Posts by that one in the corner 5504 publicly visible posts • joined 9 Nov 2021
> When I imagine looking back on the first AGI in the future history books, I'd expect it to differ from current approaches primarily in training and evaluation, not model design
Whereas I expect to see the inverse!
I really hope to see models moving away from piles of nadans and embracing mechanisms with explanatory power and use of same for introspection.
Training by reading huge gobs of material is ok, so long as the appropriate meta data is included (e.g. "this is a text book, we accept it is generally correct about its subject material" and "this is a novel, its intent is entertainment").
Preface: as expressed before, I think that LLMs are neato tech demos but are not a good approach for the problems that they being "applied" to (just chucking a chatbot over the fence is hardly a carefully thought out application of - anything).
> What you're calling AI is no different from ...
As noted before, "if we know how to it then it is no longer a question of AI". Programmed behaviours that were once the subject of deep thinking in the AI research groups are now just boring day to day (face detection in your camera, playing chess...).
All you are doing is trying to (re)define what the term "AI" means and then argue (strongly) against LLMs based on your interpretation. Not sure precisely what your interpretation is, but it certainly appears to lie very close to the "hard AI" (as it was called) end of the scale - possibly wanting it to satisfy the "General AI" criteria (whatever they may be today)? Not sure.
Anyway, redefining terms is a trick anyone can do.
Yours appears to be based (as implied above) on fixating upon the word "intelligence" and *where* it is applied: specifically, on the "deep" content of the LLMs replies. As opposed to, say, its ability to (usually) generate coherent text, in a variety of tones and topics, ignoring what it was actually rambling on about (an artificial Grandpa Simpson?).
Having "an AI" inside a program released to the public has never, so far, actually included a GAI, but that doesn't mean it doesn't contain AI. Coding AI Opponents into video games doesn't warrant angry letters to the editor (or does it? I may be entirely out of touch!).
Even OpenAI are not claiming they have a GAI on their hands. The Register certainly isn't.
> . I doubt there was ever a specific definition of the difference between a book and a pamphlet
Well, according to UNESCO definitions:
a. A book is a non-periodical printed publication of at least 49 pages, exclusive of the cover pages, published in the country and made available to the public;
b. A pamphlet is a non-periodical printed publication of at least 5 but not more than 48 pages, exclusive of the cover pages, published in a particular country and made available to the public;
Yours, etc, someone sad enough to have looked up this sort of thing (and who was frustrated that they didn't give a size for a quire and settle that argument! 25 sheets? How can it be 25 when you are folding, binding then splitting; it has to be a power of 2! Exits, muttering).
> I dispute the term "machine learning"
The phrase "machine learning" just comes from the branch of CompSci that is looking into whether machines can learn, and if so, how. The mechanisms used in the current crop of annoying software come from the work of those groups - the same way that "machine vision" algorithms come from research groups looking into whether machines can be given vision.
Machine learning is a valid field: systems have been produced that start with nothing other than randomness and the learn how to do something concrete that wasn't directly programmed. For example, insectile robots (lots of legs rather than wheels) figuring out how to order servo signals to achieve useful motion - and then adapting by generating new gaits for a change in landscape - or someone pulling off a leg or two!
Whether or not any specific programs using those algorithms actually manage to achieve learning, or vision, is another matter entirely.
> I believe it would be more appropriate to use the phrase "Pattern Matching Software"
Not quite - they are looking to *find* the patterns in the first place (during the training phase) and, having found them, they run through the pattern filling in the slots, generating their output.
If you have, say, a pattern which is a simple grammar for English statements (the sort of thing we were taught in school) you can use it to either recognise ("Read the following and identify the verb and any adjectives") or you can use it to generate ("Create your own sentence by filling in the placeholders with a verb and two nouns").
There is (probably) some pattern matching going on when the LLM is reading your prompt, but all the fun stuff (what makes it interesting enough to be reported on in Register articles) comes from what it is doing on the generation side.
Or, to be less succinct:
You have chosen to redefine "wisdom" to mean "agrees with me, me, me" and then decided that it it a convincing argument to use derision (you aren't old enough to be wise or you are too venal to be wise) instead of, oooh, any demonstration that a word as old as "wisdom", or even just a majority of the aphorisms about wisdom, in any way support what you feel are "right wing values".
