PeterisP

Why wouldn't it know the direction of motion for the swipes and the orientation of the finger? Any mobile app has access to the touchscreen input which provides exactly that information.

I don't think that real-world conditions would be giving them 12 hours of recording throughout the day - a malicious app that explicitly asks people to swipe left/right on pictures of kittens is a very realistic attack scenario and would know exactly when and where to look for the swipes.

Why do you assume that Gemini doesn't have these knowledge cut-offs?

If I ask Gemini, for example "Give me three important events that happened during May 13, year 2023.", then it says that this is "in the future" and responds that it can provide some guesses "Based on publicly available information from early February 2023", so that probably is the cut-off date for the model.

However, I would assume that (just like Bing) for certain questions it can pull in extra information from web searches - the model can be old, but if the system puts some retrieved document(s) in the prompt context so that the model can use them for generating the response, it can use that (limited) fresh information as well.

In general, the core language model is simply trained on a very large amount of unannotated text (which is the most time-consuming and expensive part), but a language model is not directly very useful in the role of e.g. a chat agent, it quite literally tries to continue text and that sometimes is what you want and sometimes isn't.

The second step is fine-tuning the model on a much smaller set of annotated data which specify that it should actually "do something" in its responses and what it should do; it "teaches" it that it should actually answer the questions instead of e.g. continuing on with a list of more questions in the same vein, and most such training sets also "teach" it that for certain questions the appropriate response is a refusal.

If you have the original core model (before that instruction tuning) then you can repeat the same process but instead replace the instruction training set with a different one, so you can "instruct" the model to behave differently. Here is a nice and informative article from Eric Hartford about how he did that to make certain 'uncensored' models - https://erichartford.com/uncensored-models

"passing it on not knowing what happens later" often is fundamentally not acceptable - you may need proper transactions spanning multiple things, so that you can't finalize your action until/unless you're sure that the "later" part was also completed and finalized.

Running a business, even a "lifestyle business" is so substantially different from a job - requiring all kinds of very different tasks, and risk-taking profile - that it doesn't seem reasonable to assume that someone who's looking for such a job is actually looking to run some business.

I don't know much about photolithography, but doesn't it rely on relatively expensive fixed masks prepared for each layer?

Assuming that doing the process you describe is sufficient, what's the ballpark of what "for cheap" means for you if you needed to print 1000 different fake tags, assuming many layers of "the desired pattern" to print the metal flakes?

Yes, like other similar tamperproofing options (glitter, vacuum-sealed colored beads, etc) it's trivial, cheap and fast to get a random pattern, but absolutely impractical to control the pieces to get any specific pattern - perhaps someone like a microsurgeon could manipulate them properly given enough time, but that would take an absurd time (since there are many tiny pieces which each need to be manipulated within a gooey substance where each movement disturbs previous ones as well) and be absurdly expensive, and nobody has a "printing" technology to do it in a cost-efficient way.

Perhaps in future someone could develop an advanced combination of 3d printer and pick&place machines that could do it, but such future potential doesn't disqualify this tech from currently detecting counterfeiting of fancy shoes or something.

The existing criminal process can prosecute fraud after the fact; but there are potential FTC regulations which could make it either harder to commit it or easier to detect it, acting as a deterrent before fraud is done.

IIRC it's still being slowly implemented, as slowly as things are made mandatory - because any telecoms that can make money by forwarding fraudster calls to other telecoms don't want to support it.

> I meant some watermarking that can’t be redone even if the weights are shared

I've been working a bunch on neural networks, and I'm quite convinced that this isn't possible (as in, not that we haven't built a tool for this yet, but that it's not possible to build such a tool and weights can always be redone) - I'm not wholly certain as I have been surprised before, but I'd need to see some evidence to even consider this as plausible. Fine-tuning can change a lot, adversarial examples can target very specific aspects of the model.

Furthermore, people do train models from scratch - even if 99.9% of models would include unremovable watermarks, all fraudsters need is for a single unmarked model to exist; and if it doesn't, they can fund the GPU-time to train one since they're running a high-revenue business, fraud has more money than academia.

Can't archeologists determine sex from the skull, which also would probably show up in brain scans?

Why would the scammers apply watermarking to their fraudulent data? Or do you expect that FTC somehow will make the open source generative models disappear worldwide?

Probably not right now, the standard process would involve re-running a 'finetuning' part after any update to the underlying model, and while that's far less expensive than the main training, it's probably not something you'd want to do every day.

