The Johnson Space Center in Houston has begun recruiting participants for the Moon & Mars Exploration Analog (MMEA) ground expedition. This unique program, scheduled to start no earlier than August 2027, involves the complete isolation of a group of four to six test subjects from the outside world for an entire year.
The main goal of the expedition is to provide scientists with the opportunity to prepare for future real missions to the Moon and Mars. Organizers are looking for US citizens or individuals with a Green Card who meet specific criteria: age between 30 and 55, height no more than 188 centimeters, and physical and psychological endurance.
Candidates must have a higher education in exact sciences, engineering, or mathematics and possess significant work experience. The selection process includes a thorough medical commission and in-depth psychological testing.
For the entire year, volunteers will be housed in the closed Pavilion 220 in Texas. Here, scientists have recreated three interconnected spaces to model different phases of space travel. Initially, the crew will reside in a two-story cylindrical spacecraft approximately 63 square meters in area, simulating a long journey.
Upon arrival at the simulated planet, participants will move into a spacious single-story Martian habitat, which covers about 158 square meters and was created using a 3D printer. This house is equipped with living quarters, a kitchen, a laboratory, and a sand area for conducting both virtual and physical surface excursions.
A compact rover designed for two people is provided for more detailed research. In this limited cabin, equipped with sleeping berths and a portable toilet, participants will spend several days in complete autonomy, allowing an assessment of their ability to cope with extreme crowding and temporary team separation.
Life in the complex will be strictly regulated according to a military schedule, similar to the routine of the International Space Station (ISS): eight hours of sleep, clearly defined working hours, physical exercise, and life support system maintenance. Furthermore, participants will be in a complete information vacuum, excluding access to the internet, social media, television, and regular phone calls.
The only ways to communicate with Earth will be through text news briefs and rare video calls with loved ones and psychological specialists. A delay of several minutes will be used to increase the realism of communication with Mission Control. During the simulation of passing the ship behind the Sun, the crew will face a complete two-week silence.
The volunteers' diet will consist exclusively of dry space rations, while fresh vitamins will be obtained from greenery grown in a small hydroponic farm. These 14 months of strict limitations and constant medical examinations will allow NASA to study aspects such as psychological exhaustion, group cohesion, and bodily function under stress, ensuring the safety of future space missions.
Additionally, it should be noted that NASA previously allocated about $600 million to three private companies to prepare for the creation of a Lunar base. The task of these companies is to deliver necessary equipment to the Moon using unmanned vehicles.
Thinking Machines Lab, a company founded by Mira Murati, former CTO of OpenAI, announced the launch of its first artificial intelligence model, named Inkling. This system utilizes an open weights structure, allowing developers and organizations to modify the model using their own datasets.
Thinking Machines Lab developed Inkling aiming for a balance between performance, cost, and adaptability, in contrast to direct competition with the market's more sophisticated proprietary models. Although Inkling contains a total of 975 billion parameters, a smaller number compared to the estimates of major closed models from companies like OpenAI and Anthropic, the company's proposal is to provide a more adjustable option for users.
The company itself stated that Inkling was trained to be a comprehensive and balanced foundational model, strong in various areas and sufficiently flexible for adaptation. It was emphasized that it is not currently the most powerful model, whether open or closed. Despite having nearly one trillion parameters, only 41 billion are activated during each query, a tactic that, according to Thinking Machines, reduces computational expenses and speeds up task execution.
This launch occurs in a context where a part of the AI industry advocates for open weights models, opposing the closed development method adopted by corporations such as OpenAI and Anthropic. Thinking Machines points out that this initiative reflects a broader sector reaction to the concept of a 'walled garden,' where systems remain under the exclusive control of creating companies.
Industry leaders, including Alex Karp, CEO of Palantir, and Satya Nadella, of Microsoft, had already issued warnings about the risk of companies compromising their own business models by feeding centralized generalist systems with strategic institutional data they do not control.
Additionally, the company mentions that this launch integrates a Silicon Valley effort to increase the availability of open models developed in the United States, offering an alternative to solutions provided by the Chinese company Alibaba and startups like Z.ai. The company observes that many American companies have been migrating to Chinese open-weights models for less complex AI tasks, seeking to reduce expenses and diversify strategies.
Inkling can be customized through Tinker, the first product offered by Thinking Machines. This cloud-based platform allows researchers and developers to fine-tune large AI models without needing to manage the supercomputing infrastructure used in training. The goal is to enable this work to be performed even from a laptop computer.
Last month, Thinking Machines and the hedge fund Bridgewater Associates published a report detailing the use of Tinker to optimize the Chinese open-weights model Qwen3-235B with Bridgewater's own data. According to the report, the resulting model outperformed GPT-5 and Claude Opus in financial document classification while reducing computing costs by over thirteen times.
Thinking Machines communicated that the pre-training of Inkling was done from scratch, utilizing 45 million tokens composed of text, images, audio, and video. In the post-training phase, intended to shape the model's behavior, the company combined distillation techniques—which use other AI models as references—with its own reinforcement learning process. The entire training process was conducted using cutting-edge Nvidia hardware.