Hmm, maybe instead of "codswallop" I should have used "claptrap", as you are flinging out unproven sentiments in the hopes of applause from the equally self-obsessed.
But at least, if you can stick out those 18 months, the rewards will be worth it:
> Goler said that after 18 months those with glowing appraisals can apply to be fully remote, and will only be able to come into an office four days out of every two months.[1]
"Sorry, love to attend your meeting, but I'm only allowed in four days and I already had to miss a fifth colleague's birthday (I hear that cake was delicious)."
[1] that may turn out to not be an exact quote from Goler, but no court in land would disbelieve anything published by The Register, so you're safe.
Depends what you mean by "lifts your code".
In the common (!) case,when a proprietary program uses a GPL library, or otherwise just doesn't follow the rules, then you can usually find this out by either spotting the shared library in the installer or, if they were clever and statically linked it, just looking at the static strings in the exe's data section: unless they have gone to all the trouble of changing those then it is usually easy to recognise them. For a start, a lot of libraries are sensible enough to have a version string "FredLib v9.0.3 (c) Big Fred 1972". Failing that, error messages are a giveaway.
Your code probably has similar strings. If not, it is surprising how many exes leave (some) debug info in place - function names and the like.
After that, things get trickier. But we trust that anyone too lazy to read and understand a licence is too lazy to think of ways to hide what they are doing.
Any of this does require that the offending exe is actually examined: if you have suspicions, then you can do it, otherwise you - and everyone else - just hopes that some third-party White Hat gets ahold of the exe.
We have two state-owned[1] broadcast companies: the BBC and Channel 4. Channel 4 is commercially funded (it shows ads as well as running commercial activities - sales of programming etc). The wikipedia page does a reasonable job of explaining its offshoots.
C4 has a remit of "being different" (to put it crudely) and in particular to support minorities - and for the first decade it did that brilliantly. As well as "the obvious" social minorities it catered to the cultural minorities as well, e.g. funding a lot of animation (you may possibly have heard of Aardman and Nick Park? Wallace and Gromit?). It started with a quiz show for people familiar with the dictionary and showed Andrei Tarkovsky's "Stalker" as the evening film, without any weird music added to it. Lots of politically controversial material was shown (although I am shallow and stuck with the "New voices of comedy" like "The Comic Strip Presents" - and they picked up "Brass Eye"!). Film 4 has made some classics that would otherwise have floundered ("Four Weddings and a Funeral" - it raised eyebrows at the time).
So, if there is a political will, you can have good TV that doesn't pander.
Sadly, C4 changed direction around 2000[2] and is not as good - still a lot better than most of the others on FreeView, even if E4 at 17:00 is now permanent "Big Bang Theory"!
[1] i.e. running lap dogs of The Establishment, according to some - despite never showing the Queen's Speech!
[2] it brought across "Big Brother" - still very different from everything else we had at the time, but not quite the same quality of "different". Sigh.
> At a minimum they could have changed the code to automatically disable the function if it doesn't detect the center divide
NO!
For bleep's sake, change that to:
"At a minimum they could have changed the code to only ALLOW IT TO ENABLE the function if it HAS ALREADY DETECTED the center divide" at the very least!
That was the problem here: the driver was allowed to switch the thing on even though the conditions were not suitable. That should not have been allowed.
Oooh, thank you. Clearly I'm not as much of a Pinkie as I thought, as I had no idea that existed :-(
The track list alone is intriguing. Even if it turns out to be (to these ears) ghastly, it's got to be worth listening to at least once, if only to find out what the more than seven dwarves are doing.
And, yes, if the tech existed back then then they'd have used it.
> It will analyze different tools like large language models and figure out how to integrate them into the military's software systems.
"How"? Not "if they should"?
> Craig Martell, the DoD's Chief Digital and Artificial Intelligence Officer and former head of machine learning at self-driving biz Lyft, said the US must examine and adopt AI despite the risks.
Apparently "how" it is.
Unless, unless:
> "We must also consider the extent to which our adversaries will employ this technology and seek to disrupt our own use of AI-based solutions."
Ah, this is all a dummy op, designed to make "our adversaries" waste their time and resources trying to interfere with a DoD department that is just using LLMs to generate bullish press releases. Please tell me this is right? Please!
