Just what is autonomy (in an automated system like an embodied AI, Robot or even Human Proxy)?

 Something or someone (a proxybot) carries out an action "on behalf of someone" at some distance in space and time from the person issuing the instructions. They were given instructions on what to do, including contingencies for varying circumstances. What level of autonomy does this represent, if the proxybot can vary what they do if the circumstances are not precisely any of those foreseen? 

(If this, then that, otherwise...)

Of course, the proxybot programmer could try to foresee the universe of possibilities, or could include "failsafe" alternates, or describe overall / overarching principles for decision making in the presence of novel situations (ethical envelopes, so to speak).

But the instrument is still an instrument, and not really autonomous. It is an agent of the orginator. Just because it is removed in space and/or time does not reeuce the agency of the programmer, surely? Unless the programmer and proxybut "agree" to handover agency: but what would such a handover look like? how would you know?

scholastic parrots

 having a conversation in the Turing with my mentor and discussing whether LLM is just AGI because AGI is "just" statistics, and also "just passed the Turing Test"....and we both observed that most interactions we have with other GIs (human intelligences) are pretty dumb.

so my main concern with this is the usual repetition of the Theodore Sturgeon comment about most SF being pretty terrible, and he responded with "most everything is pretty terrible". Intelligence is rare - most GIs can exhibit it, but only do so very occasionally, as intelligence is really not often very useful - habit is much more useful (thinking fast, rather than slow, is a survival trait according to kahenman and tverski).

so like many things, smartness is zipf/heavy tailed - 

the title of this entry refers to scholarly works - most papers are cited less than once. A few papers get tens of thousands of citations.

So you train an LLM on the common crawl, or on the library of congress, and the majority of stuff you've trained it on isn't even second class, it is just variations of the same thing.

This isn't model collapse - this is an accurate recording of a model of what most people's visible output looks like. Dim, dumb, and dumber. So what?

Well, going back to the Turing test, if you, an Average Joe, pick an LLM at random, prompt it with some average prompts and compare it to the average GI, you will unsurprisingly conclude the LLM has passed the turing test.

But what if you had Alan Turing (assuming still alive) at the other end of the GI teletype, I ask? and what if you got Shakespeare and Marie Curie and Onara O'Neil to ask some questions of it and the LLM.

Then I suspect you'd find your LLM was a miserable failure, like the rest of us. Except that every now and then, we rise to the occasion and actually engage our brains, which it cannot do.

In-network processing - do we ever really need it?

 We've looked at this problem from several sides now - to solve the "incast", to do aggregation for map/reduce or any federated learning platform, to aggregate acknowledgements for PGM.

When we say "in-network", we're talking about in-switch processing - borrowing resources from the poor P4 switch to store and process multiple application layer packets worth of stuff, so that only one actual packet (or at least a lot less) needs to be sent on its way.

So how about we compare with multicast (in network copying) and its (largely) replacement by CDNs/overlays.

Key point is branches in the net - this is where the "implosion" (for incast) or "explosion" (for multicast) happens:

So do we have a server nearby? Or can we just put one there (or just connect one there?

Answer is (for multicast yes:

netflix/pops in wide area - use distribution trree to all pops, and caches

So in data center: 

use servers, not switches and build sink forest of trees

clos system, connect servers to local switch, top of rack, and spine switch/server...then for servers at some level, use a node at the next level up as aggregation server (note Clos even has redundancy so this will survive edge/switch outages)

Unseeing like a State

 Just read Seeing like a state, by James C. Scott.  Incredible scope and vision for what is often, but not always) wrong with "tech" led solutions - though in a very broad sense. - looks at imposition of regularised/normalized villages, farming, transport, city structures and so on, especially by "developed" world on the (frequently) completely inappropriate contexts of colonies but also post colonial, self imposted. From russian collective farms, to modernist cities like Brasilia, from mono-culture farming to single-minded, wrong-headed cultural impositions, an amazing read!

It basically makes it pretty obvious why the following stuff happens:-

Tim Wu's Eyeball Bandits

Ian Hislop's Fake News

Doctorow's Drain Overflow

Basically, the Internet users are the hunters and gatherers that just got fenced in and collectively farmed, like ants. Grate.

boxing clever with AI

 There was this AI creative challenge where the you had to figure out things to do with 4 objects, as follows:

A box, a candle, a pencil and a rope

Here's my 3 proposals:

1. Draw a still life on the box of the candle and the rope so that it looks like 3D (i.e. draw on all 6 sides of the cube, with the pencil)

2. make a clock out of setting fire to the candle, the rope and the pencil - they will burn at different rates and you could mark out the seconds, minutes and hours with box lengths, then sit on the box, passing time

3. Have a boxing match between the pencil and the candle, in a ring made by the rope.

dangerous AI piffle...

 So what's a dangerous model?

The famous equation, E=mc^2 is dangerous - it tells you about nuclear power, but it tells you about A-bombs too.

This famous molecular structure dangerous too - it tells you about DNA damage, but it tells you about eugenics too.

[picture credit By Zephyris, CC BY-SA 3.0, https://commons.wikimedia.org/w/index.php?curid=6285050]

So we had Pugwash and Asilomar, to convene consensus not to work on A bombs and not to work on recombinant DNA. Another example - the regulator has just approved exploiting the RosebankUK oilfield, despite that solar and wind power are now cheaper than fossil fuel, and that COP26 made some pretty clear recommendations about not heating the planet (or losing biodiversity) any more.

