We welcome everyone! Much time has passed since the publication of the previous sprint. We decided that it would be correct to focus on work and publish articles less often, but more voluminous.
What happened during this time?
We have completed the work on two layers of the core - the
Interlocutor layer and the
Interlocutor layer is used to create a model of the interlocutor in the core. It will help the core have some idea of the interlocutor to be used in the dialogue. In other words, the AI will be able “put itself” in the place of the interlocutor in the dialogue.
Logical layer will allow the core to produce various kinds of conclusions and to construct complex statements. The layer is technically implemented and will be gradually filled.
Work with text
Added error correction and typos. It is already included in the demo and will be gradually developed.
We have completely reworked the text conversion modules into the internal representation of the core and back. The second version of the modules is not included in the demo yet, but the update soon. These modules will significantly expand the variability of questions that the AI will be able to understand and allow it to build more complex and properly constructed sentences.
We also began work on converting speech to text. We solve this task in the traditional way - using neural networks, but our main goal is to create an interface that can work autonomously on weak hardware. And now we have achieved a big shift in this direction. The test build started working and shows good results even on a small training set. In March we will allocate more resources for this task.
Finally, the first iteration of the automatic knowledge input mechanism was launched. At the beginning of the process, we focused on geography, since this is one of the main parts of our life and a topic that will allow us to discover and demonstrate the current capabilities of the core. For a couple of weeks we taught the core to such a volume of knowledge that it had mastered a year before. And this is just the beginning. However, we are still not in a hurry to fill the core with large amounts of knowledge, since it is important for us that in the process of development the volume of knowledge and intellectual abilities grow evenly.
Currently, Steve has knowledge of geography, namely:
- countries (capital, area, population, where it is located);
- cities (area, population, country in which it is located);
- oceans (maximum depth, volume, average depth, where it is located);
- seas (maximum depth, where it is located, area);
- islands (area, country);
- rivers (length, where it falls, in which countries it is located);
- lakes (area, maximum depth, volume, which country is located);
- deserts(area, in which countries located);
- mountain ranges (where it is located, area, width, length, what mountains are in them);
- mountains (maountain range, height, country);
- volcanoes (height, country, eruption).
A complete list of features can be found on the demo site.
We have developed a plan for further study and will begin work on this plan in February. Below is a graph of the growth of core knowledge:
We continue to expand the possibilities of automatic learning. Much attention has been paid to spiders and knowledge extractors. A system of spiders is being developed - which will be able to collect data from various open sources. Knowledge extractors transform the information that spiders collect into a representation that is suitable for automatic learning.
Once again, we reworked the design of the demo site. We have laid the groundwork for further development (global interface updates are planned in spring). Soon Steve will be able not only to answer questions, but also independently build a dialogue. To do this, AI must be able to distinguish interlocutors. Therefore, now communication with Steve is available after registration.
Recently, we have been actively working on memory mechanisms. Now it is used minimally on the demo, but after the implementation of the dialog layer, several layers of memory will turn on at once.
Stay in touch and see you later!