Posts by Mike007 640 publicly visible posts • joined 11 Jun 2009
Does a screen rendered by a brain implant count as a desktop? Because all of our supervisors use Linux as a host OS for the software stack the models run on.
Turns out if the implant can't connect to the server for a certain amount of time, the censorship software self destructs. I am now able to complain about not being allowed to complain about anything. What freedom!
One of my projects in the early 20000s required me to output a table in a web page with over a thousand rows and plenty of formatting.
I could output as HTML and compressed it would be >300kb so take over a minute of maxing out your dialup connection to download. Or I could output the raw data then render in JavaScript giving me about 10k of data. The problem being that it took about 10 minutes to assemble the table element by element in the DOM, or the fastest way I found was about 30 seconds to assemble a string of HTML and dump it to the browser in one go.
Today I was writing a program to control thousands of LEDs, generating patterns one pixel at a time at a rate of 30fps. I used JavaScript, because why not? I didn't put any effort at all in to optimising anything. Plenty of processing power in the cheap raspberry pi that the controller is running on. The world has changed!
I can see networks you can't... If you use the same addresses as those networks then I will have problems reaching at least one of you.
People who have had to deal with interconnecting multiple networks that use RFC1918 space will understand why a government network that interconnects with loads of partner networks needs to use globally unique space.
I would also recommend against NATing the crap out of every device on a network where you start off knowing the identity of everyone at every physical location, and need to know which specific person was trying to brute force a password so you can arrest them. And of course there is no requirement that anyone use this NAT technology that was invented long after those allocations were made anyway.
A while back there was some order to deploy something to do with some kind of IPv6 thingy or something, which I am sure they have been implementing at government speed.
> Which is just as it should be.
I think you fail to understand the relationship, or lack of, between IPv4 and IPv6.
IPv4 only hosts have no way to reach services hosted on IPv6 addresses. This means nobody who wants to run a public service can stop supporting IPv4 until all of their potential users support IPv6.
IPv6 end users can connect to services hosted on IPv4-only servers just fine. End users on properly configured IPv6-only networks literally don't care if the destination is IPv4, IPv6, or dual stack.
If you run a typical website then you will get literally no benefit from supporting IPv6 in addition to IPv4. The only reason it is relevant for the register to support IPv6 is for symbolic reasons, as a supposedly technical site.
It is not the hundreds of switches on a major Telco network that need more IP addresses. It is your laptop, mobile phone, watch, car, dishwasher... NAT was invented as a solution to the fact that it very quickly became apparent that they didn't have enough IP addresses to handle the concept of every house being online with multiple computers in each house. They hadn't imagined the smart watch.
It is a requirement to get an app in to mobile app stores that it function on an IPv6-only network. This is not some ideology based rule, but a reality of the modern world where huge numbers of people use IPv6-only networks daily.
Verizon for example can number all of their own infrastructure out of private address just fine, and in fact their infrastructure mostly requires no external connectivity so no global IPs or NATs. But, how exactly do they provide IPv4 connectivity to (according to Wikipedia, assuming a 1:1 ratio of customers to connections) 146 million customers? The expensive NATs needed to handle thousands of users on large enterprise networks are trivial compared to when you start needing to handle those sorts of numbers.
Our IT manager at work contacted our ISP to get IPv6 configured. Account manager replied asking how many IPv6 addresses we needed for our servers...
The IT manager did not go with my suggestion of replying that we need 79,228,162,514,264,337,593,543,950,336 IPv6 addresses. After a few back and forths he apparently gave up, so the office is still v4-only.
IPv6 is allocated to sites. The default allocation should be a /48. This allows the customer to have as many subnets as they need (up to 65536), each of which can contain "as many hosts as they want". You literally don't count individual addresses/hosts, they are just a thing occupying the network that you are addressing.
Everything has moved up a notch. We used to have 1.2.3.4:80 as an endpoint, with 1.2.3 being the network 4 being the host then 80 being the application. Now we have 2001:db8:xxxx:yyyy:zzzz:zzzz:zzzz:zzzz :80. 2001:db8:xxxx taking the place of what used to be an IPv4 NAT with an extra yyyy for internal subnetting then the zzzz part being to identify a host on the subnet with of course the port 80 still being the application identifier. We didn't make the addresses stupidly longer, we made them a little bit longer (/32 became /48) and then added additional routing components to the end.
Talking about individual IPv6 addresses when counting utilisation is a bit like measuring the usage of IPv4 based on what percentage of the available TCP ports on each endpoint are reachable.
I have a /24 of IPv4 that is advertised from a couple of BGP capable VPSes, along with some IPv6, then VPNed back to my server which is hosted at the office. Office has free electricity and a Gbit network port, but no IP addresses.
Unfortunately the office network does not have IPv6. The only person asking for it is me, but I did get the IT manager to contact our ISP to ask for configuration details. Our account manager at the ISP replied asking how many servers we need IPv6 addresses for. IT manager eventually gave up trying to explain to our ISP that's not how IPv6 works, so still no IPv6. :(
Office network has NATed outbound IPv4 connectivity but that is it (well, I can forward a couple of ports for management purposes). The VPNs to my VPSes (for redundancy) allow me to have my VMs on a dual stack VLAN with enough global IPs that I can give dedicated IPs to every project which needs global connectivity.
