Blog of Florian Dietz

Blog entries

  • Achieving AI alignment through deliberate uncertainty in multiagent systems 2020-06-10

    The problem of AI alignment is one of the most important questions we need to answer to safeguard humanity's future. How do you ensure that an Artificial General Intelligence will behave ethically?

    I outline a general approach to achieve this goal that counterintuitively relies on confusing the AI on purpose.

  • Interaction Networks: Using a Reinforcement Learner to train other Machine Learning algorithms 2020-05-21

    The wiring of neurons in the brain is more flexible than the wiring of connections in contemporary artificial neural networks. It is possible that this extra flexibility is important for efficient problem solving and learning.

    This paper introduces the Interaction Network. Interaction Networks aim to capture some of this extra flexibility.

    (I wrote this paper as an independent researcher, outside of university, which is why I'm publishing it on my blog.)

  • The Brain as a Reinforcement Learning Algorithm 2020-04-04

    (This article tries to make concepts from machine learning digestible to non-technical readers. I try to keep things non-technical.)

    I am continually amazed by how often I find analogies between my work as an AI researcher, and my own thoughts. When working on an AI algorithm, nothing is a more promising sign than the realization that my own brain has the same behavior as the algorithm I am developing.

    This is especially true when the behavior is irrational. If both the brain and an AI make the same types of mistakes, it is a valuable learning experience, and it is also a sign that you are on the right track.

    I would like to share some of my discoveries with others. They have helped me immensely with introspection and meditation of all kinds, and I hope they will be of use to other people as well.

  • Automation in Data Science 2019-07-08

    Why the most technical parts of my work keep getting easier, and the most irreplaceable parts have nothing at all to do with AI.

  • A Data Privacy Startup 2019-05-17

    The GDPR has created a lot of obligations for the proper handling of user data. This is an opportunity for individuals to take ownership of their data.

    I describe an idea for a very profitable startup that helps with this process, and also allows you to make money from your private data in the process.

  • Self-assessing neural networks 2019-05-07

    I invented a new type of layer for neural networks: It gives the network the ability to critically assess the reliability of its own features, to enable more informed decision making.

    Preliminary results: Use of this layer lead to a minor improvement on the MNIST dataset, but the improvement is too small to know for sure if it is useful.

    Even if it turns out that this improvement is a fluke, there are a large number of future improvements that could make the algorithm more effective.

    Addendum (2019-05-18): After testing this idea on a few more datasets and architectures, I did not notice any consistent improvements. The SelfAssessmentLayer sometimes improved and sometimes hurt performance, and I could not detect any pattern in this. It appears that the SelfAssessmentLayer is not as useful as I expected because the information it provides is sometimes misleading and/or irrelevant and the neural network has no way of knowing when this is the case.

  • Quantified Ontology: The probability that we live in a simulation 2019-03-09

    What is the probability that we live in a simulation?

    TLDR: I don't know, and neither does anyone else. There is important information that we don't have access to that no amount of introspection will give us.

  • Tracking social circles and measuring civilization 2019-02-28

    This article is basically an attempt to build a primitive but working model of Psychohistory using contemporary AI techniques.

  • Realistic and interesting aliens 2019-01-23

    Aliens in science fiction are usually very similar to humans. Often they are basically just humans with a minor cosmetic difference and a culture that is actually closer to modern-day western societies than many historical human societies were. If we were to meet life that evolved on other planets, they would certainly be far more alien than that.

    Most authors find it difficult to imagine non-human intelligences without bandying around words like 'incomprehensible', for the simple reason that they are authors: Their trade is writing stories that are interesting to humans. Why should we expect them to understand non-human intelligences, or the intricacies of evolution?

    In contrast, I am an AI researcher. I have no talent for writing captivating stories, but I frequently come up with concepts for truly alien species, because understanding these concepts is important for my research, and because designing a fictional species is fun.

    Here is a list of non-human aliens I have made up that are particularly interesting:

    • Davlash: A species with a fundamentally different method of reproduction.

    • Nophilians: A species of savants.

    • Xeltek: A benevolent pseudo-hive mind.

    • Benefactors: An artifically created species that helps others.

    • Sino: A species that achieved a technological singularity, and accidentally made life very interesting for every other species.

    • Boros: Not a species, but a civilization of humanoids. However, their technology leads to a drastically different culture compared to contemporary western civilizations.

  • The copycat effect and responsible media 2019-01-04

    The copycat effect is the tendency of sensational publicity about violent murders or suicides to result in more of the same through imitation.

    Through the copycat effect, both media companies and private individuals with strong networks can be indirectly responsible for things as terrible as suicide, school shootings, and even terrorism.

    The principle of free speech is too important to force the media to be quiet about such topics, but their reliance on sensationalism to make sales is costing lives. What then can realistically be done about this?

  • Nature vs Nurture: In favor of nurture, because nature is both powerful and lazy 2018-12-30

    Evolution is powerful and human behavior is determined mostly by nature. But evolution is also lazy and works in mysterious ways. Ironically, this means that we end up with many of our traits determined by nurture.

  • What is love, technically? 2018-12-30

    During my AI research, I noticed an algorithmic phenomenon that very closely resembles what we call 'love' in everyday conversation.

    While most definitions of (aspects of) love, such as "oxytocin" are kind of depressing, this one is actually inspiring and romantic, which is unusual for scientific theories to say the least

    Note that we actually use the word 'love' to refer to dozens of different concepts. This definition only explains one meaning of the word, not all of them: It explains the kind of deep, true affection people can develop for each other, which stays with them no matter what, even through tragedy.

