hmmm... strange.

I'm presenting this to the head of supercomputing for a NATO country on Friday because... oh wait... I am one of those educated people with access to 10 computers on the first page of the top 500. I am very sorry to disappoint.

Second I'm presenting my research and findings in a big room at Huawei Connect in China later this year... because crackpots need love too :)

Thanks for all the fish

i wouldn't pay for an OpenAI product, that's an ethical issue. And forget data sovereignty. I don't feel comfortable supporting people like Sam Altman and the CEO of Anthropic who I'm sure spends his mornings admiring children in school yards.

But that said, every university in Europe is producing the same thing and let's be honest, training new models has becomes a lot easier these days.

They buy them for the customers.

What you do is you buy a company who has a lot of locked in customers for the next 5-7 years. For the first two years, you keep the rest and vest employees who perform to the bare minimum until they can cash out. Then the customers start leaving one by one. When the cost of maintaining the product gets high, you hire 5 people from India to take over. You then wait for all service agreements to expire and you end it.

PCoIP never stood a chance. HP likes to sell boxes and lots of them. The only "long term service" they understand is high end printers. And the bitch of that is, they're not doing long term anymore. A few years tops. I buy a new used plotter for $500 every few years because HP will increase service cost so high the customers can't keep them. Oslo, Norway doesn't even have HP printer service anymore as far as I can tell. Yes, I know it's two companies... but HP and HPe are still not really doing it.

When HP spun off their engineering tools division, it was the end of the company.

The NSA almost certainly needs to investigate the possibility of a threat. As such, they'll use it test for these vulnerabilities within their test network and then possibly on a larger network. They will either deem it a threat or not.

That said, there are a hundred open models which are always les than two steps behind Anthropic. So, assume that within a month, there will be a new open model matching or outperforming this one. And it will be a MoE and it will be untraceable.

So... if Anthropic can't release theirs... don't worry. Anyone who has tried the pubic release of Qwen 3.6 35b on a RTX 3090 and a pile of MCPs knows we don't have long to wait.

It will need to be replaced, who pays: Amazon or the supplier/manufacturer ?

Why would evolution no longer apply to us ? OK: our technology, medicine, etc might mean that some pressures might be reduced but they are still there and others will appear. Eg: we do not need to be as strong, so weaker people are not so selected against, so more weaker people survive; our modern diets are causing health problems, so there will be pressure to better cope with them; ...

Into the hands of Trump supporters. Not until the facists are out of office and citizens demand the restoration of democracy and corporations be held accountable to regulations.

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

The message seems clear: If you want these features you must buy more recent models. But I ask myself: how long before these new models have features removed to get me to buy even newer stuff ?

Presumably these TVs were marketed as having these features - so, in some jurisdictions at least, this would be illegal.

Nation states will have sleeper agents who will grab a copy, send it home and go back to sleep waiting for the next big thing. How really dangerous it is - time will tell.

It depends on the audience who will read the articles that s/he writes.

If it is clickbait chasing nonsense about pop singers or film stars latest affair or wardrobe 'malfunction' then the readers are unlikely to be too critical unless you do not have enough pictures of naked flesh. Your editor & publisher will be happiest if you write lots of articles and care little if it is slop.

If you are writing about something supposed to be factual, eg: science; finance; politics; ... then the articles should be well researched & checked and any uncertainties noted in the article. You will be rewarded and applauded by your readers for insight, good context & few errors. Your editor/publisher will like many articles but accept that quality takes time. This is not quite true: if your publisher has a strong political bias then you will be expected to follow that bias and invent facts support that bias - ie lie and produce fake news. So as long as AI slop has the correct bias many will just publish it. Journalists of integrity will not want to work at such a publisher ... however journalists do have mortgages & kids so some 'bend' their professionalism.

The first Unix machine that I used had 1MB of RAM and it supported about 10 simultaneous users. OK: this was the 1980s and it was a PDP-11 (16 bit machine), but how times have changed!

BTW: Mb means megabit, MB means megabyte.

but the market capitalisation ie the total value of all publicly traded shares, but even then AMD beats Intel at £268.77 B vs £191.72 B but by not so great a margin.

Beware it shows me the market cap in British pounds, but that is where I live.