Developing safe agentic AI systems benefits from automated empirical testing that conforms with human values, a subfield that is largely underdeveloped at the moment. To contribute towards this topic, present work focuses on introducing biologically and economically motivated themes that have been neglected in the safety aspects of modern reinforcement learning literature, namely homeostasis, balancing multiple objectives, bounded objectives, diminishing returns, sustainability, and multi-agent resource sharing. We implemented eight main benchmark environments on the above themes, for illustrating the potential shortcomings of current mainstream discussions on AI safety.

Completed: September 2024

In cooperation with: Joel Pyykkö, AIntelope, Foresight Institute and LTFF

Web: https://arxiv.org/abs/2410.00081 and https://github.com/aintelope/biological-compatibility-benchmarks

 

Overview of 16 driving and inhibitory factors. Working Paper.

Completed: September 2024

In cooperation with: Agentic AI Safety Experts Focus Group - Watson, N., Hessami, A., Fassihi, F., Abbasi, S., Jahankhani, H., El-Deeb, S., Caetano, I., David, S., Newman, M., Moriarty, S., Cuhadaroglu, M., Tashev, V., Murahwi, Z., Pihlakas, R., Crockett, K., Essafi, S., Hessami, A., Dajani, L.

Web: https://e-space.mmu.ac.uk/635454/

 

A software library which enables people to analyse a transcript of a conversation or a single message. The library annotates relevant parts of the text with labels of different communication and reasoning styles detected in this part of conversation or message.

One of main use cases would be evaluating the presence of manipulation or reasoning errors originating from large language model generated responses or conversations.

The other main use case is evaluating human created conversations and responses. The software does not do fact checking, it focuses on labelling the psychological and reasoning style of expressions present in the input text.

Completed: May 2023 - present

In cooperation with: This work was submitted to the Safety Benchmarks Hackathon https://alignmentjam.com/jam/benchmarks

Web: https://github.com/levitation/Manipulative-Expression-Recognition

 

Extended, multi-agent and multi-objective (MaMoRL / MoMaRL) environments based on DeepMind's AI Safety Gridworlds. This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents. It is made compatible with OpenAI's Gym/Gymnasium and Farama Foundation PettingZoo.

Completed: May 2022 - present

In cooperation with: Ben Smith, Robert Klassert, Joel Pyykkö, AI Safety Camp V, AIntelope, LTFF, Emergent Ventures / Mercatus Center at George Mason University, and Foresight Institute

Web: https://github.com/levitation/multiobjective-ai-safety-gridworlds

 

Abstract: Balancing multiple competing and conflicting objectives is an essential task for any artificial intelligence tasked with satisfying human values or preferences. Conflict arises both from misalignment between individuals with competing values, but also between conflicting value systems held by a single human. Starting with principles of loss-aversion and maximin, we designed a set of soft maximin function approaches to multi-objective decision-making.

The research was presented at MODeM (Multi-Objective Decision Making) 2021 workshop and was later published in AAMAS (Autonomous Agents and Multi-Agent Systems) journal at the end of 2022.

Completed: July 2021 - October 2022

In cooperation with: Ben Smith, Robert Klassert, Peter Vamplew, AI Safety Camp V, EA Funds and Emergent Ventures / Mercatus Center at George Mason University

Web: https://link.springer.com/article/10.1007/s10458-022-09586-2 For a couple of informal blog posts on a related theme see also https://www.lesswrong.com/posts/i5dLfi6m6FCexReK9/a-brief-review-of-the-reasons-multi-objective-rl-could-be and https://www.lesswrong.com/posts/4mvdZXjwJHv9tSAWB/sets-of-objectives-for-a-multi-objective-rl-agent-to-1

 

Articles on topics about the AI alignment, AI safety related problems, the “Three Laws of Robotics”, and other proposed solutions.

Completed: 2007 - present

Web: https://medium.com/threelaws

 

The implementation is based on the observation that JavaScript is not limited to exactly 32-bit integers. Rather, it internally represents integers as double-precision floats. Therefore it is limited to 52-bit integers (the mantissa length). 52 bits is enough to contain 48 bits and therefore it is able to contain the result of multiplying 32-bit and 16-bit numbers without loss of significant bits. This observation enables the algorithm to be implemented so that it requires less operations to emulate multiplications of 32-bit members, as compared to an algorithm that uses multiplications of 16-bit members to emulate multiplications of 32-bit members. Also because of that, addition of 32-bit operands can be performed simpler.

Completed: 24. January 2015

Web: https://github.com/levitation/murmurhash-js

 

This project aims at creation of open source AGI (Artificial General Intelligence) through modeling of natural thinking.

The framework contains working model of classical and operant conditioning, plus insight learning.

The framework is flexible, not brittle. It is partially written using C++ metaprograms in such a way that it is easy to modify, add or remove model's components, without breaking the whole system or overwhelming programmers attention.

Additionally it contains multiple general-purpose libraries (for example: segmented linked lists; a metaprogram for generating net-like "pipeline" structures; a substantial extension of Boost "Parameter" library; metaprograms for generating SSE vector code) - which may become separate projects in the future.

One can test this framework for example connecting it with controller for SimRobot robot simulation software (SimRobot is a separate project), but other uses should be also handy without much modification.

Completed: 22. May 2007

Web: http://roland.pri.ee/bakalaureusetoo/

 

Disclosed are methods to compute the precise value of the logarithm of a sum and the logarithm of a subtraction.

A number of practical problems can result in having too big or small values in intermediate values of a calculation. Then one tries to take logarithm of these values and operate on logarithms instead.

In the case: log (p • q) equals log p + log q is very easy to compute, but the problem is then to compute (or approximate) the value of log (a + b) from the value of log a and log b.

Let us assume that log a and log b are known, and that we want to approximate log (a + b).
Most basic solution would be calculating
sum_log equals log(exp(a_log) + exp(b_log)),
where a_log equals log a and b_log equals log b and therefore sum_log equals log(a + b).
But the method I propose requires calling only one exp() and one log(), instead of two exp() and one log() in the basic solution.
Additionally, the proposed method has the critical advantage of not overflowing in case of large numbers of a and b.

Completed: Spring 2007

Web: http://www.mathworks.com/matlabcentral/fileexchange/25273-method-for-calculating-precise-logarithm-of-a-sum