First... We can't release this model because it doesn't work

Second... We need to convince the Christian right that they should use their influence to force this tech down everyone's throats.

Anthropic is going to go public, but this should be considered gross negligence because they are knowingly asking money for something they know can only decline.

Try the open models and tell me that they aren't good enough to replace Anthropic in 95% or more cases already. And how will Anthropic compete with free?

Why open models matter? Well, it's only a matter of a few years before even miniscule devices will be able to locally host AI.

Here's the next thing. You need to see AI as an onion. Neural networks are a series of layers. Last week, I was playing with running layers at differing cost levels of hardware. I uses a cluster of H200s for the outer layers and I used <$100 AI accelerators for the inner layers and I used an RTX3090 for the middle layers. I then tested coding and general nonsense like "what eyeshadow matches these earrings" questions. 85% of all questions were answered quickly on the $100 accelerator. 99% were answered with the cheapest two options. And remember, I wasn't running a small model, I was running a gigantic model sharded across a $100 device, a $1000 device and a $500,000 device. I reduced usage of the $500,000 device to almost nothing. I managed to achieve the same results at about a 20% performance drop on a 1 trillion parameter model while increasing compute density of a cluster of H200s by 100 fold.

So, what this means is that using extreme MoE models sharded properly and adding what currently is a $100 accelerator and soon will be a $5 accelerator and a thin layer in-between, assume a single RTX3090 class card for 1000 users (500 for better performance).... the case for massive inference data centers is screwed. Give me a grant and a few months, I am 100% sure I can get efficiency closer to 10,000x rather than 100x better. And no, this is not exaggerations. I would retrain the models to be spread across more... thinner layers with a LOT more experts. Of course, retraining something on the scale of a $1 trillion parameter model is expensive. What's great is, there is true value in China footing the bill for this because cutting their dependence on gigawatt data centers filled with NVidia and tons of HBM memory (possible literally) is a survival requirement.

If there's anyone in China reading this, take Qwen or Deepseek, spread them REALLY REALLY thin... then distribute the layers and open the weights. You'll make it so that companies like Huawei and the others can layers locally on devices as small as ESP32 and the distribute the layers outward. It was LM Studio's magical cross platform sharding which got me going on this. It just works. It's so simple. It just works.

It doesn't matter if it's Google, Meta, or Apple, it will happen. And the government will LOVE IT. Because it doesn't matter how good the techbro lawyer are. The government will gain access to the data. It would save many many billions in surveillance. It would place the burden of law enforcement on the techbros and they'll pay the price gladly for access to the personal data.

And if the heads up thing isn't good enough, expect everyone to start wearing cute hairclips, headphones, etc... that do the same thing. Glasses are nice for people like me. And the best part is, I would be the best of the assholes because I generally wear my glasses facing the ceiling until I need to read.

It's coming and it will be here soon. I believe even now, I could probably make a video capturing hair clip with android integration and all the fun stuff for maybe $25. And I'm sure Zuck can do it cheaper.

I think it will be funny when it becomes normalize for old men like me to wear hair clips

U.S. Patent No. 10,855,990
This is old technology, but was used extensively in JPEG and JPEG2000. All these patents are and have been long expired. There is no novel approach in U.S. Patent No. 10,855,990. More specifically, all the claims they're making in terms of the specific violations of this patent were covered in ITU-13818-2. Though ITU-https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.itu.int%2Frec%2FT-REC-H.263-200501-I%2Fen hammers the last nails in the coffin. I have read and reviewed the claim and the patent and the technologies presented in 10,855,990 are just reiterations of earlier work with scrambled wording to try and give a new name for variable sized macroblocks. They novel approach implemented in H.264, H.265 and H.266 was the method of selecting which specific pattern of "coding units" to apply. I have not checked for reuse of this, but this is neither in 10,855,990 or the claim. So, I believe they checked and found out that there was no violation. Oh and to be clear, they're completely fixated on the sharing coding parameters between blocks. Their approach is almost, barely, kinda novel, but the fact is, I'd make a strong argument that this is obvious, it's basically just macroblock grouping which has been part of standard video coding as far back as MPEG-1 and ASF. And the method applied could easily be argued to be an almost direct copy of LZW compression.

