They absolutely are stealing people's credentials. These stolen credentials power the massive network of "transfer station" relays used by China. It's in Anthropic's threat report , and also reported on by CISA.

Yes, and also different behaviors in that they can remove the safeguards. After a distillation attack, the "student" model can be trained without safeguards.

As noted above, the Chinese transit router attacks constitute a massive credential stuffing attack by means of using the "transfer stations" that rely on hacked credentials, API keys, and session tokens.

True, but using illegally obtained credentials to learn from your competition, and then conducting trade secrets and intellectual property theft so that you can profit, is definitely a crime.

Now, whether they deserve it for training their AI models using massive amounts of other people's intellectual property, and every comment me and scores of others ever posted on StackExchange is another question.

I agree with the point that they do this by training their models on everything on the internet, including our Slashdot comments. There is definite irony there. But much the same way as 23AndMe suffered a "credential stuffing" attack where DNA databases were extracted by using many logins to find matches and gather data, distillation crosses over into an attack when you use industrial-scale to intentionally break guardrails and get the AI engine to essentially spill its secrets, often by means of a credential stuffing attack in the first place. This is when it crosses the line from model distillation, like taking an open weight model and reducing it for a finite purpose, and a massive industrial-scale effort to exfiltrate as much information as possible, which would not be available in legitimate use cases. So semantically what China is doing is definitely considered a type of attack.

That would make sense if you were able to downgrade a LightSail instance to a lower tier without migrating to a new instance. You can't, you can only upgrade, so it becomes quite a bit of work. A reverse caching proxy, e.g. native CloudFront, doesn't help on the slowest part of WordPress, the dynamic PHP content e.g. wp-admin, editors like Elementor etc; and that slow running mySQL database. Realistically many users aren't even technical enough to do this and end up paying wpengine or GoDaddy exorbitant costs for hosting. Now products like Google AI studio, which deploys to Google Cloud Run, can do it for very little money in hosting and zero DB maintenance, security upgrades etc.

I've tried the LightSail bitnami instance on that $3.50 tier and it is far too slow to load and even login to the WordPress console and do anything without waiting and waiting. Page load times are abysmal. Also it doesn't include a public IP at that price point, so it's at least $5/mo for the slowest option to be available on the internet. People may not need to use Lambda, as this is for scripts/dynamic content/contact forms etc; but many do care if it runs on S3 as a simple static site. No need for a dedicated public IP either, it runs a CNAME to an AWS shared DNS record. In this S3 setup, pages load extremely fast.

If there is no such thing as a "distillation attack" why is it tested on in security certification exams around AI security and even by Anthropic itself?https://www.anthropic.com/news/detecting-and-preventing-distillation-attacks

You are right in that most of it is a binary, and it's not the same as open source because you don't get the training data and can't train a model using their software. But you can configure it and change things, fine tune it and distilled it into a smaller focused model. Things that are impossible with traditional binary software. So somewhere in between freeware and open source, maybe like binary software where much of it is in plaintext scripting files.

One word: China. Chinese AI models are already available open weight (essentially open source equivalent). Many of them were created via large scale distillation attacks against Anthropic and OpenAI. They are competitive in quality with American models. A slowdown deprives China of this key source of training and data for their models. They can then build in secret internally until the models are sufficiently advanced, and defenses against distillation significantly hardened; so they can take the lead. All the while not having to immediately show advancement over rivals to investors. I've never met a Silicon Valley company that decided to slow down for reasons that weren't also aligned with business interests.

It reads like rats fighting for a piece of debris to float on as the Titanic is sinking to me. WordPress is a bloated web platform, riddled with security problems and a vast ecosystem of mostly crappy plugins trying to get people to pay a few dollars a month for some basic functionality. It needs to be run on a fullon virtual server and can't be served from things like AWS Lambda/S3, Google Cloud Run on demand, etc; bringing up hosting costs, for small sites this is the difference between $20 a month and 50 cents a month in hosting (or free if in the free tier of cloud providers). Now with AI, it's simple for anyone to make a site (or clone their existing website) in native HTML/CSS/JavaScript/React, which loads much faster and tends to work better across devices and screen size. WordPress' days of dominance are numbered, especially with this unstable Mullenweg guy at the helm.

For sure, a (large/well-founded) university or a big corporation could run these, but I'm not sure who would call that "local" so much as private datacenter or internal. You could run it on AWS Bedrock or Azure foundry too and those don't really share data back to the AI companies, or zero data retention cloud for open weight models. I think "local" as having maybe an rtx5090 or a super up Mac studio, which will very slowly run some lightweight models but I'm not sure if those are particularly useful for anything beyond high school biology papers.

What kind of hardware is he running? To run frontier open weight model like Kimi K3 you need at least 500k-1mm in compute equipment, meaning minimum two NVIDIA DGX B300s; which are often unobtainium due to high demand. Then you need ridiculous cooling and electricity, not something you put in a garage or office server closet. Or you can use a zero data retention cloud provider and pay per usage...

I always thought the prepper bunkers sold to the middle class and rich alike were a useless scam. Most are located well outside the metropolitan areas they typically live or spend most of their time in, and require transportation to break through massive traffic and other pandemonium to get to. Helicopter pilots aren't going to care about their clients then, they're going to save their families first. Money and even gold etc will be instantly worthless, currency becomes shelf stable food, water, shelter and ammunition. The contractors who built the bunkers will know where to go and loot them, and tell their friends where to go too. Their "rightful" owners will be greeted with empty bunkers or someone else wielding a shotgun.

Did you only watch the pilot of the original series, where Captain Pike said he couldn't get used to a woman on the bridge? If you watched it further than that, it was considered pretty woke for its time. I'd also argue if more men behaved like Jean-Luc Picard, there would have been little reason for the woke movement to exist. Be a Picard, not TOS Pike (but maybe SNW Pike, he's alright).

If the financial reports, EBITDA calculations, salaries and numbers of machines managed, OSes, RMM/MDM software reports and other data from numerous MSPs also lie, then sure. Short of that, no. