StuartGray

@StuartGray@mastodonapp.uk

#UK #Software #Developer, #IT #Architect, #Web #Geek, #Python.

Keen interest in #DigitalRights and #Privacy.

Long-term interest in #AI, #MachineIntelligence, #MachineLearning, #ProcGen, #LLMs, #NLP, #DeepLearning, #ComputerVision, #ImageGeneration.

Long time #Gamer, mostly #FPSs, #RPGs, #MMOs. Wannabe #GameDev, tinker with #Unity, combining #ProcGen & #AI in fun and sometimes #NSFW ways.

#SciFi lover, #Trekkie, #StarWars, #Culture, & many more. #Book lover, #Centrist, #ProEU, #Rejoiner.

This profile is from a federated server and may be incomplete. Browse more on the original instance.

glynmoody, to Ukraine
@glynmoody@mastodon.social avatar

Ukrainian special forces ‘in operating against Russian mercenaries’ - https://www.theguardian.com/world/2024/feb/06/ukrainian-special-forces-sudan-russian-mercenaries-wagner doesn't seem like best use of scarce resources...

StuartGray,

@glynmoody I’m torn on this, 50/50, without knowing Ukraines goals & objectives - which they’re unlikely to openly publish.

I’d hope and guess it’s a targeted extension of their guerrilla campaign inside Russia - with a goal to use small targeted attacks against high value targets to destabilise, demoralise, distract, raise the cost of being in Ukraine, and ultimately redeploy scarce Russian units away from Ukraine.

It’s not a bad idea if the resources committed are minimal.

StuartGray,

@glynmoody “Russian mercenaries” are all fully state funded, it’s another branch of the state in all but name.

Russia has multiple Wagner operations in various countries around the world, some of which are easier to access than others.

It forces Russia to re-assess its priorities, potentially re divert its own scarce resources, and ratchets up the already existing tension (remember the cancelled coup?) between the Putin and members of the Wagner group.

0xabad1dea, to random

A few times I have told the anecdote that the singly most baffling thing I ever saw in a code review — not the most insecure, just the most “how could a real programmer have written this? how could this ever make sense?” thing — was simply a C++ variable “number_of_trucks” … declared as float. Unambiguously referring to real physical trucks in a fleet.

Reader, it’s been over ten years and I am blowing the gods damn whistle. I had edited that story to protect the guilty: the variable was named number_of_planes. It was shipped by a company whose name begins with “B” and rhymes with “GOING out of business.”

StuartGray,

@0xabad1dea my best code review WTF story, is the time I reviewed the code from an large multi-national outsourcing partners new on-site developer.

I forget the precise details of the change they were tasked with, except it was mostly adding a new feature with minor changes to integrate.

They wrote a new class with 20+ functions all named function1, function2, etc...

And zero comments describing their purpose.

They didn’t last long.

kellogh, to random
@kellogh@hachyderm.io avatar

when i’m looking at a post, is the like count only the number of people from my server that have liked it?

StuartGray,

@kellogh No, or at least, it shouldn't be.

Had to double check, but a recent post of mine had 2 "favorited" notifications, both not on my server, and on the web UI they both show up when you click the tiny star icon next to the count at the bottom of the post.

Post visibility is somewhat complex, but anyone who can see your post should be able to favorite it.

However, there is also a delay factor for posts & likes to replicate between servers & become visible.

kellogh, to ai
@kellogh@hachyderm.io avatar

with coding like github copilot, my workflow is

  1. generate
  2. read
  3. edit
  4. test

i get a bit confused when people say things like, “AI writes insecure code”, as if they’re not doing steps 2 & 3.

a lot of people say they don’t get much benefit from copilot. i imagine a lot of devs are in this category, if their work is mainly maintenance, then there’s not much of step 1 that’s needed

StuartGray,

@kellogh I have a theory that this is closely related to peoples default starting position on the Plan-Do-Check-Act cycle.

In short, some people naturally prefer to start tasks by planning first, where others prefer to start by doing e.g. read the ikea instructions first vs. just start assembling

My gut feel is Planner types prefer to start with text on a page & then edit it, where Doers prefer to jump in and write from scratch - so step 1 is uncomfortable for them.

https://www.lean.org/lexicon-terms/pdca/

revk, to random
@revk@toot.me.uk avatar

Is there really any good reason why the UK legislation government web site (https://www.legislation.gov.uk) asks if I want to consent to cookies?