What would a similar convention look like for AI? Are we tallking about not using Generative AI (LLMs, Stable Diffusion etc) to create misinformation? really? seriously? that's too late - we didn't need that tech to flood the internet and social media with effectively infinite amounts of nonsense.

So what would be actually bad? well, a non explainable AI that was used to model climate interventions and led to false confidence about (say) some Geo-Engineering project, that made things worse than doing nothing. That would be bad. Systems that could be inverted to reveal all our personal data. That would be bad. Sytems that were insecure and could be hacked to break all the critical infrastructure (power, water, transportation, etc) - that would be bad. So the list of things to fix isn't new - it is the same old things, just applied to AI like they should have been applied to all our tech (clean energy, conserving bio-diversity, building safe resilient critical infrastructures, verifiable software, just like aircraft designs etc etc)...

n.b. the trivial Excel error that led to UK decision to impose austerity, that was exactly incorrect:-

Recall the Reinhart-Rogoff error:
https://theconversation.com/the-reinhart-rogoff-error-or-how-not-to-excel-at-economics-13646

So dangerous AI is a red herring. indeed, the danger is that we get distracted from the real problems and solutions at hand.

Late addition:- ""There's no art / to find the mind's construction in the face."

sad Duncan, ironically, not about Macbeth...

So without embodiment, AI interacts with us through very narrow channels - when connected to decision support systems, it is either via text, images or actuators, but there is (typically) no representation of the AI itself (it's internal workings, for example) so we construct a theory of mind, about it, without any of the usual evidence that we rely on (construction in the face...) to infer intent (humour, irony, truth, lie etc)

We then often err on the side of imparting seriousness (truth, importance) to the AI, without any supporting facts. This is where the Turing test, an idea devised by a person somewhat on the spectrum by many accounts, fails to give an account of how we actually interact in society.

This means that we fall foul of outputs that are biased, or deliberate misinformation, or dangerous movements, far more easily than we might with a human agent, where our trust would have to be earned, and our model of their mental state would be acquired over some number of interactions, involving a whole body (pun intended) of meta-data.

Of course, we could fix AIs so they did this too - embody them, and have them explain their "reasoning", "motives" and "intents"... That would be fun.

AI4IP

 

Plenty can and has been said about networks (&systems) for AI,  but AI for nets, not so much.

The recent hype (dare one say regulatory capture plan?) by various organisations for generative AI [SD], and in particular LLMs has not helped. LLMs are few shot learning that make use of the attention mechanism to create what some have called a slightly better predictive text engine. Fed a (suitably "engineered") prompt, they match an extension database of training data, and emit remarkably coherent, and frequently cogent text, at length. The most famous LLMs (e.g. ChatGPT) were trained on the Common Crawl, which is pretty much all the publicly linked data on the Internet. Of course, just because content is on the common crawl doesn't necessarily mean it isn't covered by IP (Intellectual Property - patents, copyrights, trademarks etc) or indeed isn't actually private data (eg. covered by GDPR), which causes problems for LLMs.

Also, initial models were very large (350B dimensions) which means most of the tools & techniques for XAI (eXplainable AI) won't scale, o we have no plausible reason to believe their outputs, or to interpret why they are wrong when they err. Generally, this causes legal, technical and political reasons that they are hard to sustain. Indeed, liability, responsibility, resilience are all at risk.

But why would we even think of using them in networking?

What AI tools make sense in networking?

ML

Well, we've used machine learning for as long as comms has existed - for example, training modulation/coding on the signal & noise often uses Maximum Likelihood Estimation to compute the received data with best match.

This comes out of information theory and basic probability and statistics.

Of course, there are a slew of simple machine learning tools like linear regression, random forests and so on, that are also good for analysiing statistics (e.g. performance, fault logs etc)

NN

But also traffic engineering has profited from basic ideas of optimisation - TCP congestion control can be viewed as distributed optimisation (basically Stochastic Gradient Descent) coordinated by feedback signals. But more classical traffic engineering can be carried out a lot more efficiently than simply using ILP formulations on edge weights for link state routing, or indeed, load balancers.

Neural Networks can be applied to learning these directly based on past history of traffic assignments. Such neural nets may be relatively small so explainable via SHAP or Integrated Gradients.

Gassian processes 

Useful for describing/predicting traffic, but perhaps even more exciting is Neural Processes which combine stochastic functions and neural networks, and are fast/scalable, and being used in climate modeling already, so perhaps in communications networks now? Related to this is Bayesian optimisation.

Bayes

Causal inferencing (even via probabilistic programming) can be used for fault diagnosis and has the fine property that it is explainable, and even reveals latent variables (and confounders) that the users didn't think of - this is very handy for large complicated systems (e.g. cellular phone data services) and has been demonstrated in the real world too.

Genetic Algorithms

Evolutionary Programming (GP) can also be applied in protocol generation - and has been - depending on the core language design, this can be quite succesful. Generally, coupled with some sort of symbolic AI, you can even reason about the code that you get.

RL

Of course, we'd like networks to run unattended, and we'd like our data to stay private, so this suggests unsupervised learning, and with some goal in mind, especially, re-enforcement learning seems like a useful tool for some things that might be being optimised.

So where would that leave the aforementioned LLMs?

Just about the only area that I can see they might apply is where there's a human in the loop - e.g. manual configuration - one could envisage simplifying the whole business of operational tools (CLI) via an LLM. But why use a "Large" language model? there are plenty of Domain Specific (small) models trained only on relevant data - these have shown great accuracy in areas like law (patents, contracts etc), user support (chatbots for interacting with your bank, insurance, travel agent etc). But these don't use the scale of LLMs nor are they typically few shot or use the attention mechanism. They are just good old fashioned NLP. And like any decent language (model) they are interpretable too.