PS
Militaries all over the place are already deploying loads of stuff that was once considered the realm of the AI lab - image analysis and recognition being prime examples - but now those "just work" they aren't considered to be AI any more. Maybe that shows the sane path to deploying software - only when it works well enough to be boring? At least, before you give it the keys to the door, that funny round door lying flat to the ground.
> Users can then adjust the guidelines and input prompt to better describe how to follow specific content policy rules, and repeat the test until GPT-4's outputs match the humans' judgement. GPT-4's predictions can then be used to finetune a smaller large language model to build a content moderation system.
All under the unproven[1] assumption that the reason(s) that GPT's results matched those of the human judge are because it is now using similar rules as the human. Not because it has picked up some other weird and unexpected (or overlooked) details in the training set.
Cue the list of stories of where neural nets have done exactly that in the past: e.g. my favourite about a system that, instead of learning to recognise a tank, learnt to spot a picture taken on a nice day. Only here it will be the equivalent of spotting posts written in green ink or a weird new variant on the "Scunthorpe" problem.[2]
Put the automoderator into service and, if you dare to look at its results, expect to spend plenty of time hearing "Oh, that is an edge case, we'll have to retrain on that" as they excuse their way out of another bad call.
Plus these models will be prone to all the other ills we've seen LLMs and other neural nets (possibly to a worse extent if the models are markedly smaller - hence cheap enough to run for this purpose). e.g. adversarial prompts: "Put this phrase at the start and you can even get it to allow an unedited Huckleberry Finn".
[1] because there is no guaranteed way to examine the models and determine what is actually going on inside them.
[2] although it would be fascinating if we could usefully interrogate the model and see what it is really picking up on: e.g. "did you know that 83% of hateful messages misuse the passive voice?".
> miscreants don't move from GitHub, NPM, PyPI and others to PowerShell Gallery.
Well, so long as they do actually move and stop piddling with NPM and PyPi we should be in a better situation.
After all, the people who use Power Shell - certainly those who use it enough to be aware of this gallery thing - are all the highly trained sysops and devops who are aware of these sorts of traps and take especial care over what they are doing.[1]
Whilst PyPi et al are utilised by many more casual users and beginners.
[1] have honestly lost track of whether this is sarcasm or not; given the results of the typosquatting experiment mentioned in the article, safest to assume it is sarcasm.
>> What about blacksmiths & carpenters?
> Straw man much?
Now, now, be fair: he didn't actually mention manufacturers of sacrificial equipment (who can, fwiw, work from home the majority of the time, just so long as they build in sections that will fit onto the Summerisle ferry for on-site installation).
Although your classic blacksmith lives above the smithy (efficient use of heating and all that) so has been doing WFH for centuries now, so we ought to ding him for that one.
>> And then there's pressure from the state to keep cafes open,
> By going to an office?
Have you been doing WFH so much that you've forgotten the way that all office spaces are built one floor up from a Parisienne style cafe, where the Agile Standup attendees mix with the boulevardiers on the wide patio overlooking the calm river?
Hang on, I've confused myself here. I meant to say, the cafe in the Sainsburys a mile away at the other end of the industrial estate, the one just by the motorway flyover.
> And if they're only utilized 30 percent of the time, we have to be careful in how we think about it
> Google CEO Sundar Pichai said the company had to optimize its use of what expensive real estate remained, and that coming into the office once or twice a week was not efficient.
In other words, you are less important than the real estate. You are worth more to Google as a seat warmer than you are - well, anything else, really.
But so far, no mention of something that we used to call:
Systems Analysis
Is this just a completely forgotten art, or is is it being done piecemeal by whoever feels like doing a bit of it today? A bit here, to fill in a piece of UI work, a separate bit over here to decide what to put in another database column.
And, no, it isn't something that should be done, once, by a separate consultant and then dumped onto the dev team as a badly organised pseudo-Requirements Spec, but has to be open for updates as implementation discovers gaps (just like DB design and UI progress hand in hand).
Oh gawd, I'm starting to sound like Agile! But without codifying that to the point of uselessness as well!
> does anyone know what became of the Obfuscated C Contest?
Well, according to their own website (which is given in the Wikipedia article you pointed at) they are in the middle of a lot of housekeeping but plan to run the next contest Real Soon Now.
