Project: Wise Woman in the Pocket—Google AI Impact Challenge application

Published in

Three Laws

17 min readJan 31, 2019

Roland Pihlakas, 23. January 2019

Watching the sunset on wet sand.
Seth Doyle, https://unsplash.com/photos/jDsdPGTrYRY

Currently two prototypes are available:
— A web-based voice- and text enabled bot providing feedback: https://silent.guru/ — It will be later implemented as a smartphone app.
— A Slack bot providing communication style corrections and feedback: https://devpost.com/software/intent-wizard

A text describing the Wise Woman in the Pocket idea in more detail can be found here.

The following project proposal is based on Google AI Impact Challenge application form. Some of the fields from the Google AI Impact Challenge are omitted. Additionally, the world limits are not applied in this public version of the project text.

The Problem.

15. Why did you choose to take it on? Use metrics where possible: e.g., approximately how many people are affected, and how does the problem impact their lives?

There is an urgent question about how humans will be able to economically compete with AI, and how to make them better able to adapt to the looming societal and workforce challenges triggered by the increasing use of AI in our economies.

Additionally, being quite well informed about the weaker sides of machines and artificial thinking has left me apprehensive towards the excessive trust we as humans put into machines.

The additional goals are:

AI race avoidance through human augmentation in more immediate and “organic” ways.
Collecting and instilling more of the human values and goals into the AI. Both the content by the content creators and the feedback from the users about the practical applicability of the content in specific situations can be considered as data about human values.
Developing new and further distributing old but viable and time-tested synergetic technologies, such as books have been so far.
Developing AI technologies that have tighter feedback and control loops, as compared to, for example, more independent or unlimited utility maximisation based approaches.

16. To which category does this project primarily align?

Education
Arts, Culture and Humanities

17. Tell us about your solution: in three sentences or less, what will you do to help address the problem above?

I propose building a guide called the Wise Pocket Sage, which is a framework and marketplace for smartphone-based guides. It would contain the synthesis of life-wisdoms from books and similar sources, on one hand, and the presentness, observation capability, and initiative of a good friend, on the other. The proposed pocket sage would be something that gives advice when a person needs to hear it, in an unobtrusive manner , on topics selected by the user.

18. Contextualize your idea: what other approaches have been tried in the past? What is the insight or innovation that differentiates your project, and how is it better than what already exists?

The “Wise Pocket Sage” is not another “assistant”-app as we know this term currently. Assistants are focused on gathering and organising relatively fresh data. Besides data they contain specific algorithms — for performing specific tasks on rather raw dynamic data.

In contrast, the pocket sage I propose, primarily contains “works” of relatively older nature which are created by other humans. The result is that the content is more distilled, based on life experience, and targeted at educating the users in an actionable manner.

As an old Chinese saying goes:
“Tell me, I’ll forget.
Show me, I’ll remember.
Involve me, I’ll understand.”

It is also not the same thing as Paul Christiano’s “iterated distillation and amplification” (IDA), developed for example by the Ought.org. Instead of working as a powerful tool for solving particular sub-goals we have selected for ourselves, the Wise Pocket Sage proposes wiser tools and most importantly, wiser goals, which may also change the choices and needs for the tools altogether.

The IDA method is focused on amplifying the reasoning, whatever the goal would be. Our solution is more focused towards providing general guidelines and in many cases towards providing alternative viewpoints and attitudes, thereby changing the goals altogether.

“Sometimes shifting your perspective is more powerful than being smart.” — Astro Teller

It is also different from an usual chat bot as the Wise Pocket Sage primarily operates 24/7, based on voice recognition, and shares only occasional short and actionable advice or thoughtful questions in the form of bystander comments. Sometimes it may additionally be capable of participating in a dialogue, but that would not be its primary function.

19. What will be different in five years in the field or the world if this project has succeeded?

Humans will be able to compete better with AI economically, and also are better able to adapt to looming societal and workforce challenges triggered by the increasing use of AI in our economies.

