Today is the 87th anniversary of the Battle of Cable Street. #NoPasaran today, and always, but it is worth remembering why #AntiFascist commemorate that date - because it represents a diverse alliance of different groups who just wanted to live & peace and coexist against an organized, well-funded and (tacitly or overtly) state-supported Fascist gang.
And we won, and the Fascists lost. They can always be beaten.
@94caf711 And the "migration crisis". 85+% of asylum claims succeed, but there's money to be made with a convoluted overloaded bureaucracy and a need to house the poor folks stuck in limbo.
Public money handed over to landlords, hoteliers and consulting firms, and a ready-made scapegoat to distract attention from utter ministerial incompetence. An absolute twofer! What's not to like?
@f3ac7035 The Tantrum Corps that basically ragequit when it was suggested that they may actually have to face consequences if they murder people rather than it just being their Call of Duty playground?
@bf6ed480 The tech really has advanced in recent years. After years of a slightly clunky box, moving out here to BCN really helped (they even have a 24hr free service line!) and the last replacement was better again.
Occasional reminder that, along with every rancid transphobe they can possibly find, "the world's leading progressive voice" employs a Spanish correspondent who's basically a Francoist and has repeatedly advocated for jailing people for non-violent political activity.
What does this have to do with LLMs?
The fundamental question is: "What is your training set; how do you verify it; where do you expect it to be predictive?" 3/
LLMs fudge all of this, and so this becomes a "grey goo" problem.
LLMs are poisoning their own input set, and this is IMO very likely to produce a spectral line.
I think that "Pre-LLM" and "Post-LLM" will be very easy to detect.
And for anything that there are actual $$$ attached to: Never connect it to a public model.
/end
Oppenheimer's work ushered in a world which is *visible* in a scientific sense. We can see the pre-nuclear world, and the post in the record.
The Manhattan Project literally changed the world.
While its children may be stranger, and more goblinoid than the grimdark that the Cold War suggested, we still live in the shadow of Hirosima & (to a much less extent) Trinity. 2/
What does this have to do with LLMs?
The fundamental question is: "What is your training set; how do you verify it; where do you expect it to be predictive?" 3/
Someone said that: "LLMs cannot write Barbie, they cannot write Oppenheimer". This is, of course, true.
The choice of those two specific movies places them within the blast radius of LLMs, which came to prominence at roughly the same time. And that's cute -- but wait, there's more. 1/
Oppenheimer's work ushered in a world which is *visible* in a scientific sense. We can see the pre-nuclear world, and the post in the record.
The Manhattan Project literally changed the world.
While its children may be stranger, and more goblinoid than the grimdark that the Cold War suggested, we still live in the shadow of Hirosima & (to a much less extent) Trinity. 2/
@210de87f An interesting experiment I've tried: Want to find out what terms are empty, overused makeweight in your field? Start typing a general statement & watch an LLM try to autocomplete it by adding them.
@4ebb1885 This population is likely to view the convincing, authoritative view as something it is not (one reason why I think that it was massively irresponsible to package LLMs with this style of interface, and especially with a style that contains faux-selfhood). FWIW, we've known about this problem since Eliza, and yet nothing was done.
Second, and related, is the myth of algorithmic infallability, e.g.: https://emptycity.substack.com/p/computer-says-guilty-an-introduction 3/
@4ebb1885 I'm not sure where we go from here - but in general, I feel like we should be defaulting towards to caution in a way that we very much are not right now.
One technical point that I do want to push back on - I don't see how LLMs evolve from their current state of statistical bins to a "model of the world", which they currently lack. Adding more parameters, and ingesting an increasingly more polluted input set just may not give any better results than today. /4
@4ebb1885 My cocnerns are twofold.
First, which I think you addressed in your piece, is the "grain of salt", the skepticism that is required to take theoutput of LLMs and apply them.
You and I know that what they are producing is "spicy autocomplete" or "eerily accurate madlibs", depending upon your viewpoint. But the general population does not. A population that has 20+% of people that do not understand how ranked-choice voting works. 2/
@4ebb1885 This population is likely to view the convincing, authoritative view as something it is not (one reason why I think that it was massively irresponsible to package LLMs with this style of interface, and especially with a style that contains faux-selfhood). FWIW, we've known about this problem since Eliza, and yet nothing was done.
Second, and related, is the myth of algorithmic infallability, e.g.: https://emptycity.substack.com/p/computer-says-guilty-an-introduction 3/
@4ebb1885 I always enjoy reading what you write, and there are some really interesting parts in the piece - and you also, en passant, pointed out that my carbon maths was wrong for LLMs (they're not as bad as I was claiming).
I'll also take the viewpoint that these things now exist, and so it is on us (as senior technologists) to see how we can use them to help new programmers - and that might, in turn, help democratise programming - which is, effectively, still a priest class today. 1/
@4ebb1885 My cocnerns are twofold.
First, which I think you addressed in your piece, is the "grain of salt", the skepticism that is required to take theoutput of LLMs and apply them.
You and I know that what they are producing is "spicy autocomplete" or "eerily accurate madlibs", depending upon your viewpoint. But the general population does not. A population that has 20+% of people that do not understand how ranked-choice voting works. 2/
@ef3c2bd1@58fbd252 It's literally top hit. And we have dealt with so many people that just want to waste our time & energy.