U.S. Patent No. 9,924,193
I couldn't find a copy of the original text (not wearing my glasses) and frankly their description was so TL;DR that they just started making things up. Ok... here's the argument against this. This has been a core features of all DWT based compression methods since the start. It was even the reason we used DWT. JPEG2000 is almost entirely based on what they're claiming here. If I spent an hour on this one, I could tear it to pieces without even trying. And skip mode... what in the world do you think something like Google Earth is?

U.S. Patent No. 9,596,469
Encoding data in a way that would allow independent parallel decoding of different portions, bands, blocks whatever of the image ... blah blah. Back to JPEG-2000 and Google Earth and stuff like that. The first time I saw this personally was at Disney Epcot Center when there was a Cray Supercomputer on display showing off a google earth like experience. The computer was streaming data at different spatial representations in parallel to hundreds of CPU cores who were all decoding and texturizing. The number of patents filed and expired on this one tech is immense. I haven't dug up specifics, but I can guarantee that the JPEG2000 patent pool clearly invalidates this

I just doom scrolled through the rest of this. I highly doubt I'm the only signal processing and video/image compression historian out there. I'm guessing that the LLMs could easily tear this crap apart too. But I'd be willing to make a few bucks as an advisor on this. I've either worked with, against, for, on, etc... on every technology being claimed here and I did this 15-20 years ago... and the tech was already old.

The prompts provided to the LLM should be copyrightable as code and the code generated should be protected the same way compiled or intermediate code is protected.

The issue at hand is how the model was trained. And the users of the LLM should be made clearly aware by the model trainer whether the user or the model trainer is responsible for the liability related to using other peoples code for training the model.

That said, we should soon be seeing models that are trained using training courses rather than massive amounts of code. Once that happens, the models will make use of agents to search the web and learn from stackexchange or other sources how to solve problems the same way a human would. Of course, when a human learns how to do something from searching the web, simply copy/pasting other peoples code can be an issue and we have to read license restrictions. But learning how someone did something and doing it ourselves is generally safe. If a model reads an articles while searching and then learns how to do something, it should also be protected if it's not copies verbatim.

We have a lot of legalities to deal with.

1) Massive models are going to die. I don't know whether it's with wasting time with nonsense like OpenAI and Anthropic. They won't even be in business by the time the lawsuits come through.

2) Agentic models will be the focus of the future because they work more like humans. We give them the base information needed to learn and find the answers themselves. Using cloud based solutions where companies hosting the solutions keep massive amounts of data locally cached so the model can research faster could be an issue. But these will cost money and really just won't do more than local-AI will. They'll just be faster. For legality sake you'd want to avoid the cloud models since caching can be seen as theft. But, local-AI is much different. With agentic solutions, I think most legal issues are back to the same issues with copyright we always have. We just have to make sure our models which we use are following the copyright rules. If it's allowed, copy/paste. If it's not, then learn how it's done and make your own solution. In a perfect world, we'd then have a stackexchange or alternate github for AI generated code and post that different LLMs could use to learn from each other. The problem then becomes whether that would be seen as training a large model and whether they are in violation of copyright again.

So, the rest of the world can continue as usual and the US will be isolated.

Now, will the US owned TikTok or the Chinese owned TikTok win the competition.

I hope US tiktok gets banned here since I think the US just has too much influence on youth outside of the US

I expected it to take 6 more months. But, if they're playing the victim already, either it's time for them to cash in on patents or to sell off their shares while they can.

That's kind of hitting on the point but not quite.

IBM mainframes are online transaction processing systems. The language hasn't been and issue for a long time and it really doesn't take more than a few days for a programmer to learn to use COBOL. The problem is that JCL, RPG, CICS. DB2 and all the surrounding infrastructure is very confusing.

The uptime you're talking about is that a mainframe is basically a special purpose computer built specifically to make it so you can suffer loss upon loss upon loss and it will still keep processing transactions. This is because they have a specific workflow which is designed specifically to support this. I have implemented the exact same topologies and workflows using more modern tools on small computers ... because frankly you don't need big computers for this. And it works. I can scale almost infinitely and lose all but 2 nodes and it will keep chugging. But I'm not going to provide support on it and I'm not going to back my project with insurance company support.

IBM is offering a lot more than computing in the price they charge. That's the real issue. And no other company has built that up.

Great experiment but of course glass has some pretty obvious downsides. We tried something similar some years ago using a DVD laser and coated glass. It is extremely reliable and thankfully cheap. The problem is... how do you read it? There's no instructions.

The worst movie ever so see a screen... Contact... should have taught everyone in this business the rules of this.

"The medium is irrelevant if no one can read it"

and

"You have to leave Jodi Foster a key to be able to build a machine"

So, the beauty of metal is that you can etch it clearly at many levels. In fact, it makes an amazing analog media. What you do with the metal is that you calculate the analog image to etch on it. Then using a laser and the scanning head from a laser printer, you can pass the scanner across the metal and print. A4 paper size or a nice 210x210 square is great for this.

What will you print. This is interesting, you'll layer many images on top of one another at different resolutions.

The first image show clearly in mostly pictures that are readable with the human eye which will explain how the second layer can be read. It should be simple enough that the reader can build said machine by hand using simple tools. Or, they should be able to read it by hand by measuring. See, you're explaining binary and a simple table such as ASCII or a 5 bit subset of it. It should also contain a Rosetta stone to allow linguists to decipher the language.

The second layer explains that we're storing information that should last forever and is a history of our world. And it should describe how to read the third layer which is much denser and contains a more advanced machine. But because the materials required to read the machine may not be available, it also describes the method of storage as well as detailing the more advanced character set. It should describe that the card which has all jagged edges is a dictionary containing 10,000 commonly used words and their definitions. But that each layer and card will contain partial dictionaries around their edges at the 3rd layer.

Density wise, We've managed to simulate 6 layers on increasing density and complexity allowing for about 1tb per A4 sheet of 1mm platinum (I'm sure other metals will work, it was just a good starting point) and the layers of course last more or less time based on the frequency and detail of the layer. Of course, I included extensive error detection and correction on more detailed layers. The simulation suggested that we could store for about a million years (no real way to test) and that most data would remain readable.

Oh, and you could build the devices fairly inexpensively.

We've been testing this in 3 different nation's archival departments. It was really funny because when I met the people from the archives at a symposium, I asked why they weren't doing this. They were like "what" and I just blurted out the design while chewing my lunch. I mean, I thought the idea obvious and they had been wasting time on all kinds of silliness like what Microsoft was doing.

People need to watch more bad movies.

Dude, I like going to the shooting range. I think it's fun. I don't imagine I'll ever own a gun and I'm not particularly interested in making guns.
That said, I think the second amendment is definitely taken greatly out of context. We are honestly using the thoughts and ideas of people who lived 200 years ago to have the slightest idea what makes since in terms of things like state run militia. And we're also using their perspective on what makes sense when the entire population of the US was 2.5 in 1776. 2.5 million people barely counts as a single city in 2026.
3 out of 10 people I know who own guns are precisely the people I would not like owning guns. These are the people who own guns because they're really into guns and they're really into gun ownership rights. I'm pretty ok with the other 7 out of 10. So, I think gun ownership should simply be highly regulated.

So, solving a network issue, 4 senior network engineers with a combined 70 years of experience and close to 100 Cisco certifications between us (at least until I decided not to bother anymore) were trying to get a complex VPN routing issues between 3 sites working while not sacrificing security. Everything worked except this one little thing. We use hours troubleshooting... and the pizza arrived.

I configured SSH (with temporary keys) to login to all three routers and three endpoints with administrator privileges. Then I opened SSH and explained
"This is the details of how to connect to these devices. These devices are located on these sites with these addresses. These devices should be able to communicate like this, not like this. We are sitting down to eat pizza. I can't be bothered with this shit anymore. Please fix it."

So, we 4 sat around watching the 85" TV like it was a sporting event and the AI chugged as we ate pizza.

And we watched...

Right as the AI started down a rabbit hole like "If we went here, we know that it's time to start reinstalling and go from the beginning".

And then it said "The systems can now talk together as you described".

WTF!!!!!

I wiped the grease of my fingers and we checked and I wrote

"You listened to the letter of the request, not the spirit. You got two computers talking, not the networks. Please fix the whole problem"

And 5 minutes later... it was done.

The AI was better than we were and we were like "Thank god we can make it to retirement.... the kids are fucked"

If only someone though it would be a good idea to make a much cheaper design that could be tried over and over again until the kinks were worked out rather than spending obscene amounts of money on trying to get something so unimaginably complex right on the first try and it costing so much that you can't even financially budget for another try.

I have been cutting bloat like Tailwind by using LLMs. After all, Tailwind is a tool which makes life easier for me as a programmer at the cost of efficiency. I basically stopped using web toolkits now. I just tell the LLM to build from scratch what I need most of the time.