Why would the web site need cookies at all?

It provides a useful service to allow access to UK legislation.

Does it really need to track who I am? What laws are most popular with my demographic? Which SIs I like to read regularly?

What laws it can try and sell me?

FFS this is mad!

StuartGray,

@revk Not sure about this specific instance, but this was something that Cummings pushed for while he was working for No10.

He openly wanted and pushed for Google Analytics on all UK gov web sites.

I assume as part of his misguided "Government should be run like a business/startup" views.

lauren, (edited ) to random
@lauren@mastodon.laurenweinstein.org avatar

AT&T is sending out letters warning they want to kill virtually all landlines (and perhaps related data circuits where fiber is unavailable) across essentially their entire coverage area throughout California. This would have devastating effects. Related CPUC meetings will be taking place through March.

Landlines provide crucial services for individuals, businesses, and other organizations in a wide variety of situations -- not just emergencies when cellular and Internet service tends to rapidly fail, but also for vast numbers of people in areas with poor, unreliable, or in many cases (even in large sections of major cities!) NO cell service, NO fiber, etc.

Landlines often provide the only available communication in a wide variety of security and safety situations, from elevators to interior spaces of all sorts where cell service simply doesn't work.

Many disabled and other persons have crucial equipment that depends on landlines. Often they are not tech-savvy and do not have friends or relatives to help them through forced technology changes.

AT&T has been shirking its public safety responsibilities for years, while still leveraging their effective monopoly on services in so many areas.

Their new effort must be stopped. I'll have much more to say about this as the situation progresses.

StuartGray,

@lauren you might want to double check the technical details of their proposal.

There is a similar effort underway in the UK, where landlines as we currently know it are being phased out in the next 2-3 years & most people are blissfully unaware.

However, it doesn't mean ripping out lines or turning them off afaik. It's a combo of moving to VOIP-like provision over exiting lines & removing the need for battery-backed regional exchanges.

Not that this is great, but it's important to know.

StuartGray,

@lauren Apologies if you felt I was somehow implying that you don't know what you are talking about - definitely not my intention.

In the UK, BT provides a similar role and will no longer be required to provide land line installation either.

However, when you look at the usage figures for landlines, they have declined dramatically as mobile services have risen.

I don't know what the answer is, and I'm not defending current proposals, but it's obsolesence of land lines in action.

StuartGray, to llm

Thinking some more about using external plugins to aid & guide LLM output over longer generation, and had a crazy idea which might just be workable in Oobabooga;

A plugin that adjusts generation parameters based on a pre-defined list & inline commands.

It would work something like 3rd party tool calling with parameters, BUT on seeing [Tempo:+] in the output, generation is paused, generation params are adjusted, and then resume generation.

Not great, but def. do-able.

StuartGray,

The aim would be for the LLM to generate the commands to adjust model parameters, rather than intervening.

I'm not sure if a prompt alone would be sufficient for an LLM to insert the commands at appropriate points reliably. It might need a finetune, but the prompt route should be easy enough to test.

If nothing else, the commands alone would server as useful hints or reminders in the absence of auto-parameter adjustment.

StuartGray,

@kellogh Thanks for link, I hadn't heard of this one. I'll be sure to check it out. It's not the same, but it sounds very similar to a lot of LangChain features, which has its uses but is largely what I'm trying to avoid.

It's kind of hard to explain the specific issue if you don't use LLMs to write long form fiction.

In essence I want to augment LLM abilities with dedicated tools that are better suited/more efficient/more accurate than pure LLM generated solutions, to improve fiction writing.

StuartGray,

@kellogh A non writing example would calculators.

LLMs are notoriously bad at math, but you can teach them to use a calulator tool as part of their generation, passing params & receiving the answer.

Calculators & function calling are both well defined, well understood, and far more effecient & effective than training an LLM to do math.

Aside from fiction writing, I also have an interest in LLM enhanced games. An example from gaming is mapping - keeping track of a users location, rooms,

StuartGray,

@kellogh objects, available exits & routes, line of sight visibility etc...

You can try, and plenty of people have, to use an LLM to handle this, but like math, they're not very good at it and use a lot of compute power for a relatively simple task we already know how to do well - even simple, regular games do it all the time.

So, far better to make "mapping" a tool that an LLM can query, and leave the LLM to do the bits it does best.

Those are fairly obvious, straight forward examples.

StuartGray,

@kellogh However, in trying to use LLMs to generate long form fiction, I've come to suspect that there is a new class of tools they need - one that doesn't have direct analog like a calculator or gaming.

That's because the guidance & support they provide is only useful if you have the ability to write coherant text automatically, like LLMs do.

e.g. for fiction writing, a "mapping" tool equivalent would be things like high level story structure, plot, key events, tempo, character arcs

StuartGray,

@kellogh I don't doubt it's possible with things like AutoGen or LangChain - but that requires the use of a fairly heavyweight framework, and as I said, I'm specifically trying to avoid using an LLM to check on LLM output.

A single LLM agent ties up a large amount of resources, multi-agent a lot more.

I'm interested in using this in an offline, privacy & censorship free local LLM that is likely to be operated in a GPU-poor or relatively low computer-power environment.

Another comparison

StuartGray,

@kellogh would be to something like Stable Diffusion & it's 3rd party guidance addons for things that augment its image generation with 3d depth perception, human poses, temporal stability.

Using these addons requires only one image generation.

Whereas a multi-agent type approach would be the equivalent of generating multiple images, with all the time and resources required, and then combining them at the end.

Both approaches work, but one is less efficient and longer to process.

StuartGray, to llm

Something I've noticed whilst experimenting with LLMs to write fiction that I've never seen discussed; They lack the ability to pick up on & appropriately use important meta-textual patterns such as tempo & pacing.

I mostly use small, local LLMs, but the same general issue exists with ChatGPT.

It becomes increasingly noticeable the longer the output generated, and it's part of what gives LLM generated text it's "uncanny valley" feel.

1/4

gutenberg_org, to books
@gutenberg_org@mastodon.social avatar

Let´s celebrate, today is Public Doman Day!

Plenty of new titles are available now and volunteers at @DProofreaders will have plenty of work ahead.

The Unpleasantness at the Bellona Club by Dorothy L. Sayers
The Mystery of the Blue Train by Agatha Christie
The Giant Horse of Oz by Ruth Plumly Thompson
Tarzan, Lord of the Jungle by Edgar Rice Burroughs
Money for Nothing by P.G. Wodehouse
The House at Pooh Corner by A.A. Milne
Hunting for Hidden Gold by Franklin W. Dixon


1/

StuartGray,

@gutenberg_org Beware that in the UK only, Peter Pan is the subject of a perpetual limited copyright expiry exception;

https://www.plagiarismtoday.com/2015/10/21/peter-pan-and-the-copyright-that-never-grew-up/

StuartGray, to random

I'm sure and will be along to condemn this hate march on Armistice Day any minute now...

https://news.sky.com/story/fighting-reported-as-people-shouting-england-til-i-die-try-to-reach-cenotaph-13005216

Edent, to SmartHome
@Edent@mastodon.social avatar

Those of you who have a , what are your use-cases for smart plugs?

Most of my electrical things aren't just on/off.

I've got a light, a fan, and an electric blanket which can be controlled from the plug.

Other things (like TV, kettle, etc) need interaction after the power is turned on.

Have you found any good uses for plugs?

(Not interested in rants about security & privacy, thanks.)

StuartGray,

@Edent if you don’t already do this, you could consider using them the same way I use my “dumb” remote power plugs.

I took some time to figure out how to split my electrical devices into always on vs. power saving, and use the dumb remote power switches to turn off all the “power saving” devices at night/while I’m away.

I used a bunch of extension cords to collect like devices under one controllable socket, and made sure that had adequate surge protection.

johncarlosbaez, to random
@johncarlosbaez@mathstodon.xyz avatar

No. Succeeding in love is not easy, and there's no formula for it.

Here's the essence of this bad take:

"Since life is itself simply a game in disguise, having a few mathematical tricks up your sleeve can also give you an edge in the game of life."

Life is not simply a game in disguise. There are no fixed rules, apart from possibly the laws of physics. More importantly, there's no fixed definition of what counts as "winning". In fact the whole concept of "winning" doesn't apply, except in very limited realms.

I'm reminded of an anecdote I heard from the statistician Persi Diaconis. I'll probably get the details wrong, but it goes something like this:

Persi Diaconis was friends with an economist who had just gotten two job offers, one on the east coast of the US and one on the west coast. The economist was having a lot of trouble deciding which offer to take: each had its pros and cons. So Diaconis said "Hey, why not use the mathematics you're always talking about? Compute the expected utility in each case, and pick the offer that maximizes it!"