In March, the two companies formalized a long-term partnership in which Nvidia invested in Thinking Machines. As part of this agreement, the startup committed to implementing at least one gigawatt of next-generation chips to train and operate its future AI models.
Thinking Machines assured that Inkling underwent rigorous safety testing before its launch. Scenarios analyzed included the model's potential to assist in the creation of biological weapons and to support hackers in cyberattacks, and the system performed well in these tests. However, the company acknowledges that it is still investigating how the safeguards integrated into the model could be altered by third parties, a concern raised by developers of proprietary models due to the system's open nature.
On Friday (10th), Thinking Machines released its first manifesto, exposing its perspective on the future of AI. In this document, the company advocates for a decentralized development model based on local knowledge. It compares the current standard of closed-source AI labs to the concept of 'central planning.' According to the company, although the latter may be effective in specific tasks, such as mathematics and chess, it does not accurately reflect the dynamics of daily human work.
Citing economist Friedrich Hayek, the company argues that 'central planning fails not due to a lack of intelligence, but because of the nature of productive knowledge: tacit, local, ephemeral, and privately held by those who acquired it through their work.' Thus, attempting to consolidate knowledge for a centralized intelligence faces similar challenges.
OpenAI launched the Codex Micro, its first proprietary physical product, in pre-sale this Wednesday (15th). This compact keyboard was developed in collaboration with Work Louder to assist in monitoring and managing Codex's artificial intelligence agents.
Available for US$ 230 (equivalent to about R$ 1,171.09, excluding taxes), the accessory is designed for professionals who perform various automated programming tasks and need to track process progress without relying solely on the computer screen.
The Codex Micro integrates a customizable keyboard with specific functionalities for interacting with AIs. It features six illuminated translucent keys that serve as a quick monitoring panel, visually indicating the status of processes executed by Codex.
The light indicators inform whether a task is processing, paused, completed, or requires human intervention. If an agent needs a decision or encounters an error, the user receives a visual alert directly on the keyboard. This allows professionals to monitor multiple workflows simultaneously.
By pressing one of the illuminated keys, the user can access the respective Codex window on the computer. In addition to monitoring functions, the device offers configurable buttons to execute common actions, such as approving or rejecting code modifications, creating project branches, and activating voice commands. The set also includes extra covers for physical customization.
The device's structure is not limited to use with Codex; the user can configure different shortcuts and switch between command layers to apply it to other computer activities. The partnership with Work Louder ensures a limited edition while stocks last.
This launch marks an attempt by OpenAI to expand its reach beyond language model-based services. Previously, the company explored hardware initiatives, including a project with the design group led by Jony Ive, aiming for a screenless portable device controlled by audio and visuals. However, this project faced technical and design challenges, as well as uncertainties generated by a legal dispute between Apple and former company employees.
The Codex Micro thus emerges as the first physical piece officially linked to the OpenAI brand, providing a specialized tool for those who use programming agents daily.
Google announced a major restructuring of Google Images, aiming to transform its search engine into a platform focused on visual inspiration and discovery, mirroring the experience offered by Pinterest.
The company plans to change the service from a traditional search tool to a dynamic and navigable gallery of visual content available on the internet. According to Google, this change aims to expand users' navigation options. Additionally, the inclusion of Artificial Intelligence (AI) image generation serves as an alternative for those who cannot find exactly what they are looking for or wish to visualize new concepts without relying on external services.
The tech giant is adopting an approach similar to Pinterest, which has established itself as a place to organize and discover visual inspiration in niches such as fashion, decoration, and travel. This transformation is expected to increase the time users spend in the Google ecosystem and strengthen advertising revenue. Another goal is to keep users on the platform by allowing them to create images that do not exist on the web, rather than directing them to third-party tools like ChatGPT.
In the new version of Google Images, a gallery called 'For you' will be introduced, presenting customized images according to the user's interests and browsing history. Similar to Pinterest, this gallery was designed for continuous navigation and will have its content updated in real time. Users will have the ability to save materials in 'collections,' which can be organized in tabs above the main gallery, allowing them to create specific categories to save inspiration, such as travel look ideas or suggestions for decorating a reading nook.
The new Google Images will be launched in the coming weeks for desktop users in the United States, requiring access via a Google Account.
In addition to the redesign of Google Images, the company has also implemented the function to generate images directly within Search. This feature was developed to address situations where the user has a very specific idea of an image that does not yet exist online. This functionality will be incorporated into Search's AI Overviews and will use the new Nano Banana model to convert textual prompts into personalized images.
Google also mentioned that the new feature can be used to simulate changes in environments or explore design possibilities, such as visualizing a room painted red or imagining a seaside-themed bedroom. The release of image generation in AI Overviews will begin in the coming weeks, in English, for all regions that already support image creation in AI Mode.