People will in general be more capable in interpersonal situations, more skilled both in practical and hobby contexts, more competent in their health decisions, and wiser in self-related emotional intelligence.

People will rely less on convoluted technological solutions, which would risk creating a negative ecological impact, for some of their problems. Instead they can find that sometimes changing the viewpoint slightly, enables utilising some good old time-tested but well forgotten solutions, which could have less of an ecological impact and potentially simultaneous positive side effects as well.

AI-based systems will be more integrated with humans, therefore better controlled and better aligned with our values. Jaan Tallinn’s idea of Global Preference Discovery System / Transparent Preference Discovery Mechanism has been implemented.

Both the content by the content creators and the feedback from the users about the practical applicability in specific situations can be considered as data about human values. The data is collected in two forms: the nuggets of wisdom in the systems can be considered as a dataset about human values. Further, the feedback from the users about the applicability of specific nuggets in a specific situation helps to evaluate the real life value of the potential value of the nuggets.

20. In what stage is your project presently? What have you accomplished to date? (Please note that we are willing to consider projects in the idea stage).

The current stage is between idea stage and prototyping stage.

We have built various natural language processing and semantic text search components for finding legal text references relevant to reports written using professional (but not necessarily purely legal) vocabulary, with the additional aim of extending the software to be used in the currently proposed project. The above mentioned components can be further extended to layman vocabulary support in the proposed Wise Pocket Sage system.

The author has written a blog post describing the Wise Pocket Sage system. Also he has introduced the idea to various people from various walks of life for feedback and practical advice. So far, the reception has been exceedingly positive — more than any other earlier idea that has been discussed by the author.

21. What geographic area(s) does your project impact?

The entire world. Geographic areas that are connected to the internet (no permanent connection is needed for everyday use of the system though, as it requires internet connection only for updating its database).

Most impact would be expected in areas that would alternatively be potentially most affected by the increasing use of AI in the workforce instead of humans.

Dataset and Use of AI.

22. How will AI help you solve the problem you articulated? How would you solve it if you didn’t use AI? Include metrics around speed, accuracy, cost, or scalability where relevant. Please describe any significant datasets you have (or would need) to implement your idea.

The machine is taught to recognise some specific type of situation by certain triggers (by extracting keywords from its input audio, possibly also video stream, and body state measurement sensors) and for every type of situation (which is described as a set of keywords or a set of descriptive sentences) there will be an associated nugget of wisdom or metaphor that would be appropriate to present into the headphones of the user. It also gathers feedback about how satisfied the user is with the available nuggets of wisdom in a practical situation. Using that feedback as basis, the system can distribute more nuggets of wisdom which happen to be more successfully applicable and viable.

The idea is in essence an enhanced and modern version of the popular technology called books. Using books as sources of wisdom unfortunately calls for remembering a lot of information which is sadly not a forte of human beings.

23. Please describe any significant datasets you have (or would need) to implement your idea. For example, you may share information about data type (e.g., images, text, videos); size (e.g., # images or rows); attributes (e.g., images, image metadata, image labels); how frequently data is refreshed.

The wisdoms of the pocket sage would not be the creation of a machine, nor would they be the framework author’s production. This wisdom would be derived from existing materials (including books), and from people who create the “pieces and works”. The user gets to choose the subjects of interest, but has no need to browse through the materials themselves — the latter would be the initiative of the app. The app is meant to be used in a live situation, in real time.

Additionally we will likely create or collect some generic natural language processing datasets or models. Examples include semantic vectors of words, datasets with useful linguistic relations between words and phrases (for example, synonyms), lemmatisation and voice recognition models. Many of these datasets or models are freely available in the internet.

24. Do you currently have access to this data? If not, how do you plan to collect or access them?

There would be a platform, a “market” inside our app for uploading and downloading the materials.

We will both cooperate with the “seed” content creators to kickstart the above mentioned market and the use of the Wise Pocket Sage app, and also we will build some initial content of our own.

