Navigating option landscapes ✷

Make my own thinking come out of superposition

Mon, 16 Feb 2026 00:00:00 +0000

Intro

Logging some stuff into a micropost, mainly for my own sake and sanity (sasa). Thinking is action in imagined space. That is fine. There is a fundamental difference, yet and still, between imagined space and physical reality. Embodiment, the messiness if microphysics.

There is one fundamental, categorical difference between the imagined and the physical realms I want to highlight. Imagination can sustain multiple option paths (superposition) no problem, physical reality can do so at the quantum level but not at the macroscopic scale, it is the exception rather than the norm.

This means, that putting your thoughts into action by speaking out loud (even to yourself), writing it up, or do something else yet in a way which is more conventionally associated with “action”, like changing your pension plan, going to a protest, helping someone out, forces your superposition bundle of thinking to decide, for once. The point is, it is all not about the external feedback but the effect of action on yourself.

In many cases you can still change your mind later on, but importantly, forcing the collapse of your imagination waves into one single thing for a moment, makes a difference for all your thinking, as it percolates back through your brain. OK, that took longer than I anticipated, intro finito.

Predictive brain presentation

Last week I did a presentation on the “predictive brain” at a work event. There, we have a weekly meeting with the purpose of keeping in sync as a team and we are experimenting with formats of how this can be accomplished in an effective and entertaining way. Currently, the format is, one person picks “anything” they would like to rant about uninterruptedly and then do so. This was my turn, and the “predictive brain” is just the most astonishing thing I learned about in my entire life, and thus my topic.

Why is that? “Predictive brain” means a couple of things. First, it changes who’s got the lead from the world to yourself. Our organism is not a stimulus-response machine, but is actually ahead of the wav(e).

Perception is not a passive taking in of things “as they are objectively”. Perception is active prediction of what is going to happen next, from the micro- to the macro-scale. You must not underestimate how your assumptions change your perception. Testified to by cognitive illusions (eg the perceived size of the moon when near the horizon). And many more, literature to the testimony, see my review of “Surfing uncertainty” book by Andy Clark.

Predictive brain books: Seven and a half lessons about the brain

After my presentation I read “Seven and a half lessons about the brain” by Lisa Feldman Barrett on recommendation by a witness colleague. I can only recommend it, very compact read and it has all the things I cherish, plus some I wasn’t even aware of. For a super-tight summary, let me go through these three items:

Energy budget: I was “brought up” in these things by the concept of “the brain is a motion computer” and what we use for thinking is motion metaphors. Probably true to large extent. Lisa adds a concept on top, saying that the brain is a budgeting machine. What it does is to allocate the body budget. Where do we spend. Looking back, looking ahead, what is important this next second, this next hour, this next week. This is what informs its activity.
The triune brain is fake. The triune brain is the idea, originating from the 1950s roughly, that we have the reptile brain, the limbic brain, and rational control by way of the frontal cortex. It’s crap, all these people work together to create the experience of the self.
And yeah, it is predictions all the way down. One of the seven + 1/2 lessons is about how perception is prediction.

Predictive brain books: Eye and brain

And sure, I started rereading Eye and brain, a 1960s book by Richard Gregory, neuropsychologist. The books is in its 10th or so edition, extremely well informed and on the cutting edge. It undoes a lot of riddles and false understandings about anything from the physiology at the front to what happens at the back. Let me just say one thing: vision without touch is impossible, aka not happening. Realizing that the “image” of the world was projected on the retina upside-down, people were wondering, do babys have to learn this? Nyearp. Maybe, the self-image of touch is upside-down. What is the intrinsic coordinate system of the brain?

We went on to ponder the ongoing rise of fascism and AI slop, how these mindsets go together, and what we gonna do about it. I went into the realization how I have come to hate AI aesthetics, in art, in code, in culture, in business, in thinking. Good reference probably, Karen Hao: Empire of AI. The mindset is disgusting.

OK, I downloaded our twitter archives a while ago, and then what. I’ve been following this “rewilding” narrative put forward by Venkatesh Rao, which I appreciate. The rewilding is a kind of anti-social-media stance, saying well lets put up all these bespoke websites instead of insta posts, to make the web wild, distributed and more independent again. Enabled by vibe coding these fancy one-off websites. To some extent I like the notion, to some other extend I worry deeply about maintainability. Not that social media crap is maintainable, anything else but. Either way, the idea is to publish your twitter archive as a custom web site using vibe coding apprioach.

Asking myself about the next action point, I decided to do “nothing” again, rather get another book and read it, I bought “Chokepoint capitalism” by Rebecca Giblin and Cory Doctorow. Techno-feudalism, right. Work for free to increase Zuck’s or some other useless character’s value, such that they can use these resources to fuel our depression. Keep you posted.

Post mortem

I write this, and last week’s post, without diving into notes and without pasting slop. Last week’s post took me 30 minutes literally, this one was more :sweat:

Some links are missing and could be filled in, hit me up if you can’t find them using a search engine.

Comments

LLMs, coding agents, benchmarks grounding

Mon, 09 Feb 2026 00:00:00 +0000

LLMs, coding agents, benchmarks grounding

Grounded airplanes are airplanes that need to stay on the ground, not fly.

The openclaw x codex 5.2 x claude 4.6 x codex 5.3 avalanche from end of January to beginning of Feb didn’t quite leave me alone, so here goes some self-normalization attempt.

Despite big claims floating around, it is still LLMs all the way down, so there is good reason for skepticism.

There is two parts of the perceived narrative, one is the companies themselves making ungrounded claims based on ungrounded benchmarks. And then there is user claims based on their specific usage profile. On the other hand there is general longstanding critical evaluation, and my own individual experience with coding bots.

Benchmarks are good to have but they are usually far from providing a complete picture. I summarize this to myself as “the smarter the thing to be evaluated, the harder the evaluation”. This is to say, any such evaluation trying to cram the entirety into a single (or a few) numbers, is suspicious at best.

Benchmarks are being gamed (anyone remember Llama 4 release a year ago?). Benchmarks are not only gamed by the human stakeholders, but also by the models under evaluation themselves (cheating). This seems new as of last year, 2025.

The incentive for a vendor is clear, if benchmark number go up, business number go up. Good for them, but

LLM vendors seem to regard benchmarks as marketing tools to increase their valuation, and have noticed that while valuations go up if they report good benchmark performance, they dont go down if obscure academics question the validity of their claims.

Granted. For individual users, they are testing things based on their own specific context, and while things might work well there, it doesn’t say much about how well it might work for my own specific context. The smarter the thing, the harder to evaluate. Repeat.

If LLMs are anything, they are prime stereotyping machines, cutting out the long tail and professing on serving the 1% tasks that get 99% of the popularity. This popularity might be a real world usage frequency, no question. Boilerplate code. Repetive tasks. Either way, whenever I tried to do something less stereotypical, I was met with zero shot failure and had to take the thing by the hand, break it down into small digestible pieces, and work it from there. This is a type of “drift” across the “intelligence” category and I claim the “adaptation” to myself here. This happens across the board, like “we did this, and we did that, and then intelligence rained down on us”. Yeah well, no. You did all these moves in the first place looking at the results of the previous round. Right? Just kidding, but hey.

I can see that there is a point in this one off “throwaway” prototyping approaches or other simple, well defined applications with plenty of precursors (a website). I have not seen maintainability of agent’s code output being evaluated. And again, the stereotyping machine will make good interpolations between minor variations of the same thing, but it will fail without notice on anything that is not in the training distribution. By construction. Period.

This is my personal take and rant, done. Let’s try and broaden the perspective. One well-known problem with benchmarks is, that as soon as they become available, they become available for training. This can lead to something called train / test contamination, and if it occurs, it will lead to exaggerated results. Usually, neither the precise training data nor the benchmark trajectories themselves are disclosed. Will we trust any claim in this situation?

Parties are trying to avoid this type of contamination, but without transparent disclosure, we will have to take their word for it. Having seen the challenges of “knowing” and filtering large datasets I am skeptical of claims about having avoided inclusion of previous benchmarks results. It is not a binary situation of course, and partial leakage is leakage, too. In addition, there is certainly adversarial (state-owned or recreational) players actively try to poison the internet training data.