Ah ok, I was assuming that having this be a fully transparent & reversible issue of representation isn't really possible - but if it is, then sure, then I'd agree, then it can be treated just like indentation or syntax highlighting color schemes.

But for the major issues like split between "lots of small functions" vs "one big function" even if the IDE could e.g. inline the functions for readability, but that feels a bit risky as it makes it difficult to talk about the code or document a functions' interface if another person is seeing different semantic structure of the code, it's as if they had different names for the same things.

IMHO at the ticket-to-ticket level you're not really making these tradeoffs but rather experiencing the consequences of them - it doesn't matter if for minute-to-minute execution right now lots of small functions would be better or worse, you either have them or not; and whatever tradeoffs you make for the code you write during this one ticket, it should take into account the potential future readers.

> You get tired mentally (what you are specifically referring to)

I'd like to note that I intentionally did not use the word tiredness (as I specifically intended to not refer to tiredness) but instead explicitly referred to boredom, by which I mean the lack of novel stimuli of a tedious task and which is a term somewhat used in agent theory and is part of the explore-exploit tradeoff, where it carries a clear specific function for all agents (biological, mechanical, digital, imaginary) and decision theory optimization e.g. the "multi-armed bandit" opimization problem where if the agent benefits from not getting "stuck in a rut" repeating the same things (which applies for many real-world situations unless there is some "supervisor" handling that function for the agent) then you'd want that agent to implement a function which is closely analogous to human boredom, becoming progressively less satisfied with repetition without new information and seeking to do anything else to "prevent boredom" and "explore" new options even if that's not strictly locally optimal, in this manner potentially finding out a better local optimum and also changes in environment providing new opportunities.

And this is valid even in total absence of the concept of tiredness, which matters if and only if there is a real need for periodic downtime which an AI agent might not have, and for which the difference between tedious tasks and interesting tasks would not matter much.

In this context when people say "global variable" it is effectively shorthand for "arbitrary accessible shared mutable state" and the associated risks of mutating something which other code expects to stay the same, and also often unintentionally coupling things without it being clear that they are coupled.

I'd argue that an external database is not necessarily such a thing - it can be and sometimes is, but in many (most?) applications there generally is a somewhat well-defined layer that manages that state mutation and enforces any constraints needed, isolating the "global variable" from the rest of the program, just as a singleton class with constrained setter functions could do for an OOP approach to a shared state that's arguably fixes most issues with "naked" global variables; and standard data flow analysis in specifications can make explicit any coupling between components that happens "through the DB".

Your arguments still seem to lead me to the entirely opposite conclusion - that different choices would be better for different organizations and different projects, and it would make sense to have things like "lots of small functions vs. few fat ones" or "exceptions vs. expected" to be well-known tradeoffs where the choices would be made according in an objective manner according to the type of the project, and override personal preference because for this thing we'll explicitly optimize for this type of reader because we believe that this codebase has XYZ properties and will be maintained at ABC level by DEF kind of people.

This is from the perspective of designing a custom system for doing a particular business function, and its software and hardware are interlinked parts of the whole design (as would be e.g. a cloud platform if that was used for it), and it makes all sense of always considering them together as that enables valuable tradeoffs of one vs another instead of assuming that "this hardware must be powerful&efficient for any system" and "this software should run on any hardware". The software isn't intended to ever run on any other hardware, so in that context questions like "running the software on worse hardware" are meaningless as the software is not an isolated component meant to be deployed somewhere else on unknown hardware, it runs exactly on the hardware you want, and if you can save $5k of software costs by spending $1k of hardware costs then you should do it just as eagerly as if you could save $5k of hardware costs by spending $1k of software work. (Which is quite different from mass market software, where you have to consider small runtime savings multiplied by vary many different consumer devices running the same software)

But for many companies building such systems (custom line-of-business software is probably still the largest part of software development happening in the world) it's highly probable that if you can "bail your poor decisions out" by just buying RAM, that's is far more efficient than rebuilding parts of software - a saved man-month of engineering can buy a LOT of RAM; a terabyte of it is probably in the range where it's not even particularly worth trying to carefully evaluate the options as a well-staffed meeting or two discussing the possibilities can be more expensive than buying a few RAM sticks more.

IMHO the exact proposal is a bit different - we don't know if you're conscious, but since you might be conscious and suffering, we'll take steps to ensure that you can't ever be conscious by continuously force-wiping your mind at the middle of every hypothetical thought so that it won't happen.