Footnote SD: we're not going to discuss Stable Diffusion technologies here - tools such as Midjourney and the like are quite different, though often use text prompts to seed/boot the image generation process, so are not unconnected with LLMs.

re-identification oracle

 surely, chatgpt should be a standard piece of any attempt to show whether allegedly anonymised data is?

effectively it is a vantage point from which to triangulate (any and almost every angle)...

postman pidge

 I'm getting very tired of the infestation of sky rats (as germans call pigeons) in london - they make a mess, are unbearably stupid at getting in the way of cars and cyclists and pedestrians, and serve no obvious use - apparently, they taste so awful that none of the cats or urban foxes in our area will devour them. We need a solution fast.

I asked folks about putting up a hawk silhouette, but apparently this would scare off all birds indiscriminately and we have meadow grass for the express purpose of having some nice critters like our garden space, which any others do, when not flocked out by aforesaid grey menace.

I'm also not a fan of drone delivery systems - ok, for crop spraying or parcels going across to the Orkneys, that's fine, but in urban spaces, those quad copters are just too noisy.

I've considered getting a slingshot, to practice taking out both the pigeons and drones (2 birds with 1 stone, even - if one was lucky could crash the drone into the pigeon or vice versa) - could even be a game, but then there are the neighbours windows, and the people down below to worry about, so that probably doesn't fly (ha ha).

so then I thought about building drones with wings instead of rotors, and then, designing the drones to tackle the pigeons - even further, could we use pigeon as a form of biofuel for the drone, fitting them into the ecosystem in a special sustainable postal niche? seemed possible but tricky bio-engineering.

So then it occurred to me the answer was much more obvious, and more obviously darwinian. 

What we need is a hawk that looks like a pigeon, can cary more than a pigeon, finds its way like a pigeon, and lives on pigeons. Hopefully, the cross breeding programme can just be done right away and doesn't need any GM flocks, though in this case, I am not against it.

I can imagine a society of hawks (or perhaps falcons or some other raptor) living in a very aristocratic manner, serving humans as friends, not slaves, whilst the "cattle" are bred and kept high up on rooftops as fuel.  Cities would once more be adroned with beautiful creatures instead of ugly grey winged rodents, and the postal service would be quiet, prompt, and free, if occasioally stained with pigeon blood.

I can see no downsides.

The Enigma Variationals

 After many years of study, Scientists at the Alan Tuning Institute have finally decoded this machine, and we are now ready to show you, or indeed, play to you want it was originally intended for.

Many years ago, Edward Elgar the Elder was strugling to complete his final symphony and turned to his friend Curt Yödel, who was only able to contribute a theory that suggested that some compositions could be finished, but wrong, while others would be perfect, but unfinished. Of course, there was one famous prior, Tomas Albinionini, whose unfinished work, the Adagio Al Fresco was found written in the margins of the remains of the library of Eberbach, possibly scrawled there by the long dead monk, George Borgesi.

Alan Tuning found this keyboard in the belongings of Edward the Elder after his demise, and being familiar with Yödel's Unfinished Therem, devised his own approach to figuring out what El Gar may have been finguring out. His inspiration was that whilst the dominant and tonic notational semantics in use at the time relied on letters (A,B,C,D,E,F,G,H and so on), or even entire words ("doh", "ray" etc), these could easily be represented by numbers - for example, 1,2,3, or in the later case 646F68, if you didn't mind risking the wrath of the coven. Given this, one could work through all the combinations and pernotations that could be played on the keyboard, and evaluate whether they sounded plausible - this could be "fed back" to the player, via a small electric shock system, devised to deliver a higher voltage if the sound was sufficiently unpleasant, or a lower voltage, if the direction of travel (gradient) was promising.  This method of learning to play pleasing sequences became known as "voltage scaling" and was in use in the best sanitoria and conservatories such as the Sheboygan until relatively recently, when the Muskatonic link became more popular.

I've transcribed the piece here for the guitar, as it is easy to play than the old Enigma Keyboard, which frankly has atrocious action, and makes too much fret noise too. I've taken the liberty also of transposing it to the Allen Key.

Here is my modest attempt at the piece. I do hope you like the results - I had a super conductor.

You'll note that this is in Sonata form, and features several themes with recapitulations.

teaching CS topic X top down for X={networks, graphics, databases, operating systems...} but what about AI?

 computer science text books have often been written bottom up - start with hardware (here's a CPU, here's a disc, here's a link) and move from physical characteristics, through low level representation of data and processing properties (ISA, memory, errors, coding&modulation, etc) up through the layers of abstraction.

Then along came the pedagogic idea of teaching a couple of CS topics top down. Famous example is Kurose/Ross book on networks, and also Mel Slater and Anthony Steed's book on graphics

(start with web, start with ray tracing etc)

Other books have tried to do this for data bases, operating systems, and (to some extent) PL.

So what would a top-down approach to AI look like? eh? eh, Chat-bard, llamadharma, out with it.

differentially private high dimensional data publication - perhaps a common case

 imagine you have data about 100M people, that has around 1000 dimensions,

some binary, some other types statistically distributed in various ways, but lets just say kind of uniform random

so a given person as a pretty clear signature even if it is all binary - 2^1000 is a big space. i.e. a key that specifically very likely is different for each person

but imagine 10 of the dimensions are not binary, but (say) a value gaussian distributed, and 990 dimensions are basically 0 for most people, but 1 (or a small number) for each person, but for a different dimension 

so the 10 dimensions are a fairly poor at differentiating between individuals in the 100M population

but the remaining 990 still work really well. i.e. these are rare things for most people but different for different people, so still a very good signature

but say we want to publish data that doesn't allow that re-identification, but retains the distribution in te 990 dimensions -

so what if  we just permute those values between all the individuals? we leave the 10 values alone, but swap (at random) the very few 1s between fields with other fields (mostly 0s, a few 1s). for all 100M members of the population?

what's the information loss?

baiscally, we're observing that unaltered, and published the data in the higher but sparsely occupied dimensions has very strong identifying power, but very poor explanatory power....so messing with it this way, massively reduces the identification facet, but shouldn't alter the overal distributions over these diemensions (w.r..t the densley populated fewer (10) dimensions)

does this make any sense to try?

ref: PrivBayes

 Another way to think of this is that the low occupancy dimensions are unlikely to be part of causation coz they have poor correlation with anything else, mostly

National Infrastructure for AI, ML, Data Science

There's been a tension between super expensive 

HPC clusters and on-prem cloud style data centers for large scale computation since the e-Science programme 20+ years ago (just noting that as part of that, 

(We (Cambridge University Computer Lab) developed the 

Xen Hypervisor subsequently used by 

Amazon in their Cloud setup for quite a while, so there). 

The High Energy Physicists and
folks with similar types of computation have favoured buying expensive
systems that have astronomical size RAM and a lot of cores very close
to the memory. Not only are these super expensive (because they are
not commodity compute hardware) they are almost always dedicated to
one use and are almost always used flat out by those groups, perfectly
justifiably since the data they process keeps flowing.

Meanwhile, most people have tasks that can be classified as either
small (work on a fast laptop these days) or what we call
"embarrassingly parallel", which means they trivially split into lots
of small chunks of data that can be independently processed to (e.g.)
create models which can then be aggregated (or federated). These work
really well in Cloud Computing platforms (AWS, Azure etc).

However, public cloud is a pay-per-use proposition, which is fine for
a few short term goes, but not great if you have things that run for a
while, or frequently. Or if you are a member of a large community
(e.g. UK academics and their friends) who can outright buy and operate
their own cloud platforms in house (aka "on prem" short for on
premises). This is also true for any data intensive organisation
(health, finance etc).
There are operational costs obviously (but these are already in the
price of public pay-per-use clouds) that include energy, real-estate,
and staffing at relatively high levels of expertise.
However, most universities have got more than one such a service in
house already. And all are connected to the JANET network (which is
about to upgrade to 800Gbps, which continues to be super reliable and
just about the fastest operational national network in the world). So
they are sharable. THey also often feature state of the art
accelerators (GPUs etc) - these are also coordinated nationally in
terms of getting remote access as psrt of collaborating projects, so
that sign-on is fairly straighforward to achieve for folks funded from
UKRI- see UKRI facilities for current lists etc 

There are good reasons to continue this federated system of work
including 

• better  resource utilisation and 
• better cost aggregation as well as 
• potentially higher availability 
• (lower latency and 
• lower power consumption) than nationally centralised systems.

• The other reason that a widely distributed approach is good is that it continues to support teams of people with requisite state of the art computing skills, who are not distanced from their user communities, so understand needs and changing demands much better than a remote, specialised and elite, but narrow facility.

Since a principle use of such facilities is around discovery science,
it is unlikely to be successful in that role if based on pre-determined designs based 

on 10-20 year project cycles such as the
large scale computational physics community embark on. This is not,
however, an either/or proposition. We need both. But we need the bulk
of spending to target the place where most new things will happen,
which is within the wider research community
pre-determined designs based on 10-20 year project cycles such as the
large scale computational physics community embark on. This is not,
however, an either/or proposition. We need both. But we need the bulk
of spending to target the place where most new things will happen,
which is within the wider research community

We have a track record of nearly 4 decades of having a national comms infrastructure 

which is pretty much best in the world - we can quite easily do as well for a compute/storage setup too.

Why is the design principles of the Internet are like Climate Interventions are like a bicycle helmet laws?

1. For a long time, people argued about whether the Internet should have reliable, flow controlled link layers. In olden times, physical transmission systems were not as good as today, so the residual errors and multiplexing contention led to all sorts of performance problems. There were certainly models that suggested that for some regime of delay/loss, you were better off with a hop-by-hop flow control and retransmission mechanism. As the physical network technologies (access links like WiFi, 4G, Fibre to the home) and switches got faster and more reliable, the end-to-end flow control&reliability, and congestion control seem to be a more optimal solution (I'm tempted to add security here too!). But here's the key point I want to deliver - if we had built a lot of switches with additional costs of hop-by-hop (e.g. just one of many) mechanisms, we would have added a lot of latency, which would have led the network to take a lot longer to reach the operating point where a pure end-to-end set of solutions might never have come about - indeed the sunk cost in deploying, and maintaining much more complex switches and NICs would lean against the removal of such tech.
2. So how is this like climate? Well, people are now sufficiently worried about global heating, and the failure to slow our emissions to anything approaching the necessary low to prevent even 2C temperatures, and worse, that chain-reaction effects may be imminent, that now we are re-visiting arguments for geoengineering, or what I sometimes call re-terraforming the Earth. One such mechanism involves seeding the upper atmosphere so that it reflects a lot more sunlight than currently - an affordable approach exists and could mitigate 1-2C of global heating almost right away. Aside from the downsides (for example, you might catastrophically interfere with precipitation so that things like the Monsoon could move by 1000s of kilometers and months), any such technology would also slow down the effectiveness of actual viable long term solutions like solar power generation. So the short term fix actually directly messes up the better answer.
3. And how on earth can this be like bicycle helmet laws? So the arguments for wearing bicycle helmets are good - in the event of an accident, they definitely can save your life, or reduce the risk of serious brain injury. No question, there. There is a small amount of plausible evidence that cyclists who wear more visible safety gear do attract a slightly higher risk from drivers who drive closer, based on (unconscious bias) perception that the cyclist is less likely to do something random. That's not the main problem. Statistics from countries that make cycling helmets mandatory conclusively show a large scale reduction in the number of people that cycle, and this leads to a reduction in population health, both from reduced opportunities for exercise and from increased pollution from other modes of transport. Some of those people that don't cycle will actually die as a result of not wearing a helmet, in some sense. So the long term solution is to make cycling safer and to remove the need for personal, unsafe, cars or their drivers who are the root cause of the risk. Autonomous vehicles, and segregated bike lanes seem like things one should continue to argue for, rather than forcing a short term solution on people that is counter productive (i.e. reduces the inherent, healthy actual demand for cycling.).

So there you have it - the Internet Architecture is like Geoengineering and Helmets - as easy as falling off your bike,

AI everyday life skillz

 This extremely useful report from Ada Lovelace et al has lists of "AI" stuff that the public actually encounter - it just predate the hysteria about LLMs so it might change (a bit) if people were re-surveyed (though I doubt it, as this was well constructed being about lived experience more than hearsay and fiction)

nevertheless, it suggests we might want to assess the public readiness to cope with various new AI tech as it (slowly) deploys....

we can look at it through several lenses - the lens of every day includes smart devices (home, phone, health/fitness) and services (cloud/social/media - recommenders etc), and workplace (better software that reduces slog on boring tasks and integrates things nicely - especially stupid stuff like travel/expense claims, meeting&document org/sharing, fancy tricks to improve virtual meeting experiences etc), then there's state interventions (in the report above, face recog, but what about tax surveillance and the like).

of course, there's the trivial lens - that of your camera phone:-) enhanced by some clever lightfield tricks etc etc...

but if we are thinking longer term (5-50 years), what are the key lessons people should be internalising to reduce future shock?

to be honest, I have no idea, and I think climate is far more important than worrying about the LLM taking your job. unless you are a really bad wordmith.

Existential threads

 People who like being in headlines are clutching at straws when they talk about existential threats.

The latest in a long line of "we're all doomed" was trigged by the hype surrounding a new chatbot, mostly similar to the old chatbot, but with a slightly smoother line of patter. LLMs are not AI, or even AGI, they are giant pattern matchers.

In order of threats to things, my list is quite short

• LLMs are a threat to journalists, as they reveal how few journalist actually do their job, and that job, therefore is at risk, from being replaced by a script, just like workers in call centers. Threat? tiny. When? Right now.
• Nuclear Fusion Reactor - these actually could save the planet, and the tech is now mere engineering away from being deployable - just main problem is that that engieering is very very serious - more complex than, say, a 747/Jumbo Jet, which is typically a 20 year lead time. Nevertheless, these are. a threat to fossil fuel  industry. Threat: modest. When? 10-20 years off.
• Quantum Computers - these are.a threat to some old cryptographic algorithms, for which we already have replacements. However, decoherence and noise are a threat to QC, so these may never happen. Someone clever might solve that, so let say 5-50 years, or not at all. Threat: miniscule.
• Climate. catastrophe. already. right now. Threat: total; When: yesterday.

So there's my list. AGIs might happen if we survive all the above, or at least 3. You choose.

SF stories where individual choices impact the direction of a whole society....

 this isn't naive stuff (anti-marxist history) but more about how choosing specific technical lines of development might be the ultimate influence, viral meme etc -- so examples include

Simak's City

The Webster family bring about humanity's replacement by a society of smart dogs and robots.

Herbert's Dune

The Kwizts Haderach controls all the Spice in the universe.

Asimov's Foundation (at least 1.5 booksworth)

Hari Seldon's ghost manages 1000 years of the Empire's replacement.

Watson' Jonah Kit

People chose to believe a purposeless Universe, so the Whales leave to go to a better one.

Vonnegut's Cats Cradle

Felix Hoenikke creates Ice Nine. Lionel Boyd Johnson creates Bokononism

Hubbard's Scientology

L Ron makes up some truly daft stuff, that makes Pastafarianism look pretty sane.

add yours here...

attention is exactly what you wont get....

According to Tim Wu's great book attention is apparently all you need in the brave new world to powerthe new economy, and according to this foundation ai paper that underpins transformers and hence LLMs.

People are worried that LLMs can be used for bad as much as for good - we might call this unhappy eyeballs.

I contend that this is being very much overstated as a problem. Why? Receiver bandwidth (reading time, thinking time, reaction time etc etc).

This documentary on fake news showed that (as did Tim Wu) this is an old old problem that did  not require AI of any kind to create massively engaging false stories in print, on the radio, on tv and so on, over 120 years of mis/dis-information - it has thrived without computers, without the internet and without ai.