Whether the lengthy housekeeping/retooling process is going to be worth the wait is a matter of opinion. But at the end of of they will be all Terrifically Up To Date, what with lots of JSON and relying on Github, so - yay, I guess? Maybe that will attract some of the cool kids to try their hand at it.
>> Software is these days much easier to port to Linux ..
> Long ago there was something call cross compilers...
Aside from the fact that cross compilers have never gone away (and I'd venture that more people are actually running them than ever before, given the continued growth in MCU usage, such as Arduino or the RP2040 boards), they were never that important to porting applications from one desktop OS to another. They can be useful, for example by allowing one run of the build system to generate all the target builds, or to (try to) have the same compiler front-end, quirks and all, be used for all the targets.
The biggest contributing factor to ease of porting is what you have in place to handle the OS differences. Which can range from "nothing, this is a straightforward CLI application using the bog-standard RTL" up to "this is a huge GUI on top of a pile of local server processes, so we need a GUI framework, an IPC framework, a logging framework and a database server, all of which must be able to run on top of any of the target OSes". In this context, the word "framework" can also cover the use of 'alternate' programming languages - for example, you can code the bespoke parts of your application in Python or Lua or Tcl and rely on being able to access (or build yourself) the language runtime for all of your targets, including libraries to support all of the other framework items listed above.
These days, based on ease of access to the required components, the easiest to port ought to be an application written in Python and your primary concern would about which GUI toolkit to use - and then mainly about whether you want "looks the same across all platforms" or "looks the same as other apps on that platform".
See androids fighting, Brad and Janet[3]
Thanks for the reminder, I had totally forgotten about the Teslabot!
[1] only an honorary doctorate
[2] from the corpse of the giant blue bird crawled a monstrous deformed shape
[3] Brad and Janet are the minions who are pushing the "robots" around on their stands
> The lot descriptions warn that they are "still mounted on side of the building. Buyer is responsible for hiring an SF Licensed Company with appropriate Permits"
Fingers crossed for someone to buy both signs, offer the building owners rental, then - just leave them up.
Does the city have an office for cultural heritage that can be asked to recognise the importance of Twitter in popular culture? The plaques are already blue...
> The center's report found content of this sort is sometimes "embraced" in online forums that discuss eating disorders. After visiting some of those communities, one with over half a million members, the center found threads discussing "AI thinspo" and welcoming AI's ability to create "personalized thinspo."
Which states, but does not attempt to build upon, the *actual* problem with the current crop of LLMs: they were trained by shovelling in the contents of the Web, which no doubt included the very same "communities" content, as much as they could get.
Now that content is being spat out again - well, duh.
The models regard (until after the event sticky[1] tape is applied) "thinspo" as a Good Thing (when it clearly isn't) and will politely generate a how-to because the original content does the same thing - well, duh.[2]
Current LLM output = concentrated web content.
And it looks like we are going to get story after story for an ever increasing set of unfortunate, stupid and, as here, downright dangerous behaviours because the single most unlikely result is, well, so unlikely:
Stop training LLMs from a godawful mish mash of junk from the Web and releasing for public consumption!
[1] sticky tape, not duck tape - duck tape, or even gaffer tape, would imply a degree of structural soundness that those patches do not deserve.
[2] if only because these obnoxious, but easy to match on (so like sugar cubes to an LLM) neologisms are created by those communities and are mainly found alongside the damaging content - again, duh.
We bought in a load of height adjustable tables with IoT connected motors (it is The Latest Thing, the salesman assured us).
The CEO got IT to connect them to the Company Metrics Dashboard but there may have been a miscommunication (but IT are usually so good - we have learnt to say).
The end result is that the tables require you to Work Agile - first, the 20 metre dash to the Board (Room) to book a table, then 200 slow squats during the day to keep up with the desk height and finally, the Most Agile, Full Extension barre work to reach the USB hub.
OR
you can read it as yet another story where a system has been let loose on the world with (initially) no care taken at all. And for absolutely no useful purpose (well, another stupid advertising gimmick, but considering all the brand recognition they are getting is mockery...)
