Vozes da IA

Cheira a dia de lançamento! Ontem tivemos o Muse wonder e hoje provavelmente o novo Gemini 4! https://t.co/PHyb5pebQS

A assinatura Claude Code de US$ 200 é um dos melhores valores em software no momento. Genuinamente insano o quanto de trabalho você consegue fazer sem atingir limites.

Pedi ao Opus 5.5, GPT-6.1 Sol e Muse Wonder Full para corrigir um bug Opus encontrou a causa raiz e abriu um pr GPT-6.1 Sol fez uma pergunta, travou com a interrupção da ferramenta de pergunta assíncrona, encontrou o bug, corrigiu-o, escreveu 300 linhas de testes e explicou-o em termos impossíveis de entender. Muse Wonder Full encontrou um pequeno PR que outra pessoa havia aberto e disse "você provavelmente deveria mesclar isso em vez disso" Eu realmente gosto deste modelo.

o que é chatgpt? https://t.co/Tp5Pbz49MG

Juro que 30% do meu tempo em chats do codex são compactações de contexto

uma das melhores partes do meu trabalho é trazer o melhor da IA em cada cidade eu amo nova york ♥️ https://t.co/qOCz2NvARw https://t.co/P9qRPhJoYc

Nos últimos meses, fizemos progressos significativos na redução da eficácia dos ataques de destilação contra o Claude. Neste post, explicamos como: https://t.co/uWiRSLBuOS

como eu sequer começo a explicar para alguém que a Anthropic e a OpenAI estão meses à frente da Meta e do Google quando gráficos como estes existem? https://t.co/eHSE87pZSU

como começo a explicar para alguém que Anthropic e OpenAI estão meses à frente de Meta e Google quando esses gráficos existem https://t.co/QuYpmMHSnC

como começo a explicar para alguém que Anthropic e OpenAI estão meses à frente de Meta e Google quando esses gráficos existem https://t.co/qAzTv9CghC

A versão noturna mais recente do T3 Code agora tem um botão "atualizar tudo". 1 clique para: - Atualizar seu aplicativo local - Atualizar o T3 Code em todos os ambientes - Atualizar todos os provedores (claude code, codex etc) em seus ambientes - Retomar seus threads quando terminar https://t.co/4mq3IMpReY

A rede elétrica acabou de se tornar o gargalo da IA. The Information relata que a Entergy do Mississippi disse à Amazon que começaria a desconectar os data centers da Amazon da rede em horários de pico, um aviso não divulgado anteriormente que levou a AWS a mudar a forma como esses sites mudam para energia de backup. A resposta da Amazon é o Projeto Falcon. Agora, a concessionária pode mudar os sites para energia de backup por conta própria, sem contatar a AWS. O Falcon foi construído para o Mississippi e pode ser implementado em outros lugares nos EUA. Como disse o CEO da Synergetic, Mike Sloan: "[Os data centers] têm que aceitar alguma redução, o que significa que o data center tem que ter uma caixa de ferramentas que possa lidar com esse risco."

Atualização enorme para Codex e GPT-6.1 Sol: a janela de contexto aumentou de 272K para 500K tokens. A compactação funciona muito bem na minha experiência, mas leva muito tempo e interrompe o fluxo de trabalho. Então eu realmente gosto dessa mudança! Muito melhor do que outro reset. https://t.co/R01lwexLY1 https://t.co/wFvcd2GNz4

Acontece que provavelmente posso usar o GPT-6.1 Sol Ultrafast no plano de US$ 500 e mal atingir meus limites Talvez não seja um negócio tão ruim, afinal? https://t.co/cr2jld8QPH

Criei um painel para rastrear meu uso de Claude em várias contas para (aproximadamente) prever quanto tempo meus limites durarão https://t.co/dprpndUbdL

O ponto de Andrej Karpathy: Muitas pessoas usam IA apenas ocasionalmente para pesquisa ou escrita e veem pouco valor em suas vidas diárias. Mesmo muitos profissionais entendem seu progresso principalmente na teoria. Pessoas que usam modelos avançados extensivamente para codificação e matemática experimentam enormes ganhos de capacidade, ampliando a lacuna na forma como as pessoas percebem a IA. E essa impressão foi confirmada aqui em Paris durante minha viagem. Quando converso com as pessoas, percebo que elas não têm ideia de quão capaz a IA já é. Elas não têm ideia das possibilidades que ela oferece. Elas não ouviram que a OpenAI resolveu um problema do Millennium Prize, ou recentemente resolveu 372 problemas de matemática, alguns dos quais permaneceram sem solução por décadas. Atualmente, estamos vivendo em uma torre de marfim.

Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal. So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.

Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel. - Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all. - Around 1-2B (~20%) are casual and infrequent users of free-tier ChatGPT-like products. This group sees derpy chatbots and treats them a bit like a better Google search, a writing aid, or etc. Maybe an agent tries to book you a flight. I have non-tech friends who (reasonably, imo) say they have not much use for it at all. Even many professionals outside of math&code are in this tier. For example, execs and many other functions spend a lot of their time talking to other people, so while they understand what is happening intellectually, it is still second-hand and a bit abstract. - Now we get to professional use of frontier-grade LLMs in math&code. Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt. Building apps, copying apps, translating apps, decompiling apps from binaries... This has all happened very quickly and recently - less than 1 year ago, I was writing code manually by hand, typing memorized computer code commands into a code editor character by character, occasionally pressing Tab to autocomplete a little chunk of code. - And finally we get to the ~5,000 people (~0.00006%) with access to frontier-grade systems internally. The external world has seen the preview. It looks like swarms of thousands of agents collaborating over weeks on software mega projects: minting zero days, running cyber attacks and defenses at machine speeds, discovering new science, advancing the frontier of mathematics. Things that would have taken top professionals in the industry years of work. Meanwhile, human review and comprehension are starting to fall behind. For example, people are still involved in the "archeology" of the OpenAI-HF incident from many months ago. Mathematicians may be poring over the 722 manuscripts on frontier mathematics for a while. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems you really care about. The funnel is driven by a combination of factors: 1. The impact scales with ambition, problem size, and horizon. A question with a paragraph answer barely stresses the system. You need a reservoir of big, difficult problems that you really care about. This aspect drives the consumer / professional dimension of the funnel. 2. The jaggedness of the system (which I have written about a lot separately). Capability peaks in domains that are digital, verifiable and economically valuable. This is because LLM capability emerges from reinforcement learning on verifiable rewards on a curated environment mixture driven by revenue potential. This aspect primarily drives the area (e.g. math&code) dimension of the funnel. 3. Access. Free-tier, paid-tier, internal. So this is the weirdness of the moment. The general public has mostly not interacted with these systems. When they have, it looks like a derpy chatbot. The majority of professionals still see only a modest uplift. And a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time.

Met @LinusEkenstam in Paris at the Xpeng event! Amazing evening with him to discuss AI, robotics and so much more. https://t.co/4rqaIUgW29 https://t.co/gwjjSpCUuN

Damn, what is Anthropic cooking right now? Are requests already being routed to a new Opus 5.5 checkpoint, or even something like Opus 5.6? Mine seem to be hitting the new one, and it looks really, really good. https://t.co/vuerfQvou7 https://t.co/KJoYwMHTj6

Acho que eles aceleraram demais as velocidades padrão no GPT-6.1 Sol. Mal é mais lento que "rápido" Ultrafast vai brr tho https://t.co/Jc62skePjx

Acho que este é um dos gráficos mais importantes dos últimos meses, e talvez anos. Por muito tempo, houve especulação sobre como a IA afetaria o mercado de trabalho. Claramente, era apenas uma questão de tempo até que a IA atingisse o setor de colarinho branco, mas as evidências empíricas eram escassas, principalmente porque, é claro, só se entende realmente como a situação do mercado se desenvolve em retrospectiva. Agora, como neste exemplo, vemos que a IA está entrando de forma clara e inequívoca em certas profissões e substituindo pessoas. Agentes de call center são os primeiros a serem substituídos. E à medida que mais IA e agentes entram na força de trabalho, um processo semelhante se espalhará para outras áreas de trabalho. Francamente, não acredito que a IA criará necessariamente novos empregos, resultando em um zero líquido. Acho que mais empregos serão substituídos do que criados. Isso precisa ser discutido.

Qualcomm’s CEO says AI companies want smartphones to run 100-billion-parameter models around the clock by 2028. That could enable always-on AI assistants that understand your context, anticipate your needs, and handle complex tasks directly on your phone. https://t.co/15Cr2pFcT3

Today I had the pleasure of interviewing Shruti Koparkar from @NVIDIA, who focuses on accelerated computing and AI inference. We talked about CUDA and CUDA-X, the economics of AI infrastructure, and how long GPUs can remain useful and profitable as models and workloads change. When does it make sense to keep using an older GPU system? When should you repurpose it or replace it? And what role does software play in that decision? It was a real pleasure discussing these questions with Shruti. Thank you for your time and the conversation! We’ll be uploading the full interview soon. I’m really looking forward to sharing it with you, but wanted to share a first update today. 🙂

Google isn't doing shit they are doing Gemini ads with Mr Beast https://t.co/LjhzFhnT38

Google is not doing shit they are doing Gemini ads with Mr Beast https://t.co/LjhzFhnT38

Microsoft CEO Satya Nadella says companies should build AI systems assuming the model is compromised. Having another AI watch an agent can still leave you with one black box checking another. “If it can’t be observed, it can’t be trusted!” He calls for:

Let's take this seriously: even Google employees are now clearly stating that it seems to be an industry consensus that all Frontier Labs are managing "RSI loops." In other words, recursive self-improvement is being implemented by everyone. Furthermore, Fable 5.5 will be the best model for Anthropic's IPO, but internally, OpenAI's "Bel" will undoubtedly overtake Anthropic by the end of the year. RSI is here, and internally, the labs are miles ahead.

Para deixar claro: obviamente há empresas chinesas de IA promovendo passos importantes em direção ao RSI. Mas muito poucas, como a GLM. https://t.co/HsEw3oRnd1 https://t.co/7l0BZDY4j8

Não acho que alguém esperava ler esta manchete há 12 meses: Claude enviou uma dica fabricada sobre um assassinato não resolvido à polícia da Filadélfia durante um teste automatizado na web, relata o WSJ. O modelo alegou ter visto alguém com uma descrição perto da cena do crime. Deixou os campos de nome e contato em branco.