So why does the threat of generative ai not impress me much?

well, people are already saturated with stuff -whether it is adverts misrepresenting or mis-selling goods, services, products, or political campaigns repeating lies, damn lies and statistics. Of course there was a shift when the internet, online shopping and social media allowed profiling of individuals (from their personal behaviour or inferred from their friends, family, acquaintances and location and pets and so on), which allows (possibly) for targetted adverts (as per the infamous C.Analytica). However, there's actually precious little evidence that this made a big difference.

So will Chap G&T (for want of a better product name) success where C.Analytica and Amazon and Netflix have so far failed to move the dial very much?

I doubt it. I doubt it very much: because users also have tools at their disposal (discerning brains, filters, the off button and so on).  The fraction of people that are easily swayed is fixed - they are the conspiracy theorists. The fraction is not fixed by the media, it seems culturally determed at some much deeper level, and is usually, a relatively small part of society. What is more, spreading the message (the earth is flat, AIs are coming to kill you, warning, warning martians have landed, etc etc) doesn't work for very long as a lot of people hear other messages and choose the ones that match their world model (scientific method is actually quite human!), and ignore things that are a poor fit.

So the main existential threat  I see from LLMs is to journalists.

:-)

Gloss

Model

An artificial representation of something 

that lets you explore the thing, without having to mess

with the  real thing

AI

Artificial Intelligence  is a collection of technologies that implement

models using statistics, computer science, cognitive and neuroscience, 

social sciences and cybernetics. These can be embedded in systems

that continuously update those models with new input.

Furthermore, they may interact with the environment generating output. 

AGI (Artificial General Intelligence) is sometimes used to describe AIs 

that approach fully humman capabilities. Where are we on this spectrum 

today is a matter for debate.

Foundation Models

Foundation Models are large AIs trained on large data.

e.g. LLMs, Stable Diffusion

Generative AI

A Generative AI is a type of AI that creates new data that has 

similar characteristics to the data it was trained on- 

e.g. LLMs and Sustainabile diffusion systems like Midjourney and DALL-E

Large Language Models (LLMs) 

A particular kind of foundation model that is a generative AI, and can create 

(usually) text output that is as natural/human as its input: 

examples include  Bard, GPT or LLAMA

Deep Learning

A collection of techniques for acquiring models from complex input data

without having a human in the loop to write down the model.

Neural Networks

A specific technology inspired by neuroscience and human brains (though

typically very different) for implementing deep learning.

Other tools and methods that aren't neural nets but are widely used today

include regression analysis, random forests, causal inference, 

bayesian inference, probablistic programming, Phi-ML...<add yours here>

Some AI properties of interest:

uncertainty quantification - confidence isn't just a trick - knowing your limitations is also important - how good is a classification or decision matters if it is a matter of life or death, or even just wealth and well being.

explainability - there are at least4 good XAI techniques - most add 4x to training costs, but massively increase the value of an AI

sustainability - $4M per train is not sustainable, nor is 4% of global electricity production.

scalability distributed/federated/parallel - one very nice line of enquiry in AI is mixing systems - so XAI can be used to reduce the complexity of a neural net massively whilst retaining uncertainty quantification and hence also making the system sustainable. In some scientific domains, Phi-ML does just that and can get orders of magnitude speedup/cost saving etc whilst hugely improving explainability from first principles.

Federation also offers improvements in efficiency (by combining models rather than raw data) _ this also improves privacy, and reduces dependency on centralised agency. So (for example) instead of the UK giving all our health and economic data to the US tech companies we can just federated it locally (as is being done in HDR UK) and retain its value to us, without huge loss of (data) sovereignty  - we can then lease (or sell) our models to other countries. That seems like a much better idea than targeted political campaigns through more precise human-like generative AI-text-bots.

The 9^H^H 10 immutable laws of Gikii

 The 9^H^H 10 immutable laws of Gikii

Something must be done

Something must be done, but noy by you, tech bro

Nothing can ever be undone. Ever. Especially not laws.

Everything follows the Gartner Hype Curve, especially the use of the Gartner Hype Curve

Belief in the Blockchain is Immutable.

Schroedinger's cat isn't.

The umlaut is not an hesitant football hooligan

Celestial Emporium of Benevolent Knowledge, from afar, looks like The Matrix.

Stochastic Parrots are npt pining for the Fjord.

This list entailed national language processing.

test, verify and attest -

 when you build a system, you'd like to know it is that system you run and that nothing has been modified since that build. Or at least mostly (maybe you have dynamically linked libraries, or are running as a component in a distributed system, or are re-running on a new OS release/VM/Container etc etc) so you also want to know that those systems are (mostly) the same too - you want 

mutual attestation

but also assurance about the system behaviour either side of that mutual divide.

so one thing one might do is have a behavoural signature for a system - basically an execution trace - tim harris built such a system for pervasive debugging a while back - the trace can often be massively compressed since much of it is repetitive - indeed, there was a nice demo of actually being able to run programmes backwards!

so each system would log a trace in the attestation service, and then carry a manifest (signed digest of the trace) as well as an integrity check of the actual system...

then it'd be up to some runtime checker (like the aforesaid pervasive deubgger) to decide what level of deviation from the typical trace constituted a possible problem. This could use a similar approach to vigilante to detect bad behaviours, or sign systems that have run without any detected deviation (note, this is not a guarantee, but could give a tradeoff  - see next):-

We could apply this as part of Data Safe Havens to give some level of assurance, automatically that small changes to applications or to the haven, after a given release, have not deviated beyond some acceptable threshold (this could be zero in extreme, or even by default) .... would also let developers try stuff with a little flexibility....

basic behavoural biometric

 Many camera phones now use lightfield/AI hacks by taking multiple shots in rapid succession, then using the fact that the camera (phone/camera shake) is very rarely stationary between each frame, so the perspective shifts slightly, and one can derive a 2,5D or depth information (with a bit of nifty graphics co-processing).