I think the best one was when I wrote a parametric CAD program recently (not a web project, C# on the web this time), and I needed a specialized PDF export engine because the existing ones cost money or sucked. So, I told the LLM what I wanted and it did an amazing job. It coded everything I needed for high quality PDF output with font embedding that actually works with Affinity (no font embedding works with Affinity), and it did it in 30 minutes or less.

Tailwind is definitely a goner. We don't need CSS tools when the AI's can do the job more efficiently without them... and no, we really don't need to be able to read the code. We have unit tests for verification.

Give credit where it's due.

I basically stopped using Google most of the time because I could use Copilot for most things. So, I suppose if I were to measure, I google about 70% less than i used to. I mean, most of my googling was figuring out how to do things and these days, I spend my of my time telling copilot to figure out how to do things instead.

That said, I tend to Google when ChatGPT is failing. And well, it fails a lot. It's really just not a very good product.

So, then I use Gemini through Google and more often than not, it gets it right when OpenAI bombs it.

Gemini has become a better set of models than ChatGPT. I probably wouldn't even use ChatGPT if Windows wasn't so utterly intertwined with it.

That said, I pay $10 a month for AI. I have my own LLM server and it's based on a $120 graphic card and it's getting REALLY good now. I don't think I'll be using cloud llms much longer. Thinking models don't need to be big. So, a 10-16GB GPU should be enough. 24 would be nicer for a longer context length though. It's pretty funny that Qwen 2.5 7b actually outperforms the biggest and baddest models if you use it agenticly and tell it to just figure it out. It doesn't need to know absolutely everything. it only needs to know how to research and take notes as it goes along.

Is this the new version of saying something like 'Thanks boomer"?

I was able to use this fancy new invention you may have missed called a hyperlink to look at your commenting history. Apparently, like myself you spend too much time on Slashdot and make allusions to things like tape recording off the radio which even suggests we're in the same approximate age group.

The big difference is, I write shit comments the length of small novels, you like one liners. We both seem to follow the pattern of offering nothing of any value and coming off as arrogant prigs.

I see you like to say things like "I use slackware with systemd" with pride rather than shame. Do you also listen to vinyl for that true sound and use a tin percolator to get that coffee how it was meant to be flavor? I assume you also take pride in walking past the Starbucks and every time you pass one you point out that their coffee is .. actually what do you people use as an insult for mainstream coffee?

Anyway, I assure you, there was no ChatGPT involved. I'm just a sad and pathetic late-middle-aged man like yourself. I tend to prefer to embrace intelligence rather than embracing masochism so I might to hold on tight to that lovely old tech because it's my woobie. I use Slashdot as my outlet to try and see if I can get someone else to agree with my opinion anonymously... again like yourself.

  The one liners... god.. they're probably even more annoying that my books. It's like a MAGA hat, designed to be endearing to people who can't handle more than 4 letters in a thought.

What surprises me isn't that China has been doing it. I think that the UK was in fact the country which glorified spying and government espionage for decades. Even now, I still remember Gorbachev crediting James Bond as the biggest single cause for the rise of Putin. The UK LOVES spying. They think it's sexy. They love the suits and the cars and let's never forget Q, the true hero in the movies.

So... why would the brits ever be surprised that someone would be spying on them using tech?

I think what's really amazing is that they didn't realize this sooner.

Doesn't this kind of prove that the UK government is criminally stupid that they never stopped it... and yes, it is absolutely preventable. Even the US manages to secure presidential communications and we see what their leadership looks like.

So, here's the real question.

If the brits have done nothing to secure their prime minister's communications, then who else is listening?

Next, how did it happen? Did they do it using government mandated back doors in encryption? After all, only criminals have anything to hide.

Even better... shouldn't we be impressed that the technology allowed probably every government with a spy network to all spy on British leadership without each other knowing it.

Or did they all know it?

I can almost imagine spies from the US, Russia, Israel, North Korea, China etc... all sitting at a Starbucks listening to a speaker together. Hell, I can even imagine at the next table, a british spy, an american spy, etc... all listening to Xi. Do they date each other? I mean, think modern nerd 007. Does the British spy get the girl? Maybe a sexy hungarian who is in the process of spying on Putin?

What about when people need sick days? Does the Chinese spy call the Taiwanese spy and ask "Can you share your notes with me today? I need to stay in bed"