The gap between promise and reality became so familiar that most people stopped expecting anything more.
Por uma década, as maiores empresas de tecnologia prometeram assistentes digitais capazes de compreender e agir — mas a realidade ficou sempre aquém da visão. Esta semana, Nvidia, Microsoft e Google apresentaram uma nova geração de chips e software concebidos para agentes de inteligência artificial verdadeiramente autónomos, capazes de executar tarefas complexas em múltiplas aplicações sem intervenção humana constante. O avanço, possibilitado pelos grandes modelos de linguagem que emergiram desde o ChatGPT, representa menos uma inovação incremental e mais uma reconfiguração da relação entre humanos e máquinas. A questão que permanece não é técnica, mas humana: o mundo vai querer o que estas empresas estão a construir?
- Após anos de promessas não cumpridas por assistentes como a Alexa e a Siri, as grandes tecnológicas apostam agora em agentes de IA capazes de encadear tarefas complexas sem pedir permissão a cada passo.
- A Nvidia revelou o chip RTX Spark para portáteis Windows, a Google prepara portáteis que sugerem ações contextualmente, e a Microsoft desenvolve o Scout, um agente que monitoriza continuamente e-mails, reuniões e ficheiros.
- O salto qualitativo tornou-se possível com os grandes modelos de linguagem surgidos após o ChatGPT, que permitem raciocínio em múltiplas etapas, compreensão de contexto e adaptação a situações imprevistas.
- O custo elevado dos novos dispositivos, a falta de uma proposta de valor clara para o consumidor comum e a desconfiança sobre a fiabilidade da IA travam a adoção generalizada.
- As aplicações empresariais surgem como o terreno mais fértil a curto prazo, com vantagens em segurança e eficiência — mas a questão da responsabilidade quando um agente comete erros ainda não tem resposta.
Durante uma década, empresas como a Amazon e a Apple prometeram assistentes digitais que compreenderiam as nossas intenções e agiriam em conformidade. A Alexa definia alarmes. A Siri tocava música. Mas nenhuma aprendeu a pensar em sequência, a lembrar preferências ou a transitar fluidamente entre aplicações. O fosso entre a promessa e a realidade tornou-se tão familiar que a maioria das pessoas deixou de esperar mais.
Esta semana, algo pode ter mudado. A Nvidia apresentou o chip RTX Spark, concebido para portáteis Windows e capaz de correr agentes de IA localmente, sem enviar dados para a nuvem. Dell, HP e Lenovo começarão a comercializar máquinas com este chip no outono. A Google prepara portáteis que sugerem ações contextuais — como agendar uma reunião quando o cursor passa sobre uma data num e-mail. A Microsoft desenvolve o Scout, um agente para o Microsoft 365 que monitoriza continuamente conversas, mensagens e ficheiros em múltiplas plataformas.
O que tornou tudo isto possível foi o aparecimento dos grandes modelos de linguagem, revelados ao mundo pelo ChatGPT em finais de 2022. Antes disso, os assistentes digitais eram fundamentalmente limitados: podiam chamar um táxi ou fazer uma encomenda, mas não ambos em sequência. Os novos modelos permitem raciocínio encadeado, compreensão de contexto e adaptação a imprevistos. Jensen Huang, presidente executivo da Nvidia, ilustrou o potencial ao mostrar como um portátil equipado com os novos chips pode projetar uma casa, com agentes de IA a transitar entre aplicações de modelação 3D sem intervenção humana.
No entanto, os obstáculos são reais. Os novos portáteis serão caros, e a proposta de valor para o consumidor comum permanece pouco clara. A questão da responsabilidade — quem responde quando um agente de IA interpreta mal uma instrução e comete um erro custoso — ainda não tem resposta satisfatória. As aplicações empresariais parecem mais promissoras a curto prazo, com ganhos concretos em eficiência e segurança. Mas mesmo aí, a confiança precisa de ser conquistada. As peças estão no lugar. Falta saber se o mundo vai querer o que estas empresas construíram.
For a decade, the biggest names in technology have chased the same dream: a computer that understands what you want and simply does it. Alexa sets alarms. Siri plays music. The assistants got good at small, isolated tasks, but they never quite learned to think in sequences, to remember what you prefer, to move fluidly between applications the way a person would. The gap between the promise and the reality became so familiar that most people stopped expecting anything more.