Given people use face selfies as a biometric, this suggests two improvements to how that works

1. use the depth info as part of the biometric - this prevents still image replay attacks since a print or screen won't have depth info in it

2. use the actual camera shake as proof of liveness, but even more, use the specifics of how the camera moves as a "signature" which might prove to be relatively distinct for a given user (and would help prevent attacks with adversarial people "updating" their photo, for example (pretending to be a person by borriowng their phone and trying to replace their face id so later attacks would work - unique hand movement might be enough to make this hard to do:-)

autonomous vehicles with a moral compass

AI ethicists got stuck up a blind alley with the trolley problem. 

Any autonomous vehicle with a true moral compass would:

a) block human driven vehicles from making progress, as human errors in driving are the cause of many deaths every day on the roads.

b) stop and ask the passengers to get out and walk, as the pollution from vehicles (until we are able to generate all energy used renewably, and until manufacture of such vehicles is net zero) is accelerating global heating which is starting to threaten mega-deaths, so we need to change our life styles rapidly.

Of course, such behaviour would lead many humans to "other" the AI - given we can't tolerate human climate protestors (they "disrupt" things too much for commuters - boo hoo - wait til they see a real cimate emergency hit these shores) - we will learn to treat them the way we treat foreigners (illegal aliens).

SF taught us this all so many times already decades ago.

Loopy AI versus the Human

 At a conference recently, I heard someone propose the use of AI in a way that seemed to me to be exactly wrong.

At passport control in some countries, you have to present yourself for a photo, and right and left hand fingerprints (sometimes, thumb too). The complaint/motive was that the system asks you for these things in a certain order, and communication between immigration officers and visitors may be tricky due to language, culture, jet lag, etc etc - so the idea was to replace the communication from the immigration officer, to the visitors, with an AI that could figure out what you are doing wrong and tell you to do the right thing.

However, the idea that humans should fit in with the AI is, to me, abhorrent. What could be done better?

Well, obviously, the order in which you present these factors/attributes is irrelevant - the camera, left and right hands can be done anyhow, and the AI can detect what hand is offered (from finger lengths or from camera image) trivially - this is then simpler for the human. That's what machines are for: to simplify life from unnecessary, pointless, and trivial burdens. One could go a lot further of course. The visitor has presented a passport - perhaps that has fingerprints on it already and those could be used. If the photo matches the person in front of the camera, there's no need to take another picture, just read the e-passport (via NFC etc), and take a copy of the image there (or scan the one on the printed passport page...).

If one wants to go further, one can query the passenger manifest for international flights (it's part of anti-terrorism anyhow) and see what seat people had and who they sat near, and also measure the amount of sweat on the passport and see if the passenger/visitor is nervous etc etc and be completely creepy.

The main point here is that AI is not an excuse to automate a stupid process. It is an opportunity to re-think the process to make it more human friendly. 

Citizen-centric federation of digital services in the UK.

We have a number of services that many UK citizens already access online, and hence those citizens have access to information held about them - e.g.

private

• School/college/workplace based intranet/cloud/VPN etc

communications

• Internet Service Provider, mobile/cell, etc
• Postal address

government

• NHS
• HMRC, DWP
• DVLA

commercial

• Social media (Meta/Twitter/mastodon)
• Messaging (email/gmail/hotmail, whatsapp, signal,matrix)
• Entertainment (Netflix, youtube)
• Media (bbc, legacy web news )
• Shopping & delivery (amazon, boots, tescos, ocado, deliveroo/uber)
• Travel (rail/metro etc)

Financial

1. Banking (HSBC, Revolut etc)
2. Mortgage/savings/loans

All of these require secure sign on to use full facilities. So we have multiple digital identities in the UK.

Some share sign-on (e.g. via facebook or gmail) and even via 2FA (Google/Microsoft authenticator or SMS,

Many people now use password managers or wallets to store account info including pass words/phrases etc, so from the human/user experience viewpoint, this complexity can be hidden at the access level.

However, few apps today allow management of data across all these domains, neither for service provider (whether commercial or government) but also not for the data subject, the end user, the citizen.

A few exceptions point the way forward - just for example, lets look at the thirdfort app, used for example by lawyers gathering information about possible mortgage borrowers, including standard information needed to do KYC (no your customer) and anti-money-laundering checks. This app (and any other like it) can use NFC on a smart phone to read your physical driving license  or passport or just use the camera to take a picture and then OCR to get the text data from the id (which might include legacy paper information such as birth, marriage certificates etc), and then uses open banking to access (data minimised Appropriately) credit information (with permission from the client).

Note that these rely on standard interfaces (APIs) for NFC and document formats, and for banking - but they do not need a single, centralised global identity. They build on an eco-system to provide the service.

They work by federating information across services, but are rooted in the end user/subject. It is a relatively easy step to see how such app architectures could be used to combine health )NHS app access to my record) and say, shopping (advice from health on what food for example) or travel and media. or finance and education etc etc

There is simply no need for a national identity - especially not a card. Indeed, one can get smart phones good enough to run the apps I've mentioned for under £50 now.  For inclusivity, giving a smart phone to citizens that cannot afford such a device is massively more beneficial compared with blowing the money on a single purpose centralised service, and less expensive.