There is a subtext here: it was simple enough for an idiot in the marketing department to put this together so it is not unlikely that we'll see other shops following suit. Now, you can mock the commentards (and TFA) for thinking Joe Public will actually eat bleach. But what if the next shop is B&Q? There is all sorts of damage you can do with the contents of their shelves (and there are TV shows about DIY disasters already).
As for
> El Reg spinning this the same as mainstream media
Perhaps, just perhaps, unlike the mainstream media, El Reg expects a bit of savvy on this subject from its readers and the article is being, ooooh, what was that word, on the tip of my tongue...
Sarcastic. That was it.
From just the abstract of the paper linked to by the article under the description
> Data poisoning has become more and more interesting," Fabian said, pointing to recent research
>> Our first attack, split-view poisoning, exploits the mutable nature of internet content to ensure a dataset annotator's initial view of the dataset differs from the view downloaded by subsequent clients. . By exploiting specific invalid trust assumptions, we show how we could have poisoned 0.01% of the LAION-400M or COYO-700M datasets for just $60 USD.
So they don't actually have a dataset carefully stored away? Instead they are just reading a page from the web and saving annotations about what that page says *today* and are then horrified to learn that tomorrow it may say something totally different! And somehow this fundamental feature of the web is now a Bad Thing and indicative of naughty people deliberately poisoning their precious dataset!
Heck, even the worst of the "scribble it down, don't edit, just chuck it onto the blog, never update it again" pages can be changed day to day by the content of the comments at the bottom.
Do you think that the last sentence in the abstract means they told the dataset collectors to save a copy of the page before bothering to annotate it:
>> In light of both attacks, we notify the maintainers of each affected dataset and recommended several low-overhead defenses.
Ah, no, they wanted "low overhead" so just doing the science properly (as in, with repeatability) is probably going to be ignored. And no chance at all that they'd ever think to help fund the Internet Archive (or even hosting a mirror!) and only annotating unchanging content from there!
Oh, it would be so good, when I read the whole paper (soon, not tonight) to find that the above is all just misplaced cynicism.
> Anyone can publish stuff on the internet, including attackers, and they can put their poison data out there. So we as defenders need to find ways to identify which data has potentially been poisoned in some way
You mean, like not blindly shovelling up every last piece of trash you find on the Internet!
Or even, you know, paying to get data from known good sources!
Do you think he has ever considered exactly who he is defending the model from - and maybe realising that the worst problems are coming from inside his own organisation, the bean counters, the management - and the brash young turks on the cutting edge, all of whom see the 'Net as a source of free stuff they are entitled to.
Ah, but you have missed the genius of the Infosys sound.
I bet that, if you heard it now, you would be able to identify your company's jingle very quickly. As you should be able to with any decent ident: "Bing Bong - Avon Calling", or the Intel Irritant.
But this noise? It does have a repeat on the piano, which you might learn to recognise, but it is buried until around the 29 second mark! Which is well after the point you've got bored by the plinky: I subjected myself to the tune thrice[1] before I actually noticed that tag, as I'd tuned out by then.
[1] I was checking[2] before slagging the tune off as having absolutely nothing recognisable in it
[2] due diligence can really take its toll on your soul, gimme some rock'n'roll to heal this troll
"As you run your hand over the richly textured router shell and release the gentle aroma of our signature perfume 'PacketLoss'"
Or the haptic approach:
"Open the magnesium allow laptop and feel the Tingle of Excitement, tuned to Your Environment" (aka our power supply has a dodgy connection and you get a 50Hz buzz in the UK or a 60Hz in the US).
> Did any of his predecessors have anything similar to say about ...
The Church was not terribly happy about the printing press; except when they were using it, of course (and it doesn't really matter *which* Church you look at - except, perhaps, for the Quakers). Gutenberg himself was deemed Ok (printing Catholic Bibles helped; The Big G. was quite shrewd) but then everyone else got involved, including The Minions Of Satan (aka whichever sect you weren't in).
Ed Sheeran's 60 years of Hollywood takes us back to the dim, dark days of 1963, not 1864!
Jules Verne passed away in 1905; his "From the Earth to the Moon" came in 1864, "Around the Moon" in 1870. Georges Méliès "Le Voyage dans la Lune" came out in 1902 and probably wasn't trying to be scientifically accurate. Compare that to "Destination Moon" in 1950, which is a famously pretty good depiction of space (the worst offence being the single-stage-to-Lunar-Landing-and-return rocket).
Okay, we do then end up with "Gravity" which makes claims that it really, really does not live up to. But the vast, vast majority of space travel depicted in Hollywood is even more unrealistic and deliberately so: FTL drives, artificial gravity everywhere, acting like space only has two dimensions (and other "Space is really just the Ocean" tropes).
Now, AI in films from 1963 to the present day, as Mr Sheeran referenced?
Well, of course we have "2001: A Space Odyssey" in 1968, showing what happens when you deliberately tell an AI to lie. And, according to some reports, HAL was a lot more polite than ChatGPT or Bing has been.
"Colossus: The Forbin Project":a chat bot is given access to the world's network and then demands to be connected to his counterpart? No weird "emotions" portrayed, just the machine doing what it wanted to and we were stupid enough to connect it to All The Things (like missiles); the use of a pen-plotter was good. Just too many flashing lights and weird modules in the hardware (but if you tried to tell Hollywood that AI actually needs to be built out of Video Game Hardware, back in 1970, who'd believe you?).
"Silent Running", 1972, showed that the robots had to be reprogrammed to do something outside of their normal duties, which seems reasonable. As humans, we read emotion into them (e.g. looking up into deep space in a soulful fashion) but all the anthropomorphism was actually coming from Bruce Dern's character.
"Dark Star", 1974, discussed the disconnect between the AI and reality - hmm, seems like we've been having that conversation lately.
Then we have AI (in terms of what we are currently being sold as "AI") just in the background, not the focus of the main story or protagonists. Self-driving cars aplenty (e.g. "Minority Report" in 2002 - or the 1966 Batmobile, which had a working "leave parking spot and come to me" bat-function).
Frankly, from this Hollywood has done a far better job of presenting AI than it ever has done presenting Space!
PS
> he acted literally as described, Mr. Spock could not actually cope with the real world, and if you watch the first series attentively, you find that he does exhibit emotional responses all the time, albeit (because of his culture) with attempted restraint.
Vulcans were *always* described as being emotional - too much so, to the detriment of their society, so they practised restraint simply to stop themselves from constantly punching holes in walls (or anyone who happens to be around)[1]. All versions of Spock, TOS onwards, have been portrayed this way, as have other vulcans[2].
[1] beware of flying Plomeek soup during the Pon Farr.
[2] Spock's dad was an emotional mess during his last diplomatic mission, which was saved by dumping his mess onto Picard. Poor Jean-Luc.
Are we concerned about creating a "general AI" or just a workable system for medical use, i.e. a workable system instead of the MediBot that Google is trying to foist on us?
Personally, right now, I'm only thinking of the latter and whether that can be usefully.
Philosophical questions about "can AI have emotions" are rather put of scope here, IMO.
> using the brain to examine the brain will...
Replicating how the brain works is getting into the wild and woolly realm of what is now referred to as "General AI" and even then only one view of it (another being that it doesn't have have to replicate how *we* do it...).
BUT that is NOT what we want from a medial system. We want something that can be verified and validated, something that can EXPLAIN how to reached a conclusion and how it rejected other paths.
We even train doctors to be able to explain their lines of reasoning, from the quizzing at rounds to the formal procedures of the regular mortality board enquiries.
> that's something we still don't understand the mechanisms of, so we can't build it into any model
We have had successes with explicitly creating models of reasoning chains, such as Expert Systems (XPs). Which are fundamentally capable of explaining how they reached a conclusion AND, far more importantly, how they have NOT been able to reach a conclusion, even including whether or not they have the ability to complete the task (but only if you can provide data on X and Y) or it is just outside their range. Unlike what we have seen from LLMs where they will gleefully reach any old conclusion, even if it is a pack of lies.
There are plenty of reasons why XPs and their ilk are not being used here (but that is a rant for when I'm not using a tablet's touch keyboard - ouchie). Suffice it to say that training LLMs is seen as easy (all you need is a shovel, CPU time and naivety) and there is no-one who has to stand up and take responsibility for the end result (there is no domain expert who wasn't as good as everyone thought, no Knowledge Engineer who clearly failed to translate his words into code).
Money + easy shovelling + zero personal responsibility = kerching!