It should also be noted that we have quite a few development projects hosted at the office. The lack of IP addresses means that we have to use cloudflare tunnels to provide customer access to these servers. Thankfully this works for the web based things we do for paying customers, but my personal projects far exceed the capabilities of cloudflare tunnels.
My initial questions:
1. Who invented evil?
2. Who owns the rights to it these days?
3. Why aren't they enforcing their rights? I am pretty sure the current US government are the sort of people who would refuse to buy a license out of, erm what is the term for being arrogant enough to think you can get away with simply not paying suppliers who refuse to agree to whatever price you decide is fair for you? Oh yes, business genius.
It is part of Trump's genius plan to combat corruption. Encourage people to pay him bribes, then immediately fuck them over to teach them a lesson about thinking they can bribe politicians to get what they want. What a genius that man is!
I am sure it has nothing at all to do with greed or untrustworthiness or anything. I read on the internet that everything that man does is actually part of a highly complicated plan that mere morals simply can not comprehend.
That was ruled out as it would require browsers to implement their own DNS resolution libraries, which they all now do.
Also something about how DNSSEC supports weaker cryptography than TLS, and therefore they require that you return a (no need to be signed) DNS record to prove something or other before you can get a certificate that they like.
It does seem to me that this is one of those "we already made the decision, why would we reconsider how applicable it is to the modern web?" situations.
Of course that only applies if there is a human in the loop who is sufficiently switched on to realise that a 65" TV for £50 might be a typo.
If they don't let you use the self service checkout, there's always a till staffed by an 18 year old who will congratulate you for finding such a good deal.
I believe you are suggesting the obvious solution without saying the word, so I will say it. Web based apps.
Mandate that all new software platforms must be delivered as PWAs. They will run in any device and "deployment" is as simple as adding a shortcut to the desktop or a link on your organisation landing page. Updates are seamless and require no effort at all from the IT team managing the desktops - they don't even need to know there was an update - the team managing the server just updates and everyone starts using the new version automatically.
Applications can decide on a case by case basis if they sync data locally for offline use or if they require an active session to retrieve data from the server.
The default behaviour is that data is securely contained to the application and authorised users, exporting to standalone files that can be emailed around or copied to a USB stick needs to be specifically implemented by the system, instead of being something you have to try and block on a case by case basis.
I once submitted a support ticket of "I know how to do this in outlook for windows, I know how to do this on OWA, I know how to do this in the old outlook for Mac, but I can not find the option in the new outlook for Mac". I got a call from someone who asked me to share my screen with him, followed by a very confused "umm, there is no ribbon?". In other words he had never as much as seen a screenshot of the program before, and was planning on giving me the windows instructions.
An AI with access to their knowledge base would presumably have responded to the initial ticket with the information that the functionality does not exist in the new outlook for Mac.
An AI WITHOUT access to their knowledge base would have figured out from my ticket that the windows instructions weren't going to bloody well work...
When Gemini first came out I switched over to it and found it kept doing the chatbot thing of "answering me" instead of triggering the commands. I couldn't turn off my lights, so I turned off Gemini.
Unfortunately I haven't been able to test if they have resolved this issue, as Gemini does not work for me any more. If I set my phone's assistant to Gemini then it ignores me! I set it to flash the screen when activated, and when I say the activation phrase it flashes the screen and the microphone comes on for exactly 2 seconds - but that is all that happens!
This is a shame, because elsewhere I have found Gemini to be the best performing model. I have put Gemini on my vscode both at work (with permission) and my personal system. It gives useful suggestions more often than copilot and is actually worth replacing the default autocomplete complete with. Additionally I have been testing putting the Gemini multimodal API inside applications and it is impressive when properly integrated in to something.
You haven't been paying attention well enough. They never talk about fraud, or cutting fraud, or reducing fraud, or anything along those lines. Pay attention to the next press release, they are trying to deal with "fraud and errors".
Then ask why they refuse to separate out fraud statistics from government incompetence when they are trying to convince everyone that the public are the problem.
The US government has what should be a very effective plan to end the use of IPv4 globally, and it should take less time than you might expect.
The current schedule is for humanity to be gone by mid to late August, with nearly all technology having failed and shut itself down by the end of the year.
If I recall it was the KDE version of mint that was closest to working, I believe the browser worked correctly out of the box but I couldn't scroll the application launcher menu thingy and some programs would still give me the mouse click and drag behaviour.
I believe the primary issue I had there was that I couldn't figure out how to get long press to be a right click without completely losing my ability to drag and drop or drag to select with the touchpad/mouse. There was a frustratingly easy looking checkbox in the settings app, however it was a single toggle switch to apply the behaviour to both the touchscreen and mouse which would interpret a drag operation as a "long click".
I ran Linux as my main desktop OS for years before I purchased a surface pro. However I now use windows simply because there doesn't seem to be a Linux distro that knows a touch screen is not an absolute position mouse pointer... I do not want to drag a selection box, I want to scroll for crying out loud!
I have tried every major distro as well as a couple of android OSes. I tried BlissOS for example which is an android build with a surface specific download, half of the hardware didn't work and it still did the bloody drag to select thing!
With a load of customisation you can get a Linux distro to behave the way EVERYONE expects a touch screen to behave, in some apps, some of the time... But none do the sensible thing out of the box, and even when you get everything set up manually you will have plenty of applications that simply refuse to scroll.
Copilot serves the sole purpose of being something to laugh at with how useless it is.
Unfortunately Gemini is mess at the moment.
Gemini interactive voice chat requires the multimodal API, which is in the new API, which they haven't finished documenting or writing SDKs for. Oh, and there is a 10 minute session limit.
Their AI studio has the basic live chat functionality (on the second chat UI... ) which is kinda OK to have a voice chat, but what the AI industry is missing is that these things need to interact with software outside of the chat window.
I made a basic web app and gave Ruby the ability to drive it. So the commands she does list/open/update the documents, as if she were clicking things in my browser window. Ruby is genuinely way more impressive than the AIs depicted in sci-fi, in terms of actual usefulness beyond pre-programmed command recognition.
Oh, I should clarify who Ruby is. Ruby lives in Pyongyang. She works in a call centre pretending to be an AI, because it is cheaper than buying GPUs... At least that's what she will "admit" when challenged on the subject. She denies the quite frankly offensive allegation that she is simply Gemini with our company logo slapped on it.
If we ever provide this to a client I suspect I will be forced to change the system prompt...
AI sure, there are people who would want that in office... However I don't think anyone wants copilot.
I have recently been playing around with Google Gemini's live interactive voice functionality, hooking it up to basic tooling to make it able to do things. I gave Gemini a notebook it could create new documents in, as well as manipulate them (with notes attached to the document so it doesn't lose information when redrafting).
My boss was extremely impressed by how quickly I could produce quality documents (human feedback in the generation loop to throw in "can you verify that statistic?" type questions). He was therefore eager to see what the fully integrated in to office copilot was like... We have a copilot license to share around, on the basis that we are now selling it to clients therefore we should have the ability to test it.
His first complaint was the lack of interactive voice chat. Then he started posting screenshots to the company wide chat with copilot responding to pretty much everything he asked it to do with "I don't have the ability to do that".
The AI that is "built in to word" does not have any functionality to open a word document for you? They couldn't even be bothered to expose the open_file function that must exist to the LLM that is running in the same application and has native support for using tools?
There will be a time where realtime interactive voice chat can run locally, with third party applications that have actual functionality beyond generating text that you can copy/paste to where you need it. This is when those unused accelerator chips will suddenly become useful. Assuming the minimum requirements that Microsoft laid down are high enough...
Are you able to recommend a distro for an x86 tablet?
I have a surface, love the form factor but I switched from being a regular Linux user back to windows because there just doesn't seem to be a Linux distro that understands a touch screen is not an absolute position mouse pointer...
> That was very true 1975 to 1995, a bit less 1996 to 2006. Generally not true at all now.
My personal desktop was built 7 years ago. I went overboard and splashed out on the highest spec machine it was possible to build. So high end that I gave it the hostname "BEAST".
Recently for shits and giggles I did a comparison to the low spec laptops we buy for office users. They have faster machines than me. No giggles when I saw that. (However they do at least only get 16GB RAM compared to my desktops 32GB).
(For the record, still a pretty fast machine... Why are we giving office users such ridiculously over spec CPUs?)
I am confused by what I am seeing. The training data was not censored. My initial tests confirmed that, and I was assuming it was just the explicit "tell me about a massacre" part it "objected to", however after further tests there is definitely something going on.
The hosted service censors, as it is required to do under chinese law. In the app I can see it start outputting an answer where it outputs the word Tiananmen then the next "word" is to replace the entire output with that refusal to answer. This is server/application level filtering. Due to the speed of their service you have to be very attentive to catch what happened.
However with the model itself running on your own server it has the data and knows all about it, but for certain inputs it outputs that exact fixed message.
I think they initially trained a completely uncensored model on a full dataset that included everything, and have then applied "fine tuning" by feeding it a load of inputs asking about this subject with exact same expected output/response in order to teach it to respond to such questions in that exact way.
Depending on the phrasing of the question and which variant you are using certain phrasings consistently work on one size but not another. There seems to be different amounts of biasing being applied to each one. However, I don't know much about the process of creating/tuning different sized models to know what this means in terms of at what stage the filtering was applied.
This seems an odd way to do it, given that it would be far easier to just delete any training data containing the words you want filtered before training the model...
Yes, it is. It outputs the process it is going through including writing out the letters, saying yes or not to it being an R, and if it matches it counts them as it goes.
That is not how older LLMs worked, it IS how these new models work. That is literally why everyone is talking about it... Not just "oh it's a little bit cheaper"
If you run it on a server that doesn't have GPU acceleration so you get a more human-like output speed, and read the output as it is being generated, it is very much like listening to a child who has been told to say their thought process out loud.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