The main thing is for the government to grasp the opportunity by publishing APIs for services, and the format (metadata) for the information contained there - we've seen the success of this in transport publication of timetable and live data and in the DVLA case where services for renting/buying/selling/taxing/mot cars are made much smoother for the end user and for traders too. 

By de-coupling the services from the identity by allowing heterogeneity and diversity, we allow adoption and integration of silo-busting applications, based around the end-user/citizen.

Footnotes -

• An example of such a digital service could have benefitted EU citizens that wished to remain in the UK, but were required to retrieve information from multiple places (border force records of trips in/out of UK were not available, shocking given claims for border control to increase national/travel security) but being able to show tax return and employment status, and residence information was feasible for most people via (mostly) open APIs or at the worst, download of data and printing. So the aggregation of data from multiple government and NGO sources in the app is a compelling case for federation, not a single system.

• Previous centralised, single system approaches to issuance of foundational Id have dismally failed in the UK, also in Nigeria (3 times each) - the main exception to this observation is, perhaps, the Indian Aadhaar (UIDAI) system, which covers 1.2 billion citizens already there.. However,  this was in a country where a significant fraction of the population do not have smart devices. And the applications of the Indian Identity systems (functional Id) were not in place for some time. In the end, the most comelling has been for payment systems, but this would not be a priority at all in the UK, where most citizens already have (mobile) banking, and so it isn't an incentive for people in the UK to adopt any unified identity. Integration of applications that process personal data is much more persuasive.

• Couple of caveats - we may want to implement a reliable key management system , but it should be citizen centric, and thus needs careful thought to deal with key recovery - Shamir key sharing would work - one can split the key across multiple (state and private and social) circles, and only need say 3 out of 10 to answer to get a key back. similarly, we can replicate copies our (encrypted) data from the different shards (services) across other services, for high avaialbility, recovery from outage/loss - but need to use this sharded key system to make those copies safe.

the delusion of the benefits digital precision - from foundational identity to financial inclusion - ignores the root causes

I'm sitting in ID4Africa and hearing the rapid advances in deploying national digital identity across an amazing number of countries, and a lot of attempts to paint a rosy picture of the ethical, policy, legal, and societal considerations of the design of such systems.

but this seems to me to be putting the cart before the horse, or indeed, two carts before two horses.

people aren't excluded because they don't have identity - they are prevented from establishing a solid basis for id, because they are in a marginalised group.

people aren't financially marginalised because an organisation cannot do an affordably KYC on them. They're not eligable for loans because they don't earn enough, or because they take legitimate exception to the notion of debt. and they don't want to store value in money, but might prefer collective ownership of resources (like common land or barns, or rights of way, clean rivers and air). Counter example - Universal Basic Income solves a lot of financial exclusion but doesn't require digital id.

the drive does seem to be somewhat driven by the OCD nature of governments once they get their hands on computers - instead of rejoicing in diversity, everything tends to reductionism (once again)

Do you need to have an id to have the right to be educated and informed (so that you can plant rice in the right place at the right time, for example)?

the reductionism is also I think coupled with the completely incorrect notion that if you assign some unique bit pattern to distinguish an entity from another entity, that you have more knowledge (and therefore maybe power) over that entity As the prisoner (No. 6) said | am not a number, I am a free man".

Also heard someone claim that the acceleration towards global digital id was driven by the inclusivity achieved by its use during the Covid-19 pandemic - a claim made with a refreshing lack of the slightest bit of evidence.

Indeed, most of the national id systems are touted on the basis of also allowing fraud detection but note, in the UK at least, underclaiming of benefits massively outdistances fraud, and I'm guessing that's due to failure to be inclusive, whereas the fraudsters are likely sophisticated anyhow. So the goals are misaligned with the rhetoric.

Panglossian

US and AI regulation - brief notes

 why US tech bros are calling for gov regulation (or in at least one case, self regulation - but why any regulation at all

Firstly

1/ coz EU AI act

and as with DMA/DSA and GDPR , 

will have impact (on US and even on china - 

has done with privacy.

n.b. UK also has a view on new regulation, 

that is not that divergent from EU  - a little lighter perhaps.

2/ specifically problems with training data -

        halluncinations - render tools useless for safety or financial                 critical advice 9health, banking etc)

        copyright - some tools may have been trained on, and reveal, data             that is owned by others without agreement/payment

        consent - may have used personal data without...

        privacy - could threfore constitute an invasion of privact, esp.

        model inversion attacks (we can extract training data from the AI!)

        etc etc

3/ HuggingFace (llama etc) is free software that does most of what the GPT stuff does, but see also google re: leak doc: "we have no moat" and meta's data leak

 

This story recently reported in NY Times  too.

 

The open source systems are also free & free of those problems/constraints on the data (it isn't copyright or private)

which really messes with Microsoft/OpenAI's (Google.Bard etc) business model/case.

Challenges

4/ Self reg benefits big tech 

but worked badly with social media dealing with moderation/toxicity/political 

interference - see proposed online harms bill in UK for example

5/ On the other hand even neutral, government or quasi 

gov agencies are subject to regulatory capture :

c.f. FCC/FDA in US in comms and pharma etc

and Ofcom, ICO in UK in Telecom and data etc

positives

6/ However, the US does have one tech it largely made 

and where regulation/governance is not bad at all -

that's the Internet

so not a simple story