for us--is the real problem. www.cst.uni-bonn.de/en/events/ch...