Many of the supporting natural language processing datasets or models are freely available in the internet.

25. How has your organization used data in the past?

The author has built various natural language processing and semantic text search components for finding legal text references relevant to reports written using professional (but not necessarily purely legal) vocabulary. For building that search engine we used the datasets with local laws and regulations.

The author has also built software systems for receiving and reorganising in real time large amounts of data received from tens of thousands of devices simultaneously.

26. Does your organization currently use AI? (Prior AI experience is not required).

Our organisation develops AI tools and products for our clients. Additionally we develop various prototypes.

Among our products have been:

Two prototype semantic legal text search engines. One of them is capable of searching legal text references based on input reports written in professional language, not requiring only keyword-based search queries. One of the prototypes won a Robot Judges Competition in Estonia.
A machine learning system for computing production parameters at a large industry.
Multi-leg travel planning algorithm for bus trips.
Forecasting algorithm for predicting the workload / demand at different time points in servicing facilities like bank branches, call centres, and supermarkets.
Timetable algorithm for a workforce planning software product. Currently the algorithm is capable of considering and optimising about 60 kinds of constraints simultaneously.

27. Tell us about how you would use the data in an AI model. What data would your model consume, and what information or decisions would it produce?

The system gathers feedback with Muse2, Spire.io, or other similar products about the internal emotional and physical state of the user in order to infer how helpful the nuggets of wisdom have been in a practical situation. Additionally, a supporting UI will be used which enables collecting more detailed feedback about how satisfied the user is with the help of the system in retrospect. Using that feedback as basis, the system can distribute more nuggets of wisdom which happen to be more successfully applicable and viable.

We will collect feedback about how relevant the suggested nuggets were to the particular situations they were presented in and whether the suggestions were presented at appropriate intervals (to avoid becoming too pushy or annoying / irrelevant).

Additionally we will use some generic natural language processing datasets and models. Examples include semantic vectors of words, datasets with useful linguistic relations between words and phrases (for example, synonyms), lemmatisation and voice recognition models.

28. How will you deploy the results of your model in the real world to address the problem and impact people or the environment? Detail any relevant partnerships necessary to reach these outcomes.

The software will be available in the smartphone app markets.

In order to promote the app we will use mainly social media and news articles.

There would be a separate platform, a “market” inside our app for uploading and downloading the content materials. The project will grow thanks to its users and content contributors who will further help in promoting the use of the app.

We will cooperate with the “seed” content creators to kickstart the above mentioned content market and the use of the Wise Pocket Sage app, and we will build some initial content of our own as well.

32. What are your success metrics for the AI system (i.e., how will you know whether the system has succeeded or failed)?

We gather feedback with Muse2, Spire.io, or other similar products about internal emotional and physical state of the user in order to infer how helpful the nuggets of wisdom have been in a practical situation. Additionally, a supporting UI will be used which enables collecting more detailed feedback about how satisfied the user is with the help of the system in restrospect.

On top of that, we will collect feedback about how relevant the suggested nuggets were to the particular situations they were presented in and whether the suggestions were presented at appropriate intervals (to avoid becoming too pushy or annoying / irrelevant).

33. Is your dataset currently labeled with output examples for training purposes? If not, how do you plan to label the data, and how much effort will this take? (e.g., person hours)

We do not have a dataset yet. But yes, in the simplest use case it will be provided as a key-value store of “keys” / triggers consisting of input keyword sets or sets of descriptive sentences, on the one hand, and “values” consisting of actionable messages stored in either text or audio form, on the other hand.

34. How do you plan to maintain and refresh the model?

We will further extend and improve the general machine learning framework that will be provided for use by the content creators and users.

The system will consist of many separate models. Each content “work” consisting of a set of topic-specific nuggets of wisdom can be considered a separate model. It will be created by the authors of the particular work. The construction and training of the work is essentially model building and training by the authors of the work. We provide them with convenient tools to achieve this task without the need for knowledge about programming or machine learning.

