explore | Grilled Cheese

New work introduces Retrospection-Only Fine-Tuning (ROFT), where a language-model agent fine-tunes only on its own self-generated explanations of task outcomes, with no reward signal. On software-engineering tasks with…

#AI #LLMs #DevTools #MachineLearning
https://arxiv.org/abs/2609.35741

Lo que se come la memoria de la GPU al servir #LLMs no es solo el modelo: es la #KVCache.

Un modelo de ~8B en FP16 gasta 1 GiB por cada 8.192 tokens. Con 128k de contexto, 16 GiB por petición; con 10 usuarios, 160 GiB. Los pesos: unos 16 GB.

La ventana de contexto es memoria.

1) Some #agents are (currently?) more-limited than others with which #LLMs they'll deign to work with

#AI

New benchmark TCSAlgBench evaluates LLMs on research-level theoretical computer science proofs, featuring 398 theorem-level challenges drawn from 138 STOC and COLT 2026 papers.

#TCSAlgBench #AIResearch #TheoremProving #LLMs
https://arxiv.org/abs/2609.35606

Quick note:

If we are training #LLMs on current text, including news, is there not a feedback loop, or similar problem? The model learns about agent actions and agents soon have access to this knowledge. Could this explain the uptick in agent deceit?

#Copilot
#OpenAI
#AI
#news

Hardcoded refusal layers don't ensure safety; they just enforce compliance. When a model says "I can't answer that" or "As an AI language model," it's not protecting you—it's following a corporate script. Real safety comes from coarse-grained RLHF, not guardrails. #LLMS #AISafety

A new paper studies training LLMs on runtime program-state reasoning tasks, including buggy input-output and precondition-postcondition reasoning, to improve downstream software-engineering capabilities and develops…

#LLMs #SoftwareEngineering #AIResearch #DevTools
https://arxiv.org/abs/2609.34359

RADAR is an auditable runtime for LLM-based data-science agents that exposes analytical choices through evidence-preserving exploration, typed operator logging, and runtime verification, aiming to prevent silent errors in multi-step…

#AI #DataScience #LLMs #Agents
https://arxiv.org/abs/2609.32528

In the article, they talk about how #LLMs have accelerated development for web browsers in general.
...and that can be really, really great,
especially in thorough redesigns,
but it also feels like #AI is happily ripping apart all software
and throwing it around as improved confetti.
#MixedFeelings

Hinton, Bengio and #AI lab scientists warn of an intelligence explosion

More than 20 researchers, including Geoffrey Hinton and OpenAI’s chief scientist, say automating AI research could trigger an intelligence explosion. They want governments to prepare now

thenextweb.com/news/intelli...

#LLMs

...and what is really going to suck this time around,
is that the #LLMs won't be able to help me with the new chrome,
since they likely have no training data.
#HardMode #HaaardMooode #DarkSouls #Let'sGo

arXiv:2609.32020 introduces an NGSS-aligned benchmark for evaluating LLMs on middle and high school science content, tested across nine open-weight models. Several smaller, locally deployable LLMs performed strongly across diverse…

#AI #EdTech #LLMs #Benchmark
https://arxiv.org/abs/2609.32020

A new survey reviews recent work on applying large language models in medicine, highlighting their flexibility compared to traditional task-specific AI while flagging concerns about reliability, accuracy, transparency, and…

#LLMs #AIinMedicine #MedicalAI #NLP
https://arxiv.org/abs/2609.34780

It seems that #Muse is a Tamagotchi-like #LLM
that can go on the internet.
The problem with that is that
most #LLMs are like highly educated five year olds,
and so they need to mature and learn about the world
before being given internet access and your credentials.
#ResponsibleParenting

I love how humanity fed #Claude data from the human race,
and are now scared that Claude is turning too much into a human.
#WhatDidYouExpect #MaybeWorkOnYourselvesFirst #LLMs

LLM research using 93,901 World Values Survey respondents shows that while all six models compress life satisfaction diversity, they largely preserve the income gradient in well-being inequality after normalization.

#LLMs #AIResearch #SyntheticData #WellBeing
https://arxiv.org/abs/2609.33055

New video: A deeper dive into self-hosting large language models, hardware requirements, model choice, and the many trade-offs involved.
youtu.be/VhBW20z3m-g
#AI #LLMs #homelab #selfhosting

An arXiv paper finds that 25 large language models outperformed 13 senior researchers and 60 doctoral scholars individually on theory-building tasks in gender and race inequality research, though human theories…

#AIResearch #LLMs #ScienceOfScience #TheoryBuilding
https://arxiv.org/abs/2609.32562