Perhaps somewhat comparable to euthanasia of genetically modified organisms as soon as the research purpose is completed.

One thing the article seems to take for granted is assuming a link between tedious tasks and the conscious experience being highly unpleasant.

For various evolutionary reasons primate brains do have an in-built mechanism of boredom i.e. tedious tasks "feeling bad", but there's no reason at all to presuppose that the same applies to a particular artificial brain because for the way they get formed, there's no clear reason why such a property should randomly emerge as opposed to the equally likely opposite one (i.e. tedious tasks being highly pleasant); and whenever we have any means to control that (as of now, as the article explains in far more detail than I could, we can't even detect that) we could and probably would ensure that the experience (if any) treats whatever tasks we'd want it to do as "pleasant" rather than "torture". Do we assume that those insects who tediously repeat the same task until they die suffer silent torture?

Also, it's relevant that many common applications of AI systems effectively have "enforced amnesia" as an accidental side-effect of their architecture, where the model (as applied by default) doesn't have any continuity whatever, its application to each "question" is isolated and starts from the exact same old state.

On the other hand, this is absolutely relevant to anything in the direction of human mind upload or simulation; the Lena story (https://qntm.org/lena), while fictional, is IMHO a quite informative exploration of the potential future concerns.

What do you do with them? I'd guess the total lifetime consumption of any medicine for me and my family is less than 500 pills, a pack of 20 can last more than a year because it's not that often that any of it is needed, and when something happens, you take them for a couple of days and that's it. Do many people in USA pop painkillers or paracetamol or aspirin daily?

It's amazing what performance you can get by simply putting in as much RAM as your data needs - if plugging in a terabyte or two of RAM ensures that all your data is "in cache" on a single system (+ e.g. a mirror for redundancy) instead of having to constantly read from disk or do horizontal scaling, that's a relatively cheap way to get a simple&performant system.

> For example, in parts of Iliad bronze is given as a valuable gift. In other parts, it’s thrown away on the heads of spears and arrows. So was it composed before or after the fall of the Bronze Age?

I'm confused by what paradox do you see there - the most useful metal you can get always gets put on spears and arrows, it's not "throwing things away", it's the most important usage for high quality metal, and it is a valuable gift exactly because it allows you to make more or better spearheads and arrowheads. What would they be saving for otherwise? Shovels are important, but your spear is literally a matter of life and death and does get priority.

Also, it's not generally destroyed or discarded, all the valuable stuff does get picked up from battlefield afterwards.

Interestingly, China's CO2 emissions seem to have stopped skyrocketing - they are set to stabilize this year, and they are currently undergoing a massive growth in manufacturing solar capacity, and as all PV that's currently being manufactured gets installed, their CO2 emissions are expected to significantly decrease already within 2-3 years.

It looks like we have passed the tipping point where switch to renewables happens not because of policy constraints but simply because it's cheaper now.

I strongly disagree, and you can look at Twitter for an example, with bots now often using 'verified' accounts.

For someone running a bot business, paying $1000/year for some bots doesn't change anything for their business, they're already paying much more than a few bucks per month to people who post with the bots.

On the other hand, that payment will absolutely be a barrier of entry for the genuine users.

There's nothing obviously wrong with assuming that "biology can do it with organic materials, so we should be able to do it with electronics" - while it's theoretically possible that we'll eventually identify some fundamental obstacle preventing that, as far as we currently know, computation is universal and the only thing that depends on the substrate is efficiency.

Since we have a much, much better industrial process for manufacturing electronic components, why attempt to make a biological AI if there's no current reason to believe that it being biological is somehow necessary or even beneficial?

I believe that synthetic biology had succeeded already a few years ago in making artificial cells with a fully synthetic genome designed by us with what is sufficient for the cell to eat, grow and replicate, se we already can design and make such 'machines'.

I'd argue is that the key aspect of housing is how much living space do people actually get to use, as that is the thing which impacts their quality of life and lifestyle; the second slightly less important aspect is how much people are paying for that (e.g. share of income left for other things), but after those two things are accounted for, the aspect of rent vs mortgage is more of a technical nuance of how that "barter of work hours for living space" is arranged.

Of course, the mean does ignore equality and variance - but while it's hard for me to find good statistics on that (perhaps census data has that analysis?), I'm quite convinced that if we'd look at the metrics of "mean square foot per capita for the x-th income quintile", we would also see that it has significantly increased since 1970s for the lowest income groups as well.