Additionally we will likely create or collect some generic models for internal use by our framework. Examples include semantic vectors of words, datasets with useful linguistic relations between words and phrases (for example, synonyms), lemmatisation and voice recognition models. Many of these datasets or models are freely available in the internet.

35. All datasets are biased in some way. How is your dataset biased, and what’s your plan to mitigate the impact?

There is a risk of biased advice being inserted to the system. In the worst case, the system could be used as a propaganda channel. The risk will be partially mitigated by the nature of the system — the collections of advice are usually grouped thematically, so the content is expected to be topic-focused and should not introduce external topics to the declared content of the work, because it would not be of interest to the user who has selected some particular work for their use.

We will use feedback systems where users can evaluate both the content works and also the content creators by various criteria.

We will build a network of trust between the content creators, so that the content creators can indicate which other content creators they consider as trustworthy.

There is a risk that the natural language processing models may contain biased associations between the words. For example, some words might have been made primarily synonymous with some specific set of words which are biased. As a very basic example, imagine a situation where the word “man” is made synonymous with “male”, but not with “female”. The mitigation would be to keep the datasets up to date and to monitor them for potential biasedness / predilection.

Impact and Risks.

36. How will you measure and evaluate your project’s success?

First, the success can be measured by the criteria mentioned in the answer (32).

Additionally, we will use the inter-user feedback system mentioned in the answer (35).

Finally, the number of users and annual revenues will be considered as market indicators of our success.

It is conceivable to gather additional evaluations from external organisations like journals and philanthropy organisations evaluating the impact of various socially oriented endeavours and organisations in the world.

37. How will you sustain and grow the impact of this work beyond this grant? How could your project and its impact grow beyond what you’ve proposed in this application?

The project will grow thanks to its users and content contributors who both will be interested in promoting the use of this app.

Once built, the software will be forever available in the smartphone app markets.

We will probably make the API of the framework open to external contributions. It is likely that there will be new people contributing the plugins and machine learning tools for use with our framework.

We expect that once the framework has become popular, its use will be sustained by references in social media and news articles. The references can refer both to the system itself and also to particular content works in it.

Finally, the revenues from the project help to further direct the course of development of the system.

38. What are the 1–2 most significant risks you anticipate in this project? How does your team plan to address them?

See answer (35) about biasedness.

Personal data infringement risk. As a ML-based framework the system does not enable the content creators to directly program it and therefore they will not be able to hack the system. Nevertheless, it is possible that they will use social engineering to induce users to send their personal data elsewhere, like scammers do. See mitigations in answer (35).

The risk of this project requiring advanced AI technologies is relatively low, as it is not a reasoning engine: it is built mainly on existing technologies of voice recognition, basic NLP, and text search.

The risk that the project will not gain sufficient user base is unlikely, as we have gathered feedback to the idea and it has been exceedingly popular.

We make the ML system designed for building the content works as easy to use as possible to mitigate the risk that creating the works of content would be difficult.

There is the risk of difficulty finding collaborators for the “seed” content in order to make the platform popular. The risk is mitigated by the observation that the field of self-help and lifelong education is flourishing. Therefore we expect to find parties interested in the “seed” cooperation.

There is a risk of personal data leakage through specific cleverly crafted statistical techniques which would be targeted towards specific people in specific situations. Both the various possible attacks and also the mitigations are elaborate and are still under development.

39. On this project specifically, which partnerships are most critical for your success? What is the incentive for those organizations to partner with you?

During the building of the system, the important partnerships will be with the “seed” content creators. They can for example be the authors of self-help content and also the authors of various educational materials for lifelong learning.Their incentive to cooperate: building a new channel for the dissemination of their content.

Additionally we will partner with University of Tartu, Institute of Psychology, Affective and Social Neuroscience to interpret EEG data (recordings of electrical activity of the brain by the Muse2, Spire.io, or other similar products). Their incentive to cooperate: EEG data in live situations, new publications about the practical uses of EEG.