EDIT: paragraph above adjusted.

EDIT: The picture above is probably a bit too simple. Even if direct contamination can be ruled out for a moment, there is something like second-order contamination. This would mean, that the model is not trained on the current brenchmark but on last year’s benchmark. It has been optimized for a particular and narrow set of evaluations. Synthetic or made-to-order data plays a role in this. A quote,

In recent years, as the financial stakes of benchmark performance have increased, systems have been trained more often directly on benchmarks (“teaching to the test”) or on synthetic data closely resembling them. As a result, model development and data curation are optimized for benchmark success, producing systems that perform well under test conditions but degrade in real-world settings that differ only modestly from them, as shown in recent medical studies, where models remain accurate despite missing key inputs yet become unstable under minor distribution shifts, generating fluent but flawed reasoning (Gu et al. 2025).

The edit is inspired by, and quote taken from, this post https://garymarcus.substack.com/p/rumors-of-agis-arrival-have-been

END EDIT

Contamination seems rather minor though when you can just cheat on your exam. The cheating phenomenon seems to have become large during last year. For completeness, this is also called reward hacking. The issue not only pertains to benchmarking but also to in-the-field use.

Prime example for cheating? ImpossibleBench :) If you report more than 0% on this benchmark, you cheated, I love it https://arxiv.org/pdf/2510.20270v1

Silent but deadly failures https://spectrum.ieee.org/ai-coding-degrades

Do LLM coding benchmarks measure real-world utility? https://ehudreiter.com/2025/01/13/do-llm-coding-benchmarks-measure-real-world-utility/

Do LLMs cheat on benchmarks https://ehudreiter.com/2025/12/08/do-llms-cheat-on-benchmarks/

I mentioned the non-evaluation of maintainability above. Other people have much more fine-grained criticism going here, but that aspect is always included. This one goes beyond coding and touches upon these mistaken (gently said) AGI aspirations. Bottom line: right for the wrong reasons. Check Melanie Mitchell’s NeurIPS 2025 keynote writeup https://aiguide.substack.com/p/on-evaluating-cognitive-capabilities

Finally, crawling the 39C3 recordings, this presentation on attacking coding agents came to meet me going down the popularity ranking, “Agentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents” by Johann Rehberger. Agents have been set up to not trust claims without visiting the link. Agents love to click on links: https://media.ccc.de/v/39c3-agentic-probllms-exploiting-ai-computer-use-and-coding-agents

And even more finally, I got interested in a comparison of genealogical trajectories, like how we went from transistor to integrated circuit, it took 13 years (1947 to 1960) https://en.wikipedia.org/wiki/Invention_of_the_integrated_circuit

Like how we got from transformers to reliability. If things get more difficult, the timeline might extend, despite acceleration narratives. After all, a narrative is just a narrative and not science, or something.

OK, Gary Marcus’ super bowl vis-a-vis table from https://garymarcus.substack.com/p/super-bowl-matchup-anthropic-vs-openai, EDIT: because I was already about openai ads.

All this is not to say that the technology is without use, but that the end-of-humanity-your-job discourse is so. Let’s know the technology from start to end and put it to use.

Comments

flatcat audio manual

Sun, 05 May 2024 00:00:00 +0000

how to make flatcat audio work

log into flatcat, ctrl-C on ux0_serial and start ./ux0_serial_osc, or boot into ux0_serial_osc automatically via editing

vim jetpack/flatcat_setup/boot/start-tmux.sh

to point to ./ux0_serial_osc

for nupg setup, go to

cd /home/src/supercollider/scapps/flatcat

and run

python flatcat-server-2.py

this takes /flatcat messages on 8999, does some pre-processing and the projection. outgoing messages on /flatcat2nupg that can be plugged directly into the nupg channel control set.

to reload the projection matrix use this command. the density parameter (0.8) can be varied between 0 and 1

oscsend localhost 8999 /matrix_reload f 0.8

to run the audio, start supercollider / scide open the /home/src/supercollider/nupg/startup-interactive.scd file and run that. it starts the nupg and spawns the GUI. in the server window, click the boot button to start the server.

open /home/src/supercollider/scapps/flatcat/flatcat-nupg-2-synth.sc

and run the code blocks

to use NTMI, stop the flatcat-server-2.py script, as NTMI uses the /flatcat data directly

open ~/bla.sc and run the line with

NTMI.run;

again, the NTMI GUI starts

open flatcat2NTMI_basic.scd or flatcat2NTMI_better.scd and run the respective file entirely (execute buffer)

note: in ~/NTMI/flatcat.desc.scd you need to add your flatcat source address for supercollider to accept the messages. if you use flatty in ap mode this is better, actually because the source IP is always the same.

flatcat audio links: slides & videos

neural synthesis slides nov 23 https://docs.google.com/presentation/d/1hsxpIO2ab4e1rv_Tq3vTuX-VcE2S1dQQoUhKD0Cljm0/edit?usp=sharing
artist talk kunsthochschule kassel may 3 https://docs.google.com/presentation/d/1aHRP-3jBx_Ov5S937N2cYjaV8iJV9dG9ROk_iZkvZJM/edit?usp=sharing
flatcat audio initial prototype 1 https://youtu.be/8twXVcml3dg?si=QRA_NGPoQa4TK549
flatcat audio initial prototype 2 https://www.youtube.com/watch?v=piMOrglS96U
flatcat audio initial prototype 3 https://www.youtube.com/watch?v=5vYb9miMJfE
flatcat audio initial prototype performed by thomas https://www.youtube.com/watch?v=w1-2qQGfQYE
flatcat audio in the street front of lab https://www.youtube.com/shorts/Vxv2nxqixLg
flatcat audio kensington 1, https://www.youtube.com/watch?v=1zdpqvniwoQ
flatcat audio kensington 2, https://www.youtube.com/watch?v=eZg3PI-37k0
flatcat audio on-stage nnoi bwo eva, https://www.youtube.com/watch?v=dFJPtAW5Yjk
flatcat audio on-stage nnoi marcin, https://www.youtube.com/shorts/mtdXeNPpbaM
flatcat audio on-stage novilla, https://www.youtube.com/watch?v=KDftj5R7FeA&t=3780s
flatcat audio on-stage zügellosen bwo mitch, https://www.youtube.com/watch?v=KxriWdqYuO4
neural synthesis recording https://www.youtube.com/watch?v=3Knd6O4XKxg&t=264s
hellisotherpeople

Comments

Book review - Artificial intelligence meets natural stupidity by Drew McDermott and Computing taste by Nick Seaver

Sun, 12 Mar 2023 00:00:00 +0000

Incidentally reading the paper (no it is not actually a book) by McDermott and finishing Computing taste by Seaver, I decided to rather quickly than comprehensively write up a summary review of both.

Review of Artificial intelligence meets natural stupidity by Drew McDermott, 1976

The overarching topic is, that many or most of the issues raised in this article are still helpful and relevantly unresolved, when looking at contemporary AI.

One item is, that just because I name my program intelligent-program, that does not mean that it is intelligent.

Another item is an invitation to get to terms with the intricacies of language. Stochastic parrots do not understand how language is connected to the processes and events it refers to anymore than an n-gram model does.

The third item is about how progress in AI is often impeded by the speculation-after-implementation phenomenon, found in PhD theses. It leads to a state where problems that are really unsolved appear as if they are solved for future students picking topics, and thus left untouched thereafter. The claim is, that if an implementation would be accepted for getting a PhD awarded, and the thesis text would focus on the failure analysis of this one single implementation, it would be beneficial for the actual development of the field.

This article is a very entertaining read. Do not get scared by the number of pages, one third of it is references and each page is short. Highly recommended regarding its entertainment value and some insights on the intricacies and everyday hickups when dealing with language and that starts with words like ‘a’, ‘the’, and ‘is’.

Some selected quotes

p7/p15/156: [natural language yes, but first understand what language is]

In this section ] ‘have been harsh toward A[’s tendency to oversimplify or overglorify natural language, but don’t think that my opinion is that research in this area is futile. |ndeed, probably because ] am an academic verbalizer, ! feel that understanding natural language is the most fascinating and important research goal we have in the long run. But it deserves more attention from a theoretical point of view before we rush off and throw together “natural-language” interfaces to programs with inadequate depth. We should do more studies of what language is for, and we should develop complex programs with a need to talk, before we put the two together.