And the economist said "Come on, Persi! This is SERIOUS!"

.....

The article is here:

https://www.theguardian.com/science/2023/nov/05/how-maths-can-help-you-win-at-everything

and I thank @pigworker for pointing it out.

StuartGray,

@johncarlosbaez @pigworker I’ve saved this to read the link later, but I’m commenting because it always amuses me when people think games have to be won or have a winner - if that was the case, Minecraft wouldn’t exist or be so popular. And yes, it’s also a frequent criticism of Minecraft too.

I assume it’s not in the article, but there’s a concept of finite vs. infinite games. Finite games have winners, in infinite games the goal is to continue playing.

aiefel, to random
@aiefel@mastodon.social avatar

deleted_by_author

    •
    StuartGray,

    @aiefel I'm looking forward to the first owners taking pictures of...

    ...AI generated art 🤔

    jhpot, to random
    @jhpot@mastodon.social avatar

    "Grok" being the name for an AI chatbot is just the latest example of tech bros not understanding science fiction. In "Strangers In a Strange Land", the famous 1950's novel the term comes from, to "grok" something is to understand it on a deep level. AI chatbots, as we all know by now, do not grok—they bullshit.

    I'm really tired of sci-fi terms being used by people who read the books and got nothing from them beyond “lol this word sounds cool”. (See also: “metaverse”.)

    StuartGray,

    @jhpot This is Musk we're talking about. In this specific instance, he understands the meaning of "Grok" perfectly, and is using it in the same way that right-wing politicians regularly use "double speak" to mis-represent an issue.

    He both wants and expects his fanboys to think his LLM is special, different, "the best".

    It's marketing.

    In exactly the same way that Musk was publicly signing an open letter calling for an AI pause - whilst secretly registering his new AI corp.

    davidallengreen, to random
    @davidallengreen@mastodon.green avatar

    NEW SUBSTACK

    "Computer says guilty" - an introduction to the evidential presumption that computers are operating correctly

    The first in a seres of posts on the Post Office Horizon prosecutions scandal

    https://emptycity.substack.com/p/computer-says-guilty-an-introduction

    StuartGray,

    @davidallengreen The date on the act tho. Oof.

    Somebody drafted that whilst the IT industry was spending significant time & resources testing for & fixing IT systems for the Millenium bug/Y2K problem. And the IT press was covering it all.

    The timing suggests either a staggering lack of IT awareness or incompetence (in which case, why were they involved with drafting it?), or it was a deliberate, informed decision.

    mastodonmigration, (edited ) to Astronomy
    @mastodonmigration@mastodon.online avatar

    @AstroMigration Post of the Day

    Speaking of , today's (https://mastodon.social/@sundogplanets/111025157450982774) goes to astronomer Prof. Sam Lawler @sundogplanets:

    "Almost all the satellites you can see with your eyes now are Starlinks. Nicely demonstrated in this awful/beautiful photo: https://apod.nasa.gov/apod/ap220614.html "

    Thank you Prof. Lawler for continuing to remind the world of the catastrophe unfolding in low earth orbit.

    Check out @AstroMigration for more great , and

    StuartGray,

    @mastodonmigration @AstroMigration @sundogplanets Whats worse is that, like Climate Change, "today" is as good as it gets and the forseeable future is increasingly worse.

    Not only are significantly more Starlink sat's due to be launched - but there are at least two competing nation state systems in active development from both Russia & China - so imagine this number of sats multiplied by 3 or 4 within 10 years just for this one purpose.

  • All
  • Subscribed
  • Moderated
  • Favorites
  • •
  • megavids
  • rosin
  • ngwrru68w68
  • tacticalgear
  • DreamBathrooms
  • mdbf
  • magazineikmin
  • thenastyranch
  • Youngstown
  • khanakhh
  • everett
  • slotface
  • Durango
  • kavyap
  • JUstTest
  • osvaldo12
  • normalnudes
  • cubers
  • GTA5RPClips
  • InstantRegret
  • cisconetworking
  • anitta
  • ethstaker
  • Leos
  • modclub
  • tester
  • provamag3
  • lostlight
  • All magazines
    •