AI is not a black box (talking clearly)

Reader, is a black box. Your mind is not mysterious. I don't know how you think. I don't know what to do and I don't know whether your words are trustworthy and for a reason. We learn to understand and trust people in many years of understanding and interaction with others. But the experience also tells us that understanding is restricted to those who have the same trust backgrounds and trust is not required for those who have motives against ours.

Artificial intelligence – while you are strange-it is a crystal crystal clear compared to. I can investigate ai equal to its thoughts and motions and I know that I find the truth. In addition, AI equal to “Life Background”, and its “motivation”, its goal of training, especially if you do not look proper and open. While we are still scarce years in the modern Ai program, I argue that there is no end problem; On the contrary, the clarity of AI programs should be tested, “white word” white “nature, can be the basis for understanding and confidence.

You may have heard about AI 'like a black box' in two days: Ai as ChatGPT or anthropic's Claude with black boxes because you cannot check their code or parameters (black box (black box (black box (black box access). In a general sense, whether you can examine those things (white Box access), either with little help in understanding how AI works in any ordinary estimates. You can follow all the instructions that describe Chatgpt and do not get more understanding than it reads out, corollarry in a Chinese-room conflict. The mind (human), however, more opaque is more than preventing AI. Since the physical organs of behavior reduces the investigation of our human thought and models and the elements are not complete, the mind of the person is “natural”.

Since 2025, the only Nerural nerual structures are displayed – those flies – have small part of the human brain. Actually, experiments using the Magnetic Resonance Mathematical performance (FMRI) can do neural work down until 1mm⁺ Volume of brain story. Figure 2 shows an instance of neural structure held as part of the FMRI course. The necessary hardware includes at least $ 200,000, strong access to the liquid helium, and the provision of patients are willing to hold Tonononductor Spins inchins from their heads. While FMRI studies can find that, for example, the processing of visual and housing symbols are associated with the brain districts, many of our brain works grateful for real-life, unpleasant risks. Conductive assessment methods, providing low-to-to-sound scales.

Open source models (white box access), including large and added and ignored. Every small one of all neurural connections can be identified and login again and below the main space of installation. AI does not care about the process, and is not affected in any way. This rate of access, control, and repeats allows us to remove a large sign of signal where we can make a good shaped analysis. Controlling what AI sees helps us connect the common ideas into parts and processes inside and outside AI in useful ways:

• Associating with neural work in concepts like FMRI. We can say that AI “thinks” in a sense. How well can we say when someone thinks in a sense? Figs. 1 and 3 Two translations of the ideas from Gemmascope provides the Gemma2 Model Syodeleters of Google 21 model to the meetings.

• Determine the importance of something output. We can say that some part-time part was important for the production of Ai. Can we say what a person's decision is affected?

• Shipment of concepts such as athletic. This means that we can say exactly what on the neural network of concept go from installation words to the end. FIG 4 shows an example of such a way that followed the idea of ​​the language program for the title agreement. Can we do the same for people?

Of course, people call answers to the first two questions. You can ask for the employment manager what they think when learning your Réméé any important things in their decision to give you a task (or not). Unfortunately, people lie, themselves do not know the reasons for their actions, or prejudice in ways they do. While this is also true of the Generative Ai, translation methods in the area of ​​AI do not rely on AW awi's answers, true, cannot be seen, knowing, or otherwise. We don't have to trust the AI ​​to tell if it is thinking about a sense. We literally read the Virtual Probe attached to their neurons. With an open source model, this is a small, funny thing so looking for what is needed to get this kind of information (rightly) to the person.

What about the closed source of the “Black Box” Ai? More can only be excluded from the Black Box only. The models of models is known, and so is their normal construction. Their basic elements are normal. They can also be reviewed at a higher rate than the people who were to be tolerated, and in a controlled and recurring way. Repetition under the selected installation is usually replaced. Parts of models can be included or given their semantes copied by the “DISTILLATION”. So the Black-Box is not a perfect sense of understanding and trust, but the fastest way to make a very obvious AI.

People are likely to be more sophisticated thinking equipment, and therefore the above comparisons may not be visible. And we are inclined to manifest our understanding and trusting people because of our age of experience and interacting with others (disputed). Our experience with a variety of AIS is growing fast, and their skills. While the most effective model sizes grow, their common buildings have been stable. Nothing shows that we will lose the nature of its obvious workforce, just as they receive and thereafter greater than the skills. Nothing indicates that testing human brain may have shown a success that is important enough to give you little opaque intelligence. AI is not – and may not be – a black box that the person is the same feeling that they say.

Piotr Mardeziel, Head Ai, Realmlabs.Ai.

Sauradi Merow and Sauradh Shrantte contributed to this post.