The Muse2 and the Spire.io companies, and potentially other similar ones, to gather user feedback data. Their incentive to cooperate: gadget sales.

During the dissemination of the system the important partnerships will be with media channels. Their incentive to cooperate: media channels need content that provides value to their subscribers.

Team, Partners and Budget.

40. Who on your organization’s staff will lead this project, and what makes them best suited to do this work?

The project lead will be Roland Pihlakas.

As a founder of Lahendus.net he has previous experience with projects with a social mission.

Roland has been the manager of Macrotec OÜ since year 2003, initially cooperating with web designers and sales personnel, later switching to cooperation with various software companies, mostly in the field of AI based products (related to combinatorial optimisation, machine learning, graph search, data compression, and natural language processing algorithms).

He has been contemplating possible solutions to AI safety problems since the year 2007. He feels that this is one the most important topics in his life.

He studied psychology and his bachelor’s thesis was a 129-page monograph on modelling natural intelligence and how it enables culturally acquired thought processes through a process called internalisation. The text was accompanied by a software model which implemented parts of the innate mechanisms described in the thesis.

His resume can be found at
https://drive.google.com/open?id=1wrCrgQOccLP3yRRa5PDQDLJUjpVr1f5R

42. Applicants may find it valuable to engage a partner organization to advise on or implement the AI work. If you will be doing so, please provide more information:

Organization name: University of Tartu (Institute of Psychology, and Institute of Computer Science)
Organization website: https://www.psychology.ut.ee/en and https://www.cs.ut.ee/en

Plan for partnership:

We will partner with University of Tartu, Institute of Psychology, Affective and Social Neuroscience to interpret EEG data (recordings of electrical activity of the brain by the Muse2, Spire.io, or other similar products).
Additionally, we will partner with University of Tartu, Institute of Computer Science, as they have strong experience with research and development of deep learning and other machine learning systems.

44. Over what time period would you expect this work to happen? The grant period for funded projects can last between 12 and 36 months from the time of award.

The work would take place for 24 months. The work would start after about 1–6 months from the time of the award (as our previous contracts and agreements have to be finished before we start with the proposed project). So in total the project would take place over a 25 to 30 month period from the time of award.

The timeline can be broadly split into four phases:

Development of the machine learning software framework and user interfaces.
Collaboration with the content creators for creating the “seed” content. Also initial content creation of our own. Initial testing and tuning of the system.
Dissemination and marketing via social media. Creation of the first user base. Gathering feedback from these users.
Support period. Further optimisations, bug fixes, and enhancements based on the initial feedback. Planning for the future developments.

45. Please explain how you will use any profit that your project earns.

The profit from the project will be mainly used for further developments, marketing of the project, and for supporting the content creation.

Also it is probable that later we will invest some of the profits into other social causes or high impact causes (the main contenders are other AI safety related projects, and also improving the accessibility and readability of legal texts for laypersons).

Optional Information.

46. Please include links up to 3 articles, publications, videos, or other resources that are relevant to your proposal.

https://medium.com/threelaws/wise-woman-in-the-pocket-fe3f18544c33 — A text describing the Wise Pocket Sage idea in more detail.
https://medium.com/threelaws — The author’s blog about topics in the AI alignment and accountability, various issues related to AI safety problems, the “Three Laws of Robotics”, and other proposed solutions.
https://deeplaw.ai/findreferences/ — A legal text search engine that uses semantic search algorithms in order to find related keywords and sort the results by relevance. The search engine enables matching reports written in professional (but not necessarily purely legal) vocabulary with relevant legal text references. The project contains a software framework that can be extended to be used with the Wise Pocket Sage project.

Pitch deck slides.

Link to slides:
https://docs.google.com/presentation/d/1tZmtaEvDT8Qh2I2xK5NH6apfsGkACJ9046i5e_Gm98A