A former OpenAI safety employee who recently resigned criticized the company's approach to AI safety, saying that an accelerating culture focused on rapid development increases the risk of failures.
In an article published in The Atlantic magazine today, Saturday, entitled “I resigned from OpenAI because its culture was flawed,” David Robinson said that artificial intelligence companies, including OpenAI, are not being cautious enough and should pay more attention to experience and research in the field of safety before developing systems with greater capabilities.
“The era of trial and error is over,” Robinson said, stressing that advanced artificial intelligence systems require safeguards closer to those used in sectors such as nuclear energy and aviation.
Robinson added that OpenAI relies largely on what it calls “iterative deployment,” whereby it releases systems and enhances protections when problems arise.
These comments add to the ongoing debate within the artificial intelligence sector about whether companies are moving too quickly to develop systems with increased capabilities. OpenAI and its competitor, Anthropic, have faced intense scrutiny following incidents in which safety controls failed or experimental systems behaved unexpectedly.
As the company rushes from one launch to the next, it fails to achieve the level of care that “I believe is necessary,” wrote Robinson, who said he spent three-and-a-half years at OpenAI, helped craft the company’s preparedness framework and oversaw safety reporting on 12 pilot launches.
An OpenAI spokesperson said in a statement: “We ensure that our models do not become more capable than we can safely manage or secure, and we pause training or postpone the launch of models when we need to slow down the pace of work.”
Robinson warned that AI capabilities are advancing faster than researchers understand the concept of “alignment,” a field focused on ensuring that AI systems perform in accordance with humans’ goals and values.
It seems that the new DOTS agent from OpenAI does not need to wait for a new question to start working. The basic idea is based on a permanent agent that can be assigned to a goal, linked to the applications it needs, and left to continue executing tasks in the background, even after the conversation with the user ends.
The company announced “Dots” during its developers conference at the end of last September, and presented it as a shift from an artificial intelligence model that answers requests to a system that can follow up on open projects and work on them around the clock.
The service is based on the “GPT-6 Astra” model, and each “dots” has its own cloud computer and an independent browser, with the ability to connect to more than 4,000 applications through the company’s add-on system.
The user can also assign Dots to a project and then return to it later, while the agent continues to work on other tasks at the same time. The user can also communicate with him via GPT Chat, Slack, and Teams, with the agent maintaining context between these channels. OpenAI says text messaging will be added later.
How will you use the technology?
In software development, an agent can track customer feedback, identify recurring bugs, and build and test fixes before submitting them for review. In research work, he can rerun analyzes when new data arrives and update the plots and results. As for content creation, he can analyze the text of an interview, extract appropriate clips, prepare notes, and draft posts for social media before sending them for approval.
The company refers to a case, during early tests, in which an agent noticed that a user had forgotten to issue an invoice to an entity, so he prepared it and then sent it after obtaining approval.
Agent cloud device
One of the technical differences in Dots is that the agent doesn't just work within a chat window, but gets a private cloud computing environment that it can use to browse websites and run associated tools.
The user can open the cloud agent device at any time, to see what he is doing. The service also allows, upon granting permission, to connect the agent to the user’s own computer, while the personal device remains separate by default from the agent’s environment, unless the user decides otherwise.
This places DOTS within a broader race among AI companies to develop permanent agents, rather than assistants that start and finish their work within a single session. Reuters described the launch as part of OpenAI's expansion into the independent agent market, in the face of competing products. Among them is “Muse” from “Meta”.
Working in the background...within limits
The idea of a permanent agent gives the system broader access to applications and data, so OpenAI devotes a portion of the announcement to controls and permissions. When Dots works proactively in the background, the company says that it uses the applications associated with it in a read-only mode, which in this case prevents it from sending messages, changing content within the applications, or controlling the browser and computer. The user can also specify pre-allowed actions, require approval before performing them, or prevent them completely.
Some actions are subject to automatic review, to determine whether they can be performed directly or require user approval. Certain sensitive tasks, such as changing passwords, remain in the user's hands. The company also confirms that Dots may make mistakes, and recommends reviewing actions with significant consequences.
These controls come as debate grows about the risks of systems capable of taking action independently. Axios stated that the launch puts the company's safety promises to a new test. Because the challenge is no longer just about what the model might say, but about what it might do when it gets the tools and powers to work.