[I leave the OCR errors in there to emphasize the point with your pleasure.]

p8/p15/156 - p16/157:

The idea is formally pretty and seems to mesh smoothly with the way a rational program ought to think. Let us call it !’sidetracking control structure” for concreteness. The researcher immediately implements an elegant program embodying automatic sidetracking, with an eye toward applying it to his original problem. As always, implementation takes much longer than expected, but matters are basically tidy.

p8/p17/158: phd theses: if a mere implementation of version I

The solution is obvious: insist that people report on Version I (or possibly “| 1/2”). If a thorough report on a mere actual implementation were required, or even allowed, as a Ph.D. thesis, progress would appear slower, but it would be real.

p8/p17/158

When a program does fail, it should tell the explorer why it failed by behavior more illuminating than, e.g., going into an infinite loop.

p9/p19/160: the wishful control structure

The idea that because you can see your way through a problem space, your program can: the “wishful control structure” problem.

Sources

p7 to p9 refs refer to version https://www.researchgate.net/publication/234784524_Artificial_Intelligence_meets_natural_stupidity
p15/156 to p19/160 refs refer to version https://www.inf.ed.ac.uk/teaching/courses/irm/mcdermott.pdf

Review of Computing taste by Nick Seaver

A social anthropology of people comprising the music recommender systems scene. What is the mindset of people designing and building these systems?

In 2023 recommender systems are deployed pervasively and while this book is specifically about music recommender systems, a great deal of their logic and culture probably is relevant to the entirety of contemporary internet reality on Amazon, Alibaba, Facebook, Instagram, Tiktok, Twitter, etc.

The book for a tiny flash of a moment starts sounding like a lot of uninformative verbiage, but it is entirely not so and it took me by surprise, twice.

Once, by unexpectedly useful analytical metaphors (traps, parks, and gardening), and then again by some unforeseen story turns.

It is just quite surprising how content based approaches enter the stage only in chapter 4. That is only due to the books narrative in the sense that it proceeds mostly chronologically. This means, music recommender systems only turned to the actual audio content at a late stage in their development, at least as seen from 2023.

Incidentally, I enjoyed the extended discussion of Every Noise At Once by Glenn McDonald. The fact that musicical genre space (as indexed by usage patterns) is not a flat space but has wormholes that shortcircuits some genre connections against hierarchical separation is nicely brought out.

There is one item in there which I appreciate particularly, called acoustemology, a portmanteau of acoustic and epistemology. Epistemology is about how we acquire knowledge about stuff. Acoustemology then means, how we acquire knowledge about stuff acoustically, that is, through hearing. Essentially, you will often start hearing it first, when your car starts having a problem.

I have this thing going in my robotics research which I call preverbal audio perception. Learning things about the world through hearing on a level fundamentally below language, parsing and understanding streams of words. I have always been puzzled by the general preoccupation with language as opposed to this analytical listening before there are words. It is not only much more interesting but also a prerequisite for language competence (that’s a hypothesis).

The pyramid of listeners was also particularly noteworthy and first seen here.

Thanks, great read.

Some selected quotes

Introduction: Technology with Humanity

Page 8

understand algorithms as the absence of a human hand didn’t make sense, because, as Ogle said, humans were everywhere: they wrote the code, they generated the data, they made the decisions about how a system should work.

Chapter 1 Too Much Music

Page 28

Overload was such a fixture of local common sense that it took me a long time to realize that it might not be a problem at all. Why should access to a large catalog be understood as a burden rather than a resource?

Page 35

The term “information overload” is commonly credited to the futurist consultant Alvin Toffler, who, in his 1970 book, Future Shock, made a dire forecast: as demands on their attention grew, fueled by new media technologies, people would “find their ability to think and act clearly impaired by the waves of information crashing into their senses” (354).

Page 36

In the mid- 1500s, scholars worried about the “confusing and harmful abundance of books” (Blair 2003, 11).

Chapter 2 Captivating Algorithms

Page 50

The ultimate measure of functionality, Mike explains to me, is “the hang- around factor”— what keeps listeners listening.

Page 50

But new users pose a challenge: data- driven recommendation requires data, and new users don’t have any.

“If it can’t be used for evil, it’s not a superpower.”

I use the term “captology” to refer to this broadly behaviorist common sense in the software industry,

There are many ways one might measure how well a recommender system works, but for most of the field’s history, one metric has dominated all others: predictive accuracy.

Researchers sometimes attributed RMSE’s paradigmatic status to the Netflix Prize, a contest run by the online movie rental company from 2006 to 2009 that offered a $ 1 million award for the first algorithm that could beat its existing system’s RMSE by 10 percent. 8 Netflix provided a data set of about one hundred million ratings, including anonymized user IDs, movie titles, release dates, and the date each rating was made. Competitors could use this data to train their algorithms, trying to predict the ratings in the set as precisely as possible.

submitted by a team named BellKor’s Pragmatic Chaos, had achieved an RMSE of 0.8567 and established the state of the art for predictive accuracy in collaborative filtering.

But the winning algorithm did not succeed by finding one particular technique that beat all others. Rather, as the contest went on, high- performing teams combined their efforts and models together into ensembles that essentially blended the outputs of a set of algorithms. BellKor’s Pragmatic Chaos was a combination of teams named BellKor, Pragmatic Theory, and BigChaos.

The second- place team, appropriately named the Ensemble, was an amalgam of twenty- three other teams.

After the Netflix Prize, it became common sense among recommender systems researchers that the best results were achieved by systems that blended together the outputs of many techniques, like the system Mike was building at Willow.

Chapter 3 What Are Listeners Like?

it required an explicit theory of listeners, their behavior, and their variation.

“What you tell us about what you like is far more predictive of what you will like in the future than anything else we’ve tried. . . . It seems almost dumb to say it, but you tell that to marketers sometimes and they look at you puzzled” (Gladwell 1999, 54).

Riedl and Konstan wrote a book together, Word of Mouse: The Marketing Power of Collaborative Filtering (2002),

In a 2018 press tour, for instance, Netflix executives described how the service orients itself toward a set of roughly two thousand data-derived “taste communities” because “demographics are not a good indicator of what people like to watch” (Lynch 2018).

“There’s a pyramid of listeners,” Peter told me (fig.

“It’s hard to recommend shitty music to people who want shitty music,”

Chapter 4 Hearing and Counting

these graduate students think that the idea of musical similarity, on which recommendation depends, is too vague to be scientifically evaluated. There are simply too many ways for music to be similar.

Today, most of the large ensemble models used for music recommendation include, in some way, a representation of sound, mixed in with the other data sources we’ve encountered so far.

“Sound, hearing, and voice mark a special bodily nexus for sensation and emotion because of their coordination of brain, nervous system, head, ear, chest, muscles, respiration, and breathing. . .

Feld (2015, 12) argues for attention to acoustemology—to “sound as a way of knowing,” which is contrasted with scientific approaches to studying sound. To know through sound is to know in an embodied, emplaced, and relational way

Many critics and practitioners have suggested that music recommendation might be improved by focusing on “music itself”—that is, how it sounds.

100

this spectacle is veiled from the material eye, we have another bodily organ, the ear, specially adapted to reveal it to us. This analyzes the interdigitation of the waves, . . . separates the several tones which compose it, and distinguishes the voices of men and women—even of individuals—the peculiar qualities of tone given out by each instrument, the rustling of the dresses, the footfalls of the walkers, and so on.”

102

“The time domain is for losers”

107

The sociologist Jennifer Lena (2012, 6), for instance, defines genres as “systems of orientations, expectations, and conventions that bind together industry, performers, critics, and fans in making what they identify as a distinctive sort of music.” Djent is defined not by the djent, but by the community of people like Tomas, who come together on web forums or at concerts and make the genre cohere. If this is what genre is, then it is unsurprising that algorithms working only from audio data would make mistakes that appall true fans

108

This problem, I learned, is called the “semantic gap”: there is a distance between the information encoded into an audio file and the meaning that an enculturated human mind can interpret from it. 2

Chapter 5 Space Is the Place

132

listeners might be surprised to discover that their “favorite” genre had a name that neither they nor the artists involved had ever heard of.

132

“I’m cheerfully unrigorous about what a genre can be,” McDonald said in an interview, “Some of them are musicological, some are historical, and others are regional themes.

133

While genre theorists might find it odd or incorrect to call “sleep” or “deep cello” genres, for people like McDonald, the term was essentially synonymous with “cluster.” Conventional genres simply served as a prototype for thinking about the many ways that similar music might be gathered together.

133

heard the same thing: “genre recognition”—a long-standing analytic task in which researchers tried to guess a song’s genre from its audio data—was being superseded by “tag prediction,” in which software would predict which labels users had assigned to tracks on music streaming platforms like SoundCloud.

135

The conquest of genre by similarity, strangely enough, results in a situation where artists may be part of a genre without realizing it;

137

postgenre artists

143

A band like Queen, for instance, might accidentally receive a boost in popularity if Whisper’s data scrapers mistook a news article about Queen Elizabeth II for one about the rock group.

Comments

Movies about AI and robots

Mon, 23 Jan 2023 00:00:00 +0000

Movies about AI and robots

Robots without AI are just dumb machines, so quite logically robots that appear in sci-fi stories are strongly connected to AI. Here is an incomplete list of such movies which I found noteworthy in a good or bad way. I am sure I forgot some obvious ones but there will still be plenty of more obscure ones. Either way I am happy to have your remarks and pointers in the comments or messages.

I will start with a list featuring androids, which are robots that are not only humanoid but effectively so well-made that they are indistinguishable from real humans. The second list is with robots of any kind, the third one is about related movies featuring some kind cyborg or otherwise hybrid human-artificial intelligent beings. Adding a forth one which I think is also related about (it is all a) simulation hypothesis.

Movies with androids

Blade Runner (classic), a major topic of Blade Runner actually is the difficulty of distinguishing androids and humans. I thought Blade Runner 2 was quite good as well.
Real humans (Swedish original, 2012), https://en.wikipedia.org/wiki/Real_Humans
Ich bin dein Mensch (I’m your Man) 2021, https://en.wikipedia.org/wiki/I%27m_Your_Man_(2021_film)
A.I. - Artificial Intelligence, 2001, https://en.wikipedia.org/wiki/A.I._Artificial_Intelligence
Eva, 2011 https://en.wikipedia.org/wiki/Eva_(2011_film)
Love, death and robots, EP14S1, Zima Blue, https://lovedeathrobots.fandom.com/wiki/Zima_Blue

Unwatched

Westworld (both original and series), https://en.wikipedia.org/wiki/Westworld_(film), https://en.wikipedia.org/wiki/Westworld_(TV_series)

Other robot or AI movies

Chappie (CHAPPiE), 2015, by Neill Blomkamp, very nice and plausible story about an actual developmental robot in the wild, https://en.wikipedia.org/wiki/Chappie_(film)
Ex machina, 2014, https://en.wikipedia.org/wiki/Ex_Machina_(film)
Hitchhiker’s guide to the Galaxy (robot Marvin)
Final space (robot KVN, among others)
Her, 2013, by Spike Jonze, Man falls in love with chatbot, had to stop watching in the middle because it felt awkward, https://en.wikipedia.org/wiki/Her_(film)
Wall-E
2001: A Space Odyssey HAL. HAL’s embodiment is the spaceship? OK. https://en.wikipedia.org/wiki/2001:_A_Space_Odyssey_(film) bwo Joe Gilmore.
Interstellar, 2014, they have these clonky robots in that film, https://interstellarfilm.fandom.com/wiki/Interstellar_Wiki
Nummer fünf lebt (classic)

Unwatched

Hi, Ai - Liebesgeschichten aus der Zukunft (hi ai, documentary, bwo udk 111) https://www.imdb.com/title/tt9081788/
Hardware, 1990, Indie cyperpunk, https://en.wikipedia.org/wiki/Hardware_(film)
Brian and Charles, 2022, https://en.wikipedia.org/wiki/Brian_and_Charles

Ghost in the Shell (1995 original), more a cyborg tale, https://en.wikipedia.org/wiki/Ghost_in_the_Shell_(1995_film)
Paprika, 2006, https://en.wikipedia.org/wiki/Paprika_(2006_film)
Pantheon, 2022, TV series, https://en.wikipedia.org/wiki/Pantheon_(TV_series), GitS reference included. bwo sansculotte
Robocop, 1987, (classic), dead person is revived as a cyborg

Simulation hypothesis

Welt am Draht (World on a Wire), 1972, Fassbinder, Two part TV show, great style and mood. Story by Daniel F. Galoye, 1964, Simulacron-3, https://en.wikipedia.org/wiki/Simulacron-3
The thirtenth floor, remake of Welt am Draht / re-adaptation of Simulacron-3
Matrix trilogy, for completeness

Aliens and alien robot embodiment

Under the skin, 2013, https://en.wikipedia.org/wiki/Under_the_Skin_(2013_film), bwo Luke Harby
Terminator 1,2,3, including the time travel this is close to an alien robot

Other

Star Wars of course, but this one is so big I will leave to specialist sites / people.

Comments

chatGPT reveals itself to be ungrounded by its comment on proprioception

Mon, 23 Jan 2023 00:00:00 +0000

A one-shot interaction with chatGPT asking for comment on proprioception

Refraining from the usual screenshot-based commentary but otherwise standard, chat session starts with a prompt by an agent

Oswald Berthold:

what do you know about proprioception and why it is so important for safe AI?

Which is reacted on by the other agent. Will dissect the original four-paragraph reply with immediate inline comment.

chatGPT:

Proprioception is the sense of the position and movement of the body, and it is important for safe AI because it allows a robot or AI system to be aware of its own body and its position in relation to its environment. This is important for a number of reasons.

This is entirely correct, proprioception is the sense of position and movement of the body. Most importantly, this refers to that part of the sensation which is conveyed even with your eyes closed, a point missed by the AI.

First, proprioception helps a robot or AI system to navigate its environment safely. For example, if a robot has proprioception, it can detect when it is about to bump into an obstacle and take appropriate action to avoid the collision.

Proprioception does help a robot to navigate its environment safely, you are good there chatty. The example on the other hand is old-school in terms of framing it as obstacle avoidance. Framed more accurately, proprioception allows us to detect obstacles, but not so much in the “when-we-are-about-to” way, instead more in the “when-it-has-already-happened” way. The special thing about proprioception then is, that it is incredibly fast. It is so fast, that it allows you to adapt your speed or direction of motion so quickly, that many times you will manage to influence the action (motion) consequence, while you are bumping into something. Consider the timescale of these events, it is in the few-milliseconds range.

Second, proprioception can help a robot or AI system to interact with its environment in a more natural and intuitive way. For example, if a robot has proprioception, it can use its sense of touch to detect when it is holding an object and adjust its grip accordingly.

This is correct again on a superficial level, but misses the point in the same way as for the previous items, in the framing. Proprioception does not happen primarily on the object level, but instead on the level of the individual motion. So the salient stimulus here is the moment when the motion does not yield the predicted response, because it is hitting the object (general class of external entropy by way of condensed matter physics), but this applies to each individually moving articulation of the body, and foremost, their integration needs to be taken into account.

This is quite informative since we know how the response was informed (average internet discourse), so the issue is there but it is discussed in wrong terms, entirely.

Overall, proprioception is an important capability for any robot or AI system that is designed to interact with the physical world, as it allows the system to be aware of its own body and its position in relation to its environment, which is necessary for safe and effective operation.

This only repeats what we said two paragraphs ago, and is cyclical reasoning in its reference to safe operation. Two slightly vague paragraphs of response do not count as “a number of” reasons, so again we are cyclical in the argument.

There certainly is a larger discussion going on about common sense and an understanding of physics which we require of trustworthy AI. This discussion aligns very well with our own framing in terms of grounding.

References

chatGPT errors: submit form: https://researchrabbit.typeform.com/llmerrors
chatGPT errors: database: https://docs.google.com/spreadsheets/d/1kDSERnROv5FgHbVN8z_bXH9gak2IXRtoqz0nwhrviCw/edit#gid=1302320625

Comments

Book review - Unthinkable - An Extraordinary Journey Through the World's Strangest Brains by Helen Thomson

Sun, 20 Nov 2022 00:00:00 +0000

Summary review

Unthinkable is a 2018 book by Helen Thomson which looks at nine different people and their brains.

Human brains are not all alike, That much is clear to everyone. What is less clear, is the sheer amount of how different they can actually be. All the time still being OK brains, each, in the sense of that they cope. People’s body heights for example are also different, but the amount they differ by is limited. There are no people that are 10cm or 3m in height. With brains, this is not so. While there is something like roughly similar ‘normal’ brains for many people, there are also wildly differing brains and there is hardly any limit on how different from the majority any single brain can be.

As such, it is a story about diversity, and diversity in as extreme a form as it possibly gets. It highlights nine individuals with very specific and rare conditions affecting their brains, and thus, their experience of themselves and the world. Tons of psychology and neuroscience lessons for use in robotics included for free.

Conditions include 1/ not forgetting, 2/ loss of location, 3/ synaesthesia (auras), 4/ switching personalities, 5/ constant hallucination (audio), 6/ becoming a tiger, 7/ becoming unreal, 8/ waking up dead, 9/ feeling other people’s pain. For all the details I refer you to the book itself. Instead, let’s pick a few gems of insight and narrative and comment on them.

Selected and annotated quotes

Maybe it is still better to read a book on paper, make less highlights and just have fewer problems?

85/86 synaesthesia, Vilayanur Ramachandran, round blob vs. jagged piece, naming them bouba vs kiki

In this experiment on synaesthesia the close relationship between perception of shape and sound or language comes to the front. Love it.

96 the dress. a classical example. straightforward indication of two entirely equally valid modes of perception, prevalent in the population. Beside the two main modes, there is a few lower but equally stable side modes, too.

I have seen this come through back then but wasn’t so impressed as this time round. The picture is really very good and-just-works for everyone I have shown it to. Myself I seem to be able to see both black/blue and white/gold while others always see it in the same colors.

143 David Rosenhan, 1973 experiment, schizophrenia psychiatry admission experiment

Even if the Rosenhan’s rigor in that study is subject to veritable criticism, the idea is interesting and it certainly just makes for a great story.

145 Ganzfeld technique sensory deprivation, ping pong balls on eyes and white noise on the headphones.

Very interesting sensory deprivation experimental technique. I have heard about this one previously, have tried it once I think but found it stupid at that time and didn’t get good results because probably I just was too impatient. Things to do.

146 Oliverc Sacks, the brain doesn’t tolerate inactivity

Oliver Sacks is well known and just putting some more matter on the the sensory deprivation technique of creating hallucinations, this is a crisp way of putting it. If in a deprived state without any changes in stimulation, the brain amplifies its own random fluctuations into fully fledged but spurious experiences.

161 Schizophrenia is one of the most complex of human disorders affecting 1 in 100 persons.

Noteworthy phenomeno and noteworthy numbers. Given Dunbar’s number as a reference (probably suboptimal choice), that means each of us know’s one and a half people affected by schizophrenia.

191-197 Interoception & all related

This is a bit terse. It highlights in detail the essential importance of interoception for a passable experience of self at all. Need to go back for details but don’t have access to the book right now.

213 Seth, Chinese woman brain scan entire cerebellum is missing

This year I met a cat (Earl) that was said to have no cerebellum, and that cat was certainly moving in a weird way. It quite took me that there also is people with that condition, so I researched this particular fact. Also note the man with only 10% of his brain below.

A new case of complete primary cerebellar agenesis: clinical and imaging findings in a living patient, https://academic.oup.com/brain/article/138/6/e353/269537
This Woman Lived 24 Years Without Knowing She Was Missing Her Entire Cerebellum, https://www.sciencealert.com/this-woman-lived-24-years-without-knowing-she-was-missing-her-entire-cerebellum
Woman of 24 found to have no cerebellum in her brain, https://www.newscientist.com/article/mg22329861-900-woman-of-24-found-to-have-no-cerebellum-in-her-brain/
An adult woman was found to be missing her cerebellum, https://www.washingtonpost.com/news/speaking-of-science/wp/2014/09/11/an-adult-woman-was-found-to-be-missing-her-cerebellum/
Related: Meet The Man Who Lives Normally With Damage to 90% of His Brain, https://www.sciencealert.com/a-man-who-lives-without-90-of-his-brain-is-challenging-our-understanding-of-consciousness?:_sciencealert-lates___=

218 default mode network

I just love the default mode network and note down every single reference I find to it. You should check out the default mode network if you haven’t.

246 Compassion training to overcome empathy, willfully extend warm feelings towards everyone in the environment

Suffering from a bit of over-empathisizing usually I found it interesting to again learn about the distinction between compassion and empathy, and then, that you can train for compassion in favor of empathy.

References

eot

Comments

Book review - Survival of the Friendliest - Understanding Our Origins and Rediscovering Our Common Humanity, by Hare & Woods, 2020

Thu, 20 Oct 2022 00:00:00 +0000

Summary review

Survival of the friendliest by Brian Hare and Vanessa Woods examines the question about how humans became the dominating hominid species over the last 300000 years or so. This is an interesting question since there were other hominids around during that time (eg Neanderthals), and most of them were physically stronger and had larger brains. It puts forward the idea that this is due to domestication, essentially selection by friendliness, leading to skills of cooperative communication among humans, and then also dogs, and bonobos. Cooperative communication made the decisive difference between humans and the other hominids, which were just not as good at it. It makes sense that working together as a group allows the species to perform better overall, as a population. But how did domestication come about if there was no one around that would domesticate us? The center bit in this is self-domestication, which occured in humans, and also dogs and bonobos, as is suggested and explained in the book. And I am not a dog lover.

Putting friendliness on the test stand, the text includes a section on the psychology of groups and identity that outlines an analysis and approach at how to bridge the mindset divide we experience as a global society. It connects quite well incidentally with Braving the wilderness by Brene Brown which I recently reviewed here. Also, there was a very good post by Cory Doctorow just very timely about “how minds really change” with an interesting book review of The Persuaders in there.

I think this is all very hot and interesting, despite some criticism of the self-domestication hypothesis which hinges on the fact that Belyaev’s foxes were preselected and had seen previous breeding as they were imported from a Canadian fox farm, back at the time.

To see why this is so, let’s look at some of the details involved in this bit of evolutionary trajectory, starting with eyes 👀. It is only humans and their domesticated animals which have white sclera. White what?!! Sclera is the name for what is white in your eyes. If the white stuff wouldn’t be white, we wouldn’t see the color of our irises. Cool, no? Anyway, the white sclerae also allows us to see where fellow humans are actually looking at. This is super essential for intention recognition and thus cooperation, seeing what someone else is going to do just by looking at their face. This ability cannot be underestimated and can be tested for its presence in animals by hiding treats under one of two cups and giving a hint to where the treat is hidden just by the animal trainer looking at it. If the animal can use this cue, it will find the treat more often on first trial rather than not. Most other animals’s sclera have the some color as their irises, and the reason is strategic. Most other animals don’t want you to be aware of where they are looking and what they’re going to do next. That’s true even for their own social life with their own kin. Not so much friendliness there, really.

There is a hallmark experiment started by Dimitri Belyaev in the former Soviet Union. It consisted in breeding foxes by selecting foxes that are allowed to have offspring entirely by their friendliness. Friendliness there is defined by having no fear of interacting with humans. Turns out, that 1/ this works really well and 2/ that by selecting for friendliness we also get a number of interesting side effects, like nice & splotchy variations in fur patterns (which provide less camouflage), curled up tails, smaller muzzles, smaller brains, and voila, white sclerae.

Now the question arises of how this got bootstrapped. Even for dogs it is most likely that some kind of self-domestication took place with wild wolves before humans took up on breeding dogs. It happened simply by the fact, that wolves who were more friendly, that is, less afraid of getting close to humans, would start to seek out human settlements foraging for food more often than those wolves who were more afraid. This is the self-domestication hypothesis in action for wolves and dogs.

Dogs for example also coevolved the enzyme which allows humans to digest and break down starch, which is a predicament to leverage the benefits of agriculture, one of the underpinnings of current human civilization. Either way, self-domestication hypothetically also took place for humans starting something like 80000 years back approximately, leading to the domesticated and friendly version of our predecessors that we are.

A corner stone of our rich social life is theory of mind. Theory of mind is a functional lump of brain activity which allows you to reason about the activity of the other mind in the other person and what they are up to, by thinking about them in terms of their mental state. This is realized by the theory of mind circuit, a set of functionally connected brain areas. Check up on the concept as it is really exciting. My favorite feature of theory of mind is so-called “false belief”. This skill comes up in infant humans around the age of four (check) and allows them (us) to understand that someone else is mistaken. :shrug:

A lot more interesting things can be said about the theory of mind circuit in our brain but here, let’s take it to one particularly relevant point. Where does our ability to afflict mindless and utter industrialized violence to other people come from if we are so friendly? It all starts with the process of dehumanization. If we don’t perceive the other as human but rather as subhuman, none of our friendliness criteria, empathy, and so forth apply. Dehumanization can easily be identified in all populist rethoric, essentially along the lines of “these animals on the other side …”. In-group / out-group thinking. The thing is, if we perceive someone, anyone, else, another fellow human, an animal or anything alive, as inferior, our theory of mind circuit goes blank (that can be measured), and then this is really bad. Because it remove all the empathic barriers to our applied ethics and that makes us as cruel as it gets.

Here the story also connects with Kate Darling’s New breed and our relation with animals and whether we dehumanize them or not and then what we are able to do to them.

Good and wild stuff, thanks.

Selected and annotated quotes

pg 3-4 By age four, you could guess someone’s thoughts so cleverly that for the first time, you could lie. You could also help someone if they had been deceived.

You will spend the rest of your life wondering what other people are thinking.

Theory of mind allows us to engage in the most sophisticated cooperation and communication on the planet. It is crucial to almost every problem you will ever face.

This is the last part on the schedule with which we acquire theory of mind as a baby and small child. The last sentence nicely emphasizes the importance of that function and concept.

pg4 You can teach only if you can remember what it is like not to know.

A more general one but a gem at putting the focus on the part that does not know. It is such an important state of mind and we look back on ourselves, there was always a time when we did not yet know something that we now know. We learned everything that we know at some point since being born (or since shortly before that).

pg189 If there is an essential lesson to be learned from psychology research, it is that we are not always aware of what shapes our attitudes and behavior.

I just love psychology and how it does just that, tell weird and funny stories about how behavior is shaped in ways that we would not know from pure introspection, aka the introspection fallacy.

More to come, check back.

References

Comments

Book review - The New Breed, What Our History with Animals Reveals about Our Future with Robots by Kate Darling

Mon, 19 Sep 2022 00:00:00 +0000

Summary review

Totally outstanding book bringing an entirely new spin to how we can imagine, and thus invent and build, a symbiotic and meaningful life with robots.

While fantasizing about the vision for jetpack cognition lab, on Jan 1st, 2020, I was noting down “a new breed of robots”, with my own idea going more down the lines of the self-learning capacity we want to build into our robots. Fast-forward to Sept 27th, 2021, when I first seem to have encountered The New Breed: What Our History with Animals Reveals about Our Future with Robots, by Kate Darling, a researcher of Human-robot interaction, robot ethics, IP theory & policy at MIT Media Lab. Finally I was ready to dive in few weeks ago and I was not disappointed. All this “new breed” talk is about how we get to a new way of thinking about robots, both for the general public and for roboticists. We need to elevate our thinking here in order to get a new generation of robots on the market and really make a difference to the positive for many people. Darling’s book is of great help in spinning that story.

In summary, Darling wittingly and successfully proposes to use our view of animals as a blue print for robots as they could be. The book is brimming with exciting details about our relationship with animals and the way our psychology gets in the way of establishing firm ethics for dealing with animals and establishing their rights. This relationship is projected onto the situation with robots, how to design them, how to treat them, and whether it is OK to kick them. It seems with animals we are mostly doing all of the kind stuff for our own psychological and emotional well-being. In other words, whether with animals or with robots, the way we treat them is more about the effect, real or imagined, this has on us than on the other creature itself.

Kids growing up in Alexa households start shouting commands rather than asking for stuff.

I really enjoyed both the general proposal for contemporary robot story-telling put forward in the book, as well as the prose itself, which is simple, readable, and stealthily funny. Totally recommended for roboticists and everyone else alike. In addition, it could be a very interesting read for people interested in animal rights & vegetarianism, as it unrolls the history of animal rights in (mostly) western societies.

Annotated quotes

This being not the only thing that got me excited, let’s isolate some core points that did so too, and look at them more closely in turn using some quotes from the book.

Public imagination about robots is strongly centered on the idea of human-like robots, which look like we do and can do everything we can do, only better
The state of robotics is nowhere near the science fiction
Robots allow us to learn something about ourselves.

Public imagination and story-telling

Imagination is largely fueled by the stories that we know. Robot stories are until recently science fiction, mediated by literature and movies. There are some classic lingering stories like the Golem, Frankenstein, Karel Čapek’s play, and maybe Metropolis. Influential later examples include the Star Wars and Terminator universes, among many more. Already,

As technology critic Sara Watson points out, our stories, too often, compare robots to humans. [p. xiv]

which is nourishing a situation where

Many people are not thrilled by the anticipated robot takeover. Our concerns are particularly centered on the idea of creating something like us, with humanlike agency, that will take our steering wheels and harm us or our children. [p. xii]

But creativity and wild imagination to the rescue. Since

When we assume that robots will inevitably automate human jobs and replace friendships, we’re not thinking creatively about how we design and use the technology, and we don’t see the choices we have in shaping the broader systems around it. [p. xiv]

From an analytical perspective it can be said, that

Inexperience with robots and their inner workings may make certain machine behavior seem magically lifelike to people. [4. Robots versus Toasters > p. 92]

but the notion of inexperience is too general to be actionable.

I’ve heard curse words directed at the public intellectuals who extol the dangers of robot takeovers, and complaints that the big-name alarmists are mostly physicists, philosophers, and CEOs who don’t have in-depth knowledge of artificial intelligence or robotics. [1. Workers Trained and Engineered > Page 6]

As roboticists it is our responsibility to share our experience, a purportedly more differentiated one. In doing this we certainly do not want to go about it by lecturing. Rather we want to rely on a universal currency and provide alternative stories of the future. Darling picks up on animals for this purpose, which is a great choice. Everyone knows animals. They provide infinitely varied example and inspiration of ways of being different than the human one while being unquestionably real. In doing this, a lot of ambiguity comes to the surface that was present while our current relation and general view of animals was shaped, historically. The close look she takes reveals that a lot of ambiguity still persists. And all of this makes for some awesome stories. In the meantime,

No, the oxen did not replace the farmer. But yes, the oxen does something farmer-like.
No, hunting dogs did not replace the hunter. But yes, the hunting dog does something hunter-like.
No, adoption of pets did not replace friendship. But yes, the pet does something friendship-like.

OK, and yes you guessed it, while not replacing pets social robots do pet-like things where a real animal is not an option.

So, if we do think creatively, we can come up with all these robot ideas where the robot meaningfully integrates with our activities, partially supplanting things we do ourselves, and at the same time extending our overall capacity with things that we could not do before. This is a whole lot about how we think about organizing our activity and tasks. House cleaning is complex and parts of it can be done by a robot. Garden work is complex and parts of it can be done by a robot. Companionship is complex and parts of it can be done by a robot. Everything we do is complex, and parts of it can be done by a robot. Art, relaxation, being social.

State of robotics

Awesome stories, it was said. Let’s start with some funny ones that the book adds to a growing list of robot, AI & tech fails.

Our advancements in artificial intelligence, as amazing as they are, haven’t gotten anywhere near understanding how to create the adaptable, flexible general intelligence that a human, even a toddler, has. [1. Workers Trained and Engineered > p. 14]

Robot fails

The vacuuming robot Roomba represents the most successful type of consumer robot to date with iRobot selling 20 million units in 2018. These robots have real utility within a narrow scope. Early versions and apparently also more recent ones will fall down stairs. But hilariously they are also known to mindlessly distribute dog poop if they come across a pile. The robot cannot sense this and just keeps doing its thing.

The company had invested in a modern art piece: a robotic office copy machine that was designed to wander the halls, randomly creating and spitting out copies of nothing. I only got to see it once, because, sadly, it wasn’t able to recognize stairs and eventually fell down them. [1. Workers Trained and Engineered > p. 9]

The absurdidity of this robot generates some entertainment points, and then another story of robots and stairs.

Musk had promised to produce five thousand Model 3 electric cars per week in 2018, but Tesla couldn’t even make half of them. What went wrong? According to analysts, the robots, while able to work consistently and precisely, weren’t able to recognize the litany of minor defects that can happen during the manufacturing process—slightly crooked parts, for example—leading to problems down the line. [1. Workers Trained and Engineered > p. 13]

The “resolution” of the robots perception and calibration was not sufficient for all levels of scale, so these micro-defects evaded the system.

Let’s add this one from 2022. At a public chess tournament in Moscow a chess-playing robot pinched then fingers of a seven year old kid and apparently broke them.

We have been calling these grounding fails. With grounding we mean the completeness of the connection of an agent’s actions out through the sensorimotor layer and down to physical reality. As the Tesla story clearly tells, physical reality does have its special surprises, even for our own reasoning at times.

Ourselves

What does this tell us about ourselves?

People are more talented than we give them credit for. [1. Workers Trained and Engineered > p. 12]

For us it is easy to do all these things, sweep the floor withouth falling down the stairs. Play chess without physically injuring our opponent, even if our hands touch while over the board. It only appears easy to ourselves though, because we simply cannot perceive our internal activity that makes sure this things are taken care of.

This is the great introspective fallacy that some many, including roboticists of every generation, fall victim to.

That is a grand one and there is more things robots can teach us about ourselves and the skills we actually have but do not perceive them as such. This only becomes obvious when you try to replicate these skills in a machine. [tbd]

Movement

This bit on movement perception really kicked in then. Did you know that there is brain circuitry tuned movement in particular? Thanks to this

… physical robots trigger another piece of our biological hardwiring: our perception of movement.

Our sense of movement is so deeply engrained that it responds to entirely abstract situations.

In a seminal study from the 1940s, psychologists Fritz Heider and Marianne Simmel showed participants a black-and-white movie of simple, geometrical shapes moving around on a screen. When instructed to describe what they were seeing, nearly every single one of their participants interpreted the shapes to be moving around with agency and purpose. They described the behavior of the triangles and circle the way we describe people’s behavior, by assuming intent and motives.

This highlights our incliniation to anthropomorphize, and more particularly attributing agency to anything that moves, except maybe ballistics. But our perception is even more nuanced, allowing us to perceive psychological states in the movement of an “agent”.

What brought the shapes to life for Heider and Simmel’s participants was solely their movement. We can interpret certain movement in other entities as “worried,”“frustrated,” or “blinded by rage,” even when the “other” is a simple black triangle moving across a white background.

All of this makes sense in evolutionary terms and also tells us about how deeply this capacity is rooted in our brain.

Many scientists believe that autonomous movement activates our “life detector.” Because we’ve evolved needing to quickly identify natural predators, our brains are on constant lookout for moving agents.

It is basically identical to our inclination to see a face in anything with two round things next to each other, which we take as eyes. The technical term for this is pareidolia.

The researchers also found evidence that animal detection activated an entirely different region of people’s brains. Research like this suggests that a specific part of our brain is constantly monitoring for lifelike animal movement.

All quotes from [4. Robots versus Toasters > p. 97-99]

This realization allows us to create extremely simple robots that still have a profound effect if they manage to activate these motion sensitivity in ourselves. The stuff discussed above only refers to visual stimuli. With robots this can easily be extended to the tactile realm to create haptic stimuli and let you perceive life-like motion with your eyes closed. The simplicity of this approach allows us to create many robot designs which are far from any shape we know, and in this way keep expectations at a minimum and be able to surprise.

References

Comments

Book review - Risksavvy, Surfing uncertainty, Braving the wilderness

Mon, 15 Aug 2022 00:00:00 +0000

Review of three books all in one

After reading Gigerenzer’s Risksavvy I started out doing a meta-review of five books I’ve had on decision making and risk management. Yukkes, topic proved too sticky and I got stuck. To break the stuckness, and new books were coming in waiting to be reviewed, I decided to go for a five-sentences-for-each-book approach. Here we go, on average more than five sentences but within reasonable bounds. Hope you enjoy, all three recommended.

Book review: Risiko by Gerd Gigerenzer

A recommendation by rakso2

Risiko, aka Risksavvy as the original English title, by Gerd Gigerenzer is a book about decision making.

The decisions we make every day a number of times are usually connected with uncertainty, and thus risk.

Take the later train and not waste time or take the earlier train and leave some maneuvering space to still take a cab, if the train is delayed or cancelled, Berlin example.

Essentially we are making decisions in face of three types of outcome scenarios: certainty, risk (which is known uncertainty), and unknown uncertainty. The most important bit is to determine, which of the three scenarios we are in, in order to then select the appropriate decision strategy. If we don’t watch our back, we might be taken by an illusion, and treat a risky situation as if we would be dealing with certainty, through the zero-risk illusion, or treat a totally uncertain situation as if we knew the uncertainties, by the turkey illusion. This is illustrated in the picture below, taken from the German edition of the book. The risk- and uncertainty scenarios each have a couple of domain examples where they apply. The bottom line is that just about any domain with true relevance like love, economy, or health are in fact subject to total uncertainty. Alas, on average these domains are treated as if they were about known uncertainties, which leads to gross calamity.

Risksavvy makes a clear voice for the use of heuristics, in face of the last of these three scenarios, in favor over optimization and the purely rational stance as taken by Kahneman.

I met Gerd Gigerenzer earlier this year and we discussed decisions, flatcat, and his new book.

The definition of a startup by Eric Ries is “an institution to develop a product under circumstances of extreme uncertainty.”

The book is very well written, sits on a sound body of quantitative results and reasoning, and is loaded with incredible stories from pension plans, health care, and more. I read the German edition but trust in the translation.

In additional Gigerenzer launches the clear idea of putting the relevant subjects (statistics, psychology) into primary school curricula, which I fully support as a sustainable policy. Quote from Alison Schrager

Gigerenzer believes we should teach risk and probability literacy the same way we teach reading or basic math.

Annotated set of the top five quotes.

Since I read German edition on paper, it is not verbatim quotes but my own compressed notes, then translated to English.

Some actual heuristics

Always ask what is the reference class, percent of what? p16
Always ask, what is the absolute increase in risk? p18
In case of a conflict between reason and strong emotion, do not use arguments but use a strong emotion with an opposite sign. p25
Checklists in aviation vs. in hospitals. If in the hospital and you are able to, always ask “do you use checklists?”. The difference here, is that on the airplane, the crew is also stakeholder in the safety, where in the operating room, the crew is not. p76
Ask a doctor not for recommendation but what he/she would do. p81
Ask waiter not for recommendation but what he/she would eat. p90
Satisficing, one-reason decision making, good for dating, hiring, and similar. p191

Some philsophy

Three rules: everyone can learn to deal with uncertainty, experts are part of the problem rather than the solution, less is more p27
Intuition and rules of thumb: assuming intelligence is necessarily conscious is a huge mistake. p46
Many experts and the public: more is better (information, computation, optimization). But, less is most often more. p58
Heuristics: accuracy-effort trade-off p58
A system that does not make mistakes is not intelligent. Intelligence means going beyond existing perceptions and taking risks. p64
Representative survey in the EU, is the following statement true? “Ordinary tomatoes do not contain genes, whereas genetically modified tomatoes do”. 36% false, 41% true, 23% don’t know. p102
Financial experts predict worse than dice (EUR - USD rate, stock prices). 10 years of always 50/50 probability to predict correctly, there remain about 10 “experts” from an initial 10000 who always predicted correctly. p120
Fear of personal responsibility leads to second-rate decisions. p121

These are not heuristics but jocular illustrations of some of the problems.

