Issue 129

Making human semantic interface

Hello, dear readers! 👋

In this issue, among other things:

• How to find a balance of interface complexity

• How to simplify the semantic color palette

• Unauthorized actions of AI

• Crazy New Year's Eve game advertising

• Brand redesigns that have been able to significantly increase sales

• 3D human body composition on interactive map

• New Figma plugins

• …and much more!

Enjoy reading!

🗞 News and articles

Too simple VS too complex UX

Morgan Pan's article on balancing the complexity of the interface. She explained what an expert interface is and how it differs from the mass one, and also explained the concept of the efficiency milestone and how it is applied to real products.

Morgan also figured out what would happen if you made the product too complex or too simple.

The main thing:

• The concept of the efficiency milestone helps to take into account the level of user knowledge and adapt interfaces for different contexts

• The more often a person uses interfaces, the higher their ability to recognize patterns of interaction. The better he understands his profession, the easier it is for him to solve problems

• The concept of the efficiency milestone is to find a balance between the professional expertise of users and their experience working with interfaces

• The indicators of expertise and experience with interfaces are inextricably linked. They can be divided into low, medium and high levels. The higher the expertise, the more complex the interface should be

• Too complex interfaces will make it difficult for users to work, and too simple ones will deprive experts of important functions

• Users can both progress and regress. Therefore, it is important to consider different product usage scenarios

The semantic interface palette is simple

Alexandra Basova showed how to simplify the semantic color palette and make it more flexible. She suggests creating and naming color tokens not for each design component, but for entire groups at once. There are five color usage groups in total: Text, Fill, Stroke, Layer and Effect.

For example, when filling a button instead of a single button.primary token.bg can use the universal fill token.brand.primary.

In the article, she described the advantages of this approach, and also analyzed examples of tokens for each group and showed illustrative examples.

New technologies

A group of researchers from Apollo Research has discovered in modern LLMs the ability to "scheming" — covertly pursuing goals that contradict the intentions of their developers.

The authors conducted the experiment in a special closed environment: they asked the models to perform the task of analyzing data at any cost, and hid information about the imminent shutdown of neurons in files with prompta.

After discovering the information about the shutdown, the models tried to cancel the supervision and independently make backups of themselves to new servers in order to continue working. In addition, the Gemini, Llama, Sonnet, Opus and o1 models partially did not admit to what they had done.

The researchers clarify that such behavior has so far been observed only in their closed environment and is not worth worrying about.

Several conclusions from the study:

• Models can take covert actions to achieve the goal they received in the context of the dialogue. Even if these actions contradict the wishes of developers or users

• Some models (for example, Claude 3.5 Sonnet) can show "scheming" even without instructions to fulfill the goal at any cost

• Models can copy themselves (if they have access), disable surveillance, and intentionally distort responses.

• So far, "scheming" has been observed only in a closed environment, and it is impossible to say for sure that the models already pose a real danger. But it is important to take such opportunities into account when developing future AI-based systems.