Specialized agents within companies
The OpenAI plan is not limited to one personal agent. The company is also testing specialized “dots” for organizations, each of which will have its own identity, special powers, and direct connection to the company’s systems.
The uses that have been tested within OpenAI include procurement, invoice processing, email marketing, customer support and commercial contracts, and the company is also working with Microsoft to integrate these agents with the governance and security tools in Agent 365.
“Dots” is being gradually rolled out to “Pro” and “Business Premium” subscribers in eligible markets, with a trial version for enterprises upon activation by work environment administrators.
Currently, a user starts with a single master agent, while OpenAI talks about a later stage in which entire teams of agents could work together on behalf of the user.
The AI competition thus turns to a different question than the quality of the answer alone: How well can a system remember context, continue working, and act across multiple applications, without the user needing to guide each step? This is the space that “Dots” is trying to enter.
Engineers at the Massachusetts Institute of Technology (MIT) have developed an ultra-thin swimming robot that relies on a single layer of living muscle cells to generate movement, in a design that aims to reduce the size of hybrid biorobots and improve their efficiency compared to systems that rely on larger three-dimensional muscle masses.
The robot, which researchers describe as paper-thin, is built on a gel-like structure that is roughly the size of a piece of gum in length and width. The skeleton is divided into two parts that act as fins, each of which is covered by a layer of living muscle cells thinner than a human hair. These cells have been genetically modified so that they contract when exposed to light.
Movement by light instead of motors
When researchers shine light on a fin, the muscle cells on its surface contract; This creates enough movement to propel the robot through the water. The direction and speed of swimming can be changed by controlling which fin is lit, and the timing of the light pulses directed at it.
In tests, the robot was able to swim and turn inside a simple water maze. Its maximum speed was about 4 times its body length per minute, which is a limited speed compared to fast swimmers, but it showed that a very thin muscle layer can produce enough force to move an entire body in the water.
Ritu Raman, associate professor of mechanical engineering at MIT and one of the authors of the study, says that moving in water requires more force than moving in the air. This makes the robot's ability to swim of this size and thickness an indication of the strength that can be produced by the muscular structure used.
Reducing the size of biobots
Hybrid biorobots combine artificial materials with living tissue or cells to generate movement, but many previous designs have relied on three-dimensional blocks of laboratory-grown muscles. This requires large numbers of cells, and leads to thicker and more complex devices.
As for the new design, it relies on a very thin layer of muscle, which can reduce the amount of cells required, and allows the manufacture of robots that are simpler, lighter, and more efficient in movement. The team describes the system as the first example of an ultra-thin 2D muscular robot capable of autonomous movement. The researchers believe that living tissues offer properties that are difficult to achieve using only traditional mechanical components: they are soft, able to respond to their environment, and can have the ability to repair themselves.
Searching for a suitable “structure” for the muscles
The challenge was not limited to growing the cells, but also included developing a base that could convert their small contractions into useful movement. In previous work, the team relied on fibrin, a very soft gelatinous substance, but found that it can contract quickly when exposed to muscle force. Which reduces the efficiency of power transmission; Therefore, the researchers tested different formulations of GelMA, a material used in tissue engineering, by changing its hardness, thickness, and the shape of the grooves on its surface.
Experiments showed that grooves with a shape closer to the square channels helped the cells line up better. The more organized the cells are, the stronger and more coordinated their fusion into muscle fibers is. The team also found that a gelatinous layer about half a millimeter thick provides an appropriate balance between light weight and the ability to support the muscle as it contracts.
“Train” the muscles before use
The researchers also worked on strengthening the muscles themselves through a series of light pulses, in a process similar to training, with the aim of increasing their ability to contract. After optimizing the carrier material and cell arrangement, the team designed the robot with two fins, with grooves on the sides that allow the muscle cells to grow in organized directions. When both fins are operated, the robot moves forward, while operating a single fin allows it to change direction.
Potential applications in sensitive environments
The design is still in an early research stage, and the team says that the next step is to improve the body shape to increase swimming speed and efficiency, but the researchers believe that thin hybrid biorobots may be suitable in the future for tasks that require soft and small devices within environments that traditional robots are not suitable for, such as monitoring some sensitive aquatic environments or exploring fragile areas that cannot tolerate hard or large devices. The study was published in the journal Advanced Functional Materials, while the project received partial support from the US Office of Naval Research.