  • The dead end in neural network research 2018-12-30

    Contemporary Artificial Neural networks are a (very profitable) dead end.

    Hardware acceleration and hyperspecialized algorithms have allowed neural networks to become much more successful than other AI approaches. However, they have also ensured that neural networks are overspecialized. There are some tasks for which they are not suitable at all, but researchers have no incentive to change this.

    If someone were to invent a revolutionary new technique to improve artificial neural networks, if the improvement does not benefit from hardware acceleration, then it won't be adopted. Even though the end result of continuing that line of research would be amazing in the long term, nobody would ever even try because the immediate results just can not compete with the performance of the overspecialized contemporary neural networks.

    In terms that should be clear to AI researchers: If we view AI research itself as an optimization problem, then we are currently stuck in a local optimum. All our research to improve neural networks further only drives us further into the local optimum. The global optimum, which represents a General Artificial Intelligence, will not be achieved unless we change course.

  • Applications of cryptography 2018-12-13

    With the rising popularity of blockchains, cryptographic techniques are slowly starting to be used for practical purposes besides the obvious. There is still a large number of possible applications for these technologies that haven't been explored yet. Here I mention one application that the government could use.

    I'm not publishing the rest of my ideas since I might want to make a startup out of some of them in the future, and also because I have some serious concerns about what might happen if the wrong people implement it.

  • Probabilities, uncertainty and learning 2018-12-13

    Probability values are deceptive. Many people feel intuitively that a raw probability value 'is just math' and does not capture all there is to say on a subject. This sentiment is correct: By giving only a probability value, a lot of relevant information can be glossed over. What is the missing information that is not captured by the probability number, and how can it be quantified?

  • Effective criminal punishment 2018-10-07

    Many crimes carry punishments with them that are in no way useful to our society. Politicians are incentivized to create laws that sound good, not ones that actually address the problem they are supposed to solve.

  • Metrics of economic performance 2018-10-07

    This article is basically a rant against popular ways of measuring economic performance, and when they go wrong.

    This is important because measures tend to get used as metrics when enough people believe in them, and flawed metrics lead to bad decisions.

    (Nothing here should be new to someone who studies economics professionally, but far too many non-experts make these mistakes even when they are otherwise smart people.)

  • Why is there no word for this? 2018-10-07

    I find myself wanting to use the following concept in conversation very often. There should be a word for it:

    "while taking into account both the [probability/frequency/amount] and the [effect size]"

  • A theory of humor 2018-09-13

    Lots of people have tried to give a scientific explanation of humor, but I have found that all the explanations I have read so far are missing something and fail to account for one oddity or another.

    Most theories of humor try to pare everything down to only a single overarching explanation for everything. While it would certainly be neat if that was possible, nature is messy and evolution does not usually develop a single solution to deal with a single problem.

    This is my own theory, inspired by my research into the way the brain works. It is similar to the incongruity-resolution theory, but with an added evolutionary basis.

    I call it FDIS for "flaw detection and ingroup signalling". This is because I believe that humor serves more than one purpose and those two are the most important ones.

  • A Graph-based Explanation of Physics 2018-06-15

    A short exploration of a different way of looking at physics.

    It is not a theory, because it does not make predictions. It is a different way of looking at things.

    I have found that this made a lot of things we normally consider weird a lot easier to understand. Both quantum physics and general relativity make intuitive sense in this framework, while other everyday aspects of life become more complex instead.

    I am not a physicist myself, but I have been told by several physicists that the idea is interesting and may be worth exploring.

    Hopefully someone with the necessary background in physics will take a look at this and derive some value from it.

  • The Accessible Mind 2018-06-08

    I expect that the technology necessary to accurately detect lies will become available in the next couple of decades.

    The impact of such a technology on all aspects of life would be enormous.

  • FeedbackFunding 2018-03-26

    A financial instrument intended for fast-growing startups to get funding during their growth phase.

    I was thinking about the best way to get funding for my startup and none of the existing systems were a perfect fit. So I invented my own.

    Feedback has been overwhelmingly positive.

    Unfortunately, actually implementing this system would be a lot of work for me and I am still busy with my startup itself. I also have no idea about the legalities involved in creating a new system for funding a startup.

    Hopefully I can convince someone else to implement the system, then sign up with them as a customer. It would be very easy for an established bank or crowdfunding platform to offer this service.

    I call this financial instrument FeedbackFunding because it is a way to get funding in a continuous fashion, so that the startup can use earlier investments to improve itself and thereby convince people to invest more, in a feedback loop.

  • Enslaving Gods and Granting Wishes 2018-03-26

    I read about an outside-the-box solution to the Hardest Logic Puzzle Ever and took it as a challenge.

    I came up with an even better solution to the problem, which doesn't just solve the problem, but also mind-controls a god as a side-effect, giving you the ability to have arbitrary wishes granted.

About me

  • I am the founder of the startup, which aims to make solving problems as easy as google made answering questions, by combining the programs of many developers in a gestalt AI.

  • I have an interactive CV.

    I built this thing when I first taught myself html. This CV is very detailed, but only displays details when you zoom in. This way, you can choose to look at each topic in exactly as much depth as you want, and skip the parts that don't interest you.

    The presentation style was supposed to be the best of both worlds: Give a quick overview and also provide full details about any topic the reader cares about.

    Unfortunately, the fact that nobody else uses this presentation style means that people get confused and have to take a minute to learn the controls, which defeats the whole point.

    If people were actually used to this way of presenting information, I am convinced it would be very effective and popular. As it is, it's kind of pointless, though.

  • Contact: