long shot nerd query:
anyone in Los Angeles have a telescope with a Hydrogen Alpha Quantum Filter?
asking for a bird-photographer pal in #LongBeach who wants to photograph crows flying across the setting sun. TY!
(boosts welcome)
Here's the thing about progress in autonomous robotics: in the past, it has been driven almost entirely by advances in hardware (especially sensors like Hesai's new camera [1] and manipulators), sometimes by new algorithms (e.g. SLAM) or architectures (e.g. subsumption), and recently by advances in simulation capacity (e.g. NVIDIA's Omniverse).
Progress in autonomous systems has almost never been driven by advances in principled understanding or engineering of behavior design, or in performance evaluation.
The scientific method is essentially unused. Instead, we have an unending supply of "Look, ma, no hands!" papers, focused on a demonstration or simulation of an idea that an author hopes will be used by others, and almost never is.
No consistent performance evaluation or meaningful comparison, just simulation after simulation, demo after demo, "experiment" after "experiment" consisting of "it doesn't work well (or at all) before I introduce my thing, and then I introduce my thing, and now it does!"
No theory of system design. Not even rules of thumb, except those transmitted verbally, from advisor to student.
Robotics progress will be slow until this is addressed, but can be brute forced by liberal application of extreme funding.
Extreme funding is happening in robotics right now, just as researchers adopt LLMs.
Be wary of announcements of major advances in robotics in the next few years - chances are they're not transferable or robust.
1. https://www.techinasia.com/news/chinas-hesai-launches-spatial-camera-for-ai-powered-robotics
Once again, frustrated by #LongCovid support apps.
After a month of worsening symptoms, I have finally managed to establish a baseline combination of exertion and targeted rest that enables me to sleep at night.
Once I have slept well for at least three nights in a row, sustaining improved sleep becomes easier. As long as I maintain the routine and walk back any changes that worsen my sleep, my symptoms stop worsening and (over the course of months) my capacity eventually improves.
Today I hit that third day, when the improved sleep begins to have an impact.
I am not "stable" or "doing great".
I have hit the first milestone that will eventually lead to stability.
It is a milestone.
But it is not time to celebrate yet.
If I was walking along a curb or a balance beam, trying not to fall off, I would be at the point where I was about to lose my balance, and I have stopped moving forward and successfully regained my balance.
But now I have to start moving forward again.
If I was digging myself out of a hole, I would have managed to get out from under an overhang, so the dirt was no longer falling in my face and I was no longer in danger of being buried. But I wouldn't be out of the hole. I wouldn't even be a little way out of the hole. I would just have hit the "not making things worse" stage.
Stop treating this like I'm done. I need reinforcement to keep working at it, not a gold star for finishing.
I have no objection to security updates. I have minimal objections to functionality updates.
But if another app changes its fucking *interface* under the cloak of "security" I will lose it!
When my long covid symptoms are bad, ANY change to my environment triggers substantial cognitive exertion to consciously update my world model and a massive surge of adrenaline that requires intense emotional exertion to manage.
** Even a minor shift is too much. **
Do not change the spacing, font, shapes, outlines, font style, words, offsets, or placement of any and all elements.
Do not change the transparency, translucency, colors, textures, or weight of any elements.
DO NOT CHANGE THE UI WITHOUT PRIOR WARNING AND CONSENT!
This is about #Apple updating an iPhone app without warning. This morning, I opened it, there was a flash, and the data entry interface was just different enough that all my muscle memory was wrong.
I feel like a toddler having a meltdown.
#LongCovid sucks, but #UI changes make everything suck more.
Just finally went to the Omarchy website to see what the marketing hype is selling.
Now I understand the money even less.
As I get it (from looking at the site for all of ten seconds before the weird pixel art stuff made my eyes hurt), the selling point of the OS is "non-programmers should be able to install Linux and customize it by vibe-coding it into whatever they want".
If this is actually what is being sold,
what is all the fundraising money *for*? Isn't the *user* doing all the work of making it work and customizing it by feeding it prompts of what they want? They could ship total garbage and it wouldn't matter, because the *whole point* is that the user changes it and makes it work.
They literally don't need to do anything at all besides hosting something on a site.
Unless *they're* paying for the tokens for the users to modify the OS? Which seems, frankly, like a crap business model.
Or unless the OS doesn't actually work without spending tokens every time you want it to do something? Which is a great (albeit evil) business model for the people selling tokens, and completely unusable for OS customers.
Now I am *more* confused.
I just learned that someone I know is starting their PhD developing an agentic LLM for preventing a medical condition.
And I know many PhDs are driven by "what grants the PI was able to get funded".
But ... WTF does that even mean?
Let's say we're talking about a chronic illness that is impacted by patient actions - cardiac health and exercise/diet, for example.
Are they trying to build guardrails to use the psychosis-generating, skill-degrading machine's propensity for creating obsessive behavior to "nudge" patients into good habits? Or are they trying to build a health advisor that only gives good advice?
Either way, I foresee nothing but ethical and engineering problems built in to the research on a scale I had previously thought reserved for lethal weapons systems and in-home elder care robots.
Another entry in the decline of #IEEE.
Until now, my IEEE membership has come with an @ieee.org email account. For decades, it was an alias I could forward wherever I wanted. Then it was an account I could forward. Then it was some kind of Google thing (never really understood that) and I had to log in to google to send emails to mailing lists.
Today I was informed that they are separating the email accounts from the main subscription and charging $40/year for them. There is no option for an alias.
There are a lot of other google-related things that fee includes, none of which I use, and none of which I even knew were being provided.
All of my ieee-related services - access to collaboration tools, membership in technical committees, journal and society subscriptions, access to Overleaf - is all mediated through this account. They have literally spent the last four years forcing people who aren't members to get these emails in order to use freely accessible online services.
I am so frustrated and I'm struggling to know how to get this across to them, since the last email I sent to their help desk got a "you must give us personal information to prove your humanity before we will respond" auto-reply.
I *hate* this timeline.
New paper from @patientled!
If you run different clustering algorithms against the same set of patient data, you don't get the same sub-groups!
The abstract conclusion:
"The strong dependence of patient clusters on algorithm choice suggests that single-method Long COVID phenotyping may produce incomplete or unstable subgroup definitions. Clustering methods may impose artificial boundaries on a smoothly varying symptom landscape, especially in studies capturing fewer symptoms. Phenotyping efforts should assess clustering robustness and avoid overinterpreting single-method results. Our multi-method analysis highlights the importance of considering the full breadth of patient symptoms when evaluating treatments."
https://academic.oup.com/ooim/article/7/1/iqag010/8707854
#LongCovid #Phenotype #MachineLearning #UnsupervisedClustering
This is demonstrably false. The emergent pile-constructing behavior is only intended by the designer, not by the robot. The *robot* intends to pick things up and put them down, to move in a random direction. It has no awareness of the concept of a "pile" (it can literally only represent a maximum of two objects at a time internally - the one on the ground and the one in its gripper). It has no sense of place (there is only "here" and "next place"). It has no way of knowing where objects may or may not be found. It does not have intent *at the level of* "pile".
Just like LLMs do not have intent *at the level of* consciousness. LLMs have intent at the level of tokens.
We could argue that LLMs exhibit emergent behavior at the level of consciousness, but that presupposes that consciousness in humans is the product of words, that the internal function that creates consciousness in humans is the *product* of language, not that one *purpose* of language is to convey the results of consciousness.
It seems much more plausible to me that consciousness is the result of the advanced modeling of not only other agents in the world, but of oneself, that is needed to predict the behavior of other humans. This could happen entirely independently of language.
Until LLMs are demonstrating advanced predictive modeling and representation of not only others' behavior but of their own behavior (not just words), I will be reluctant to consider consciousness as an explanation.
2/3
There is a fundamental flaw in the idea that consciousness can be defined "functionally" (i.e. "it is what it does").
Consciousness is a purely internal state - we infer it in others through communication, but it is fundamentally something someone experiences about themself, not something defined by how other people perceive them.
Because it is an internally defined property, it *cannot* be defined on the basis of "what it does". It must be defined, at least in part, on the basis of "what it knows".
To take a simple example from robotics:
I have a robot with one gripper and two sensors that moves through its environment. One sensor (A) tells the robot whether its gripper is full or empty. The other sensor (B) tells the robot whether or not it is in the presence of something its gripper can pick up (an object, P).
The robot has three rules.
1. If no P, move in a random direction.
2. If gripper is empty and P sensed, pick up object and then move in a random direction.
3. If gripper is full and P sensed, drop object and then move in a random direction.
This robot has no concept of anything in its world except P/not P, gripper full/empty, and "move in a random direction".
Yet, externally, the effect of these rules is that the robot builds piles of objects [1].
Attributing consciousness to an LLM is the equivalent of saying our robot *intends* to build piles of objects.
1/3
Triggering article: https://www.slowboring.com/p/nobody-knows-what-theyre-talking
Outstanding explanation of the problems with Tuller's Wired article on Long Covid: https://darthfoo.substack.com/p/the-patient-who-isnt-supposed-to
Longshot, but looking to connect to someone who took Eng 101 at Arizona State University in Spring or Fall of 2020 who received the following email (or one like it). Boosts greatly appreciated.
#Caturday blep!
RE: https://fediscience.org/@oatp/116489885682133529
This paper is *advocating for* LLMs to be given access to paywalled papers as training data while actual human researchers outside academia have to go hat-in-hand to authors begging for preprints, because LLMs not having access to high quality human-generated work will be bad for the LLMs.
I can't even with this timeline.
My Verification of Autonomous Systems book is finally out!
In addition to contributed chapters on theoretical and practical aspects of verifying autonomy*, the editors contributed an introduction on stakeholder perspectives and a chapter on the specific challenges autonomy introduces to verification.
Each chapter is cross-referenced to e challenges it addresses in a handy table to help readers find the information most useful to them.
I'm really proud of all our authors! So many thanks to everyone who contributed^, especially my co-editors, Mae Seto, Don Sofge, and John Sustersic!
E-book now available for download and hard copy coming soon!
* This book is about verifying autonomous embodied robots, not LLMs or AI models.
^ No AI was used in the generation of this book's content.
Link: https://link.springer.com/book/10.1007/978-3-031-88546-4
#Robotics #AutonomousSystems #Verification #VerificationOfAutonomy #Robots #AutonomousRobots #CyberPhysicalSystems #RobotTesting #FormalMethods #Modeling #Simulation
My first #Mosstodon!
Patient Tips for LC Care Providers:
When a #LongCovid patient says "X makes my long covid symptoms worse,"
DO NOT SAY "What symptoms?"
If we're struggling enough to seek care that won't exacerbate our symptoms, we're carrying lists of 60+ symptoms, one of which is likely to be a combination of cognitive dysfunction and inability to respond to stress.
Asking "What symptoms?" triggers us to (a) forget to mention the important ones like PEM and cognitive impairment and (b) only mention the 3 that come to mind first, missing at least 5 symptoms that are relevant to whatever course of treatment you're about to recommend.
Instead, work through the body, asking what symptoms worsen overall and in each system.