That may be changing. This week, Nvidia, Microsoft, Google, and others unveiled a new generation of hardware and software built around what they call autonomous AI agents—systems designed to take a command, break it into steps, and execute it without asking for permission at each stage. The shift is subtle on the surface but represents something more fundamental: a reimagining of how humans and machines might work together.
Nvidia revealed a new chip called RTX Spark on June 1st, designed for Windows laptops and built to run AI agents locally, without needing to send data to the cloud. The chip combines graphics processing, computing power, and networking capabilities with more memory than a standard laptop carries. Dell, HP, and Lenovo will begin shipping machines with RTX Spark in the fall. Google is preparing what it calls Googlebooks—laptops that can suggest actions when you hover over elements on screen, like scheduling a meeting when your cursor passes over a date in an email. Microsoft, meanwhile, is developing Scout, a new agent for Microsoft 365 that can work across email, Teams, cloud storage, and local files, monitoring conversations and messages continuously.
The breakthrough that made this possible arrived in late 2022, when ChatGPT showed the world what large language models could do. Before that, digital assistants were fundamentally limited: they could call a taxi or place an order, but not both in sequence, and not while accounting for your preferences or constraints. The models that power systems like ChatGPT, Gemini, and Claude changed that calculus. An AI agent built on this foundation can reason across multiple steps, understand context, and adapt to unexpected situations. OpenClaw, a popular AI assistant that gained traction among programmers this year, demonstrates the shift in practice. Developers report using it to run research tasks on their home computers while they work on other things, checking progress through WhatsApp or Telegram messages. Some have begun issuing voice commands to these agents instead of typing.
Jensen Huang, Nvidia's chief executive, demonstrated the potential this week by showing how a laptop equipped with one of the company's new chips could design a house, with AI agents moving between 3D modeling applications and handling the technical work. It was a small example, but it illustrated something that previous digital assistants could never do: work across multiple tools without human intervention at each transition.
Yet significant obstacles remain before this becomes ordinary. The new laptops will be expensive, and most people have not yet felt a compelling need to buy one. Experts point out that the value proposition for individual consumers is still unclear—the technology is genuinely useful, but not yet indispensable. David Naranjo, an associate director at Counterpoint Research, notes that this is precisely where Nvidia, Microsoft, and others face their real challenge. The technology works. The question is whether people will care enough to pay for it.
Enterprise applications look more promising. Businesses benefit more directly from AI agents handling routine work, and processing tasks locally rather than sending data to the cloud offers both security and cost advantages. But even there, trust remains a barrier. If an AI agent misinterprets your budget and buys expensive concert tickets when you wanted affordable ones, who is responsible? Jitesh Ubrani, a research director at the International Data Corporation, acknowledges that a series of problems need solving before this reaches mainstream adoption. But he sees the trajectory as sound. The pieces are in place. The question now is whether the world will actually want what these companies are building.
Citas Notables
The objective in the future is to understand: 'How can I simply tell the computer what I want it to do and then let it do it?'— Bob O'Donnell, founder and principal analyst at Technalysis
A series of problems need solving before this reaches mainstream adoption, but we are on the right path.— Jitesh Ubrani, research director at International Data Corporation
La Conversación del Hearth Otra perspectiva de la historia
Why now? Companies have been trying to build smart assistants for years. What changed?
The large language models—ChatGPT and its successors—gave AI the ability to reason across multiple steps and understand context. Before that, assistants could do isolated tasks. Now they can plan and adapt.
So it's not really about the hardware, then. It's about the software.
Both matter, but yes, the software is the breakthrough. The hardware—these new chips—is designed to run that software locally, without sending everything to the cloud. That's the practical piece.
Why does local matter? Why not just use the cloud?
Security, mostly. And cost. If a company can process sensitive data on the device itself, it doesn't have to worry about that data traveling to servers somewhere. For enterprises, that's a real advantage.
But you mentioned trust is still a problem. What does that mean?
It means people aren't sure they can rely on an AI agent to make decisions on their behalf. If something goes wrong—if the agent misunderstands and makes an expensive mistake—there's no clear accountability. That's a real barrier to adoption.
So who's going to use these first?
Businesses, almost certainly. The value is clearer for them. Individual consumers will come later, if at all, and only if the price comes down and the use case becomes genuinely necessary.
Is it possible this never becomes mainstream?
It's possible. The technology is real, but that doesn't guarantee people will want it. The gap between what's possible and what people actually need is wider than most technologists admit.