LLM

Nova técnica revoluciona limitações de memória RAM.

Desvendando ambientes de aprendizado por reforço para inteligência artificial com Stefano Fiorucci

Nova técnica revoluciona o uso de inteligência artificial em dispositivos móveis.

Treinamento de IA nunca foi tão simples, graças à Apple!

Modelos de linguagem não sentem — mas organizam o comportamento como se sentissem. Essa afirmação não é metafórica.

Claude Code opened to any LLM — OpenAI, Gemini, DeepSeek, Ollama, and 200+ models via OpenAI-compatible API shim · Decentralized git repository on gitlawb.

Inteligência artificial: ameaça à inteligência humana?

Revolutionários em inteligência artificial resolvem o problema de memória da IA!

The site, whose policies are subject to change, has struggled with the issue of AI-generated writing.

Você acha que o Claude está “gastando sua cota rápido demais”? Talvez o problema não esteja no modelo… Tenho visto muita gente reclamando disso nos últimos meses. Principalmente quem usa plano Max…

Mistério sem nome da inteligência artificial que viralizou na internet domina as manchetes e acende o debate sobre privacidade digital.

Guia definitivo para IA local: hardware, configuração e modelos para rodar no seu ambiente

Artificial intelligence models are multiplying fast, and competition is stiff. With so many players crowding the space, which one will be the best — and who decides that? Arena, formerly LM Arena, has emerged as the de facto public leaderboard for frontier LLMs, influencing funding, launches, and PR cycles. In just seven months, the startup went from a UC Berkeley PhD research […]

The company has raised $12 million in seed funding to build an AI operating system for enterprise.

Artificial intelligence models are multiplying fast, and competition is stiff. With so many players crowding the space, which one will be the best — and who decides that? Arena, formerly LM Arena, has emerged as the de facto public leaderboard for frontier LLMs, influencing funding, launches, and PR cycles. In just seven months, the startup went from a UC Berkeley PhD research […]

Apresenta os modelos GPT-5.4 Mini e Nano da OpenAI, destacando sua velocidade, capacidade e aplicações potenciais em IA, com comparação aos modelos anteriores.

Por IBM Technology

Por Alex Ziskind

Por Nick Saraev

Por Universe of AI

Por Cole Medin

Por Matthew Berman

The company says the new model will reduce the "cringe" that's been annoying its users for months.

More and more senior researchers in AI seem to believe that LLMs aren’t the path to achieving AGI. The latest to join this...

Por Software Engineer Meets AI

The company open sourced an 8-billion-parameter LLM, Steerling-8B, trained with a new architecture designed to make its actions easily interpretable.

Por Universe of AI

Por AI Jason

Por AI ProgBr

AI coding tools have enabled a flood of bad code that threatens to overwhelm many projects. Building new features is easier, but maintaining them is just as hard.

Por David Shapiro

Por Tom Delalande

The vision of human-level machine intelligence laid out by Alan Turing in the 1950s is now a reality. Eyes unclouded by dread or hype will help us to prepare for what comes next.

Por AI ProgBr

Acabei de testar o Claude Sonnet 4.6 em produção. Opus 4.5 não conseguiu completar. Sonnet 4.6 entregou. Mesma qualidade. 40% mais barato. Artigo completo com breakdown técnico, benchmarks…

LLMs have recently helped find solutions to a number of minor longstanding problems. But a new plan called First Proof is really putting them to the test

Por Dracon Dev & AI

Por NetworkChuck

Por Steve Eisman

Large language models (LLMs) have fundamentally transformed automated software development by enabling direct translation of natural language descriptions into functional code, driving commercial adoption through tools like Github Copilot (Microsoft), Cursor (Anysphere), Trae (ByteDance), and Claude Code (Anthropic). While the field has evolved dramatically from rule-based systems to Transformer-based architectures, achieving performance improvements from single-digit to over 95\% success rates on benchmarks like HumanEval. In this work, we provide a comprehensive synthesis and practical guide (a series of analytic and probing experiments) about code LLMs, systematically examining the complete model life cycle from data curation to post-training through advanced prompting paradigms, code pre-training, supervised fine-tuning, reinforcement learning, and autonomous coding agents. We analyze the code capability of the general LLMs (GPT-4, Claude, LLaMA) and code-specialized LLMs (StarCoder, Code LLaMA, DeepSeek-Coder, and QwenCoder), critically examining the techniques, design decisions, and trade-offs. Further, we articulate the research-practice gap between academic research (e.g., benchmarks and tasks) and real-world deployment (e.g., software-related code tasks), including code correctness, security, contextual awareness of large codebases, and integration with development workflows, and map promising research directions to practical needs. Last, we conduct a series of experiments to provide a comprehensive analysis of code pre-training, supervised fine-tuning, and reinforcement learning, covering scaling law, framework selection, hyperparameter sensitivity, model architectures, and dataset comparisons.

Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning remain elusive. Recent reasoning models outperform comparable instruction-tuned models on complex cognitive tasks, attributed to extended computation through longer chains of thought. Here we show that enhanced reasoning emerges not from extended computation alone, but from simulating multi-agent-like interactions -- a society of thought -- which enables diversification and debate among internal cognitive perspectives characterized by distinct personality traits and domain expertise. Through quantitative analysis and mechanistic interpretability methods applied to reasoning traces, we find that reasoning models like DeepSeek-R1 and QwQ-32B exhibit much greater perspective diversity than instruction-tuned models, activating broader conflict between heterogeneous personality- and expertise-related features during reasoning. This multi-agent structure manifests in conversational behaviors, including question-answering, perspective shifts, and the reconciliation of conflicting views, and in socio-emotional roles that characterize sharp back-and-forth conversations, together accounting for the accuracy advantage in reasoning tasks. Controlled reinforcement learning experiments reveal that base models increase conversational behaviors when rewarded solely for reasoning accuracy, and fine-tuning models with conversational scaffolding accelerates reasoning improvement over base models. These findings indicate that the social organization of thought enables effective exploration of solution spaces. We suggest that reasoning models establish a computational parallel to collective intelligence in human groups, where diversity enables superior problem-solving when systematically structured, which suggests new opportunities for agent organization to harness the wisdom of crowds.

Por Pourya Kordi

💡 More intelligence means more context — and less human friction. Welcome to Context Engineering 2.0. This paper dives deep into one of the most fundamental, yet underdefined, frontiers of modern...

Por TecnoSpeed

tecnospeedti on September 9, 2025: "A cada conversa com o Rodrigo Palhano, a cabeça borbulha de ideias — e várias delas super práticas pro dia a dia. 🚀 Dias 7 e 8 de outubro, no #TecnouPdate em São Paulo, ele vai estar no palco trazendo ainda mais provocações sobre IA e o futuro da linguagem natural. ✨ E pode anotar: um lançamento importante vai acontecer. 👀 Garanta sua vaga agora e não fique de fora dessa transformação (link na bio)! 🚀 #Tecnoupdate #IA #LLM #ODS".

O GPT-5 está dando realmente o que falar. Mas, se nos abstrairmos do hype midiático em torno dos LLMs, que repete ad nauseam narrativas sobre avanços revolucionários, e fizermos uma análise técnica racional, vamos identificar fatos básicos. Os fatos são que os LLMs persistem como modelos estatísticos de autopreenchimento e não agentes cognitivos. As alucinações persistem, pois são inerentes à arquitetura Transformer, e a ausência de grounding semântico, faz com que os seus outputs sejam otimizações de distribuição léxica e não representações de intencionalidade. Ponto. Assim, se alguém se sentir surpreso com as falhas do GPT-5, isso apenas revela desconhecimento dos limites arquiteturais do transformer, como ausência de modelo de mundo, incapacidade de inferência causal e dependência crítica de viéses humanos.   Os que atribuíram "confiabilidade" a sistemas probabilisticamente instáveis estão vendo agora que foi uma bola fora.  A indústria alocou capital em narrativas, e não em saltos arquiteturais. Os transformers são máquinas de correlação superficial. Não são o caminho para a AGI. E estão batendo no teto. Os ganhos agora são marginais, com escalonamento paramétrico ≠ progresso qualitativo.

Por OpenAI

Por AIM Network

🚨 Some are calling LLMs "Shoggoths": uncontrollable creatures that are tamed with fine-tuning techniques. Post-training has its limitations, and researchers have recently detected