Paul Volcker: The invention of the ATM is the only useful innovation ever made by banks. p140
Nation of lawyers, created by lawyers for lawyers. p238
Make predictions all the time, but don’t keep records. p122
Banks: 3-6-3 business model, savings get 3% interest, loans yield 6% interest, be at the golf course at 3 o’clock p289

Translated with https://www.DeepL.com/Translator (free version)

Book review: Surfing uncertainty by Andy Clark

A recommendation by Guido Schillaci a Roboticist and AI reseacher

Andy Clark is a prolific advocate of the radically predictive view of brain function, the predictive brain hypothesis.

The work revolves around expounding the implications of the predictive view of the organization of neuronal activity, as exemplified by the publications of Friston and colleagues.

While alternatively called the free energy principle of brain organization or the active inference principle, or yet something else, this view is the first to make a categorical difference by flipping the whole idea of behavior from the stimulus-response principle over to the prediction-error principle.

That means that animals are not driven by the stimuli coming in from the environment whereupon they are free to choose a response. Rather, they are actively seeking out stimuli through their actions and shaping their experience just a tad ahead of the passive flow of things.

Do you realize what this means?!!

It completely breaks the traditional view of stimulus-response, sense-think-act, and so forth by a categorical shift. In computational terms, this make a statement about what is the entry point of the sensorimotor loop. This image has of course huge implications for the design of robots, if not interpreting human actions and experience in a better way, especially for agents that are designed to surf on new situations they never before experienced, through the use of exploration (trying stuff out) and learning (sorting into what worked and what didn’t). Another picture below taken from my thesis (p35) illustrates this transition based on a transformation of the traditional agent-environment diagram.

It means that every single piece of data counts starting from the beginning of learning in order to try out the most informative thing next at any given time.

Plus all the hierarchical stuff you need to check out on your own <3

There is plenty of other implications and welcome simplifications within the nascent theory of adaptive behavior which makes this view so essential.

The book is awesomly written, although it took me a bit to lock onto the groove and that did pay off, it focusses on the predictive processing formulation within the set of available formulations, and comprehensively explores all corners through many angles making using of plenty well-chosen examples, often visually presented. Super-recommended.

Annotated set of the top five quotes.

There are so many good ones that I will restrict myself to those from the preface. The first three are about shifting the locus of agency, that is, where is the entry point of the sensorimotor loop.

To deal rapidly and fluently with an uncertain and noisy world, brains like ours have become masters of prediction—surfing the waves of noisy and ambiguous sensory stimulation by, in effect, trying to stay just ahead of them. Page xiv

A skilled surfer stays ‘in the pocket’: close to, yet just ahead of the place where the wave is breaking. Page xiv

Perceiving, imagining, understanding, and acting are now bundled together, emerging as different aspects and manifestations of the same underlying prediction-driven, uncertainty-sensitive, machinery. Page xiv

These three bring home the essential point about how the predictive processing view goes hand in hand so nicely with the challenges of self-learning robots.

Predictions concerning these more inward-looking channels will prove crucial in the core account of action, and in accounting for feelings and conscious experiences. Page xv

In an especially satisfying twist, it will turn out that meat that constantly attempts (using a multilevel inner organization) to predict the plays of (partially self-caused) sensory data is nicely positioned to learn about those regularities themselves. Page xv

The prediction task, as we shall see, is thus a kind of ‘bootstrap heaven’. Page xvi

Book review: Braving the wilderness by Brené Brown

A recommendation by Marjan Sharifi

Brené Brown is a researcher, storyteller, and (currently enraged) Texan who’s spent the past two decades studying courage, vulnerability, shame, and empathy.

After checking out her work, aeh, website, the book “Braving the Wilderness - The Quest for True Belonging and the Courage to Stand Alone” spoke to me with the stand alone bit, so I dived right in.

The story and book are great, while it took me a bit to accept some of the context, but hey. That’s why we are here.

It is about belonging, how a sense of belonging is necessary for the health of a human being, and what happens if you end up, like many of us, without a clear sense of belonging to any specific group, clientele, off-the-shelf identity, thing, or what have you.

An alternative, exemplified through the quote by Maya Angelou, “You are only free when you realize you belong no place—you belong every place—no place at all. The price is high. The reward is great.”, the story takes us through the quest of finding belonging to yourself, from where everything of true signficance starts.

The place where we need to go, stand, and generally hang out is called the wilderness. Going out there, being vulnerable, and showing yourself.

The proposal is to use the concept sequence of BRAVING: Boundaries, Reliability, Accountability, Vault, Integrity, Nonjudgment, Generosity.

Another set of four piece rule, even more practical, goes like 1. People Are Hard to Hate Close Up. Move In. 2. Speak Truth to Bullshit. Be Civil. 3. Hold Hands. With Strangers. 4. Strong Back. Soft Front. Wild Heart.

These things are explained in the book of course in more detail, also with vivid examples. I think that the wording is well chosen though, and the terms speak for themselves and thus can be quoted and used just so to reasonable effect. Recommended.

TIL: bullshit

Annotated set of the top five quotes.

This is the inspirational quote from Maya Angelou, used by Brown as a light house guide for the story.

You are only free when you realize you belong no place— you belong every place— no place at all. The price is high. The reward is great.

At the peak of unbelonging.

What kind of shit is it when you don’t even belong at AA?

And some calls to self-action.

Belonging to ourselves means being called to stand alone—to brave the wilderness of uncertainty, vulnerability, and criticism.

What all wilderness metaphors have in common are the notions of solitude, vulnerability, and an emotional, spiritual, or physical quest.

Rather than pitching wild and innovative new ideas that could potentially change everything, we’re staying quiet and small in our bunkers and loud in our echo chambers.

References

Gerd Gigerenzer, Risksavvy. How to make good decisions, Penguin, 2013
Andy Clark, Surfing Uncertainty: Prediction, Action, and the Embodied Mind, Oxford University Press, 2015
Brené Brown, Braving the Wilderness: The quest for true belonging and the courage to stand alone, Ebury Digital, 2017

Navigating option landscapes ✷

Make my own thinking come out of superposition

Intro

Predictive brain presentation

Predictive brain books: Seven and a half lessons about the brain

Predictive brain books: Eye and brain

Social fallout and where to go next

Post mortem

Comments

LLMs, coding agents, benchmarks grounding

LLMs, coding agents, benchmarks grounding

Comments

flatcat audio manual

how to make flatcat audio work

flatcat audio links: slides & videos

Comments

Book review - Artificial intelligence meets natural stupidity by Drew McDermott and Computing taste by Nick Seaver

Review of Artificial intelligence meets natural stupidity by Drew McDermott, 1976

Some selected quotes

Review of Computing taste by Nick Seaver

Some selected quotes

Comments

Movies about AI and robots

Movies about AI and robots

Movies with androids

Unwatched

Other robot or AI movies

Unwatched

Not directly AI but highly related, brain interfacing and mind upload

Simulation hypothesis

Aliens and alien robot embodiment

Other

Comments

chatGPT reveals itself to be ungrounded by its comment on proprioception

A one-shot interaction with chatGPT asking for comment on proprioception

References

Comments

Book review - Unthinkable - An Extraordinary Journey Through the World's Strangest Brains by Helen Thomson

Summary review

Selected and annotated quotes

References

Comments

Book review - Survival of the Friendliest - Understanding Our Origins and Rediscovering Our Common Humanity, by Hare & Woods, 2020

Summary review

Selected and annotated quotes

References

Comments

Book review - The New Breed, What Our History with Animals Reveals about Our Future with Robots by Kate Darling

Summary review

Annotated quotes

Public imagination and story-telling

State of robotics

Robot fails

Ourselves

Movement

References

Comments

Book review - Risksavvy, Surfing uncertainty, Braving the wilderness

Review of three books all in one

Book review: Risiko by Gerd Gigerenzer

Annotated set of the top five quotes.

Book review: Surfing uncertainty by Andy Clark

Annotated set of the top five quotes.

Book review: Braving the wilderness by Brené Brown

Annotated set of the top five quotes.

References

Comments