So, I'm sorry if you felt bad.
My project is live. The beast is slain. I was right.
But nothing makes up for looking out of the corner of your eye and, just for one second, thinking that there's a comment on a thread from an old friend - a mazel tov for the sort of technical problem that they used to relish.
There isn't. There never will be. Their memory is, and ever will be, a blessing. As will their god-tier snark.
But we have to go on, and we will.
@4ebb1885 I always enjoy reading what you write, and there are some really interesting parts in the piece - and you also, en passant, pointed out that my carbon maths was wrong for LLMs (they're not as bad as I was claiming).
I'll also take the viewpoint that these things now exist, and so it is on us (as senior technologists) to see how we can use them to help new programmers - and that might, in turn, help democratise programming - which is, effectively, still a priest class today. 1/
With yesterday's release put to bed, I'm just finishing off prepping a conference talk, and am apparently once again at the "Oh, fuck, I have far too much material" stage. In truth, though, this is always preferable to the "Aaargh, I have nothing worth talking about." stage that inevitably precedes it.
@0b13872d I hadn't heard of it until I went to Uni - one of the terms was referred to as "Michaelmas Term".
Now, goose on St Martin's Day (Nov 11th) is a different kettle of fish...
My #Barcelona apartment has both an enclosed interior terrace with two doors out to it & a large front balcony overlooking the street.
The different bits of Outside don't connect up & I'm pretty sure my little guy thinks we live in Castrovalva - he keeps asking me to open both doors and then running through the terrace and looping round a few times. #CatsOfMastodon (Not remotely sorry for the #DoctorWho deep cut either).
Still experimenting with Copilot for #Java development, and writing. I typed this: "As Java developers, there is an analogue to this " - and here's what its autocomplete suggestion was: "As Java developers, there is an analogue to this in the Java ecosystem: the ability to work at different levels of abstraction within the Java ecosystem."
One of the most striking things about it when using it for writing is the repetitive, circular nature of a lot of what it suggests. 1/
Now, because I know a certain amount about how LLMs work, this isn't actually a surprise to me - and I've known about this effect since reading van der Linden's rather overlooked classic "Expert C Programming" back in '97 or so, but it still seems to surprise a lot of the breathless hype merchants. /2
An adolescent pigeon is back in the small windowbay by the kitchen again today. I have no proof, of course, but I like to think that it's the same one that hatched there a couple of months ago.
@f0fe8ed3 You should also be aware of the research done by New Relic in this area (e.g. https://newrelic.com/resources/report/2022-state-of-java-ecosystem) - it is based on direct reporting on telemetry that their customers send to them and is *not* self -reported. It also has a very large sample size (10s of millions of production JVMs).
Feel free to DM me if you want more details or an introduction to folks there who may be able to provide you with a more up-to-date dataset.
@0b2d82b6 I, for one, cannot believe that a man who would use the details of his sexual relations with women in order to try to embarrass & humiliate a vulnerable 78-year old man on live TV, for the sole purpose of entertaining some of the worst people in the UK; could possibly have done the things that he's credibly accused of.
In short, a useful tool, although not as useful to me as I suspect it will be to others, b/c I don't spend a lot of time doing the sort of coding that it can help with anymore.
I'll glad take the increase in writing quality it's giving me, but anyone who claims that this is going to put programmers, especially senior ones, out of a job simply doesn't know what they're talking about. /2
Oh, yeah, and it straight-up hallucinates - classes and enum values that just don't exist started being suggested almost immediately.
This is less of a problem in a language like #Java where the IDE will immediately flag that, but in more "helpful" / dynamic languages, I can see this being a significant source of bugs, especially in code written by more junior developers.
Results from using #Github #Copilot so far: Coding - modest improvement in the areas where Copilot is useful (boilerplate, low-level implementation, things I would have used SO for) - perhaps 25% decrease in time spent.
Writing (yes, it has started suggesting in my asciidoc files as well): Zero increase in speed, but an increase in quality - as I look at what it's proposing and think: "Good Lord, I can do better than *that*". 1/
In short, a useful tool, although not as useful to me as I suspect it will be to others, b/c I don't spend a lot of time doing the sort of coding that it can help with anymore.
I'll glad take the increase in writing quality it's giving me, but anyone who claims that this is going to put programmers, especially senior ones, out of a job simply doesn't know what they're talking about. /2
The need to service those debts has further knock-on effects, such as reducing food budgets, leading to decreased nutrition, higher levels of stress, etc, etc.
In short - it's nothing close to the whole story, but there are real effects here, with real consequences, that we've known about for a long time.
The "Vimes Boots" formulation of it may be a bit glib, but there are still plenty of people who have never thought about any of this that it could a useful starting point for. /4
@adac0016 Is there a ferry route to Cyprus, or would it have to be a privately chartered boat? I'm pretty sure there is only 1 ferry from Italy to Malta.
@ceafb3f6 Is the answer: "Cocaine, and being surrounded by a pod of people who consistently tell you that your half-baked dorm room ideas are the profound truths of the universe"? (I.e. the same as basically every other billionaire)
Oddbean new post about login | logout
Notes by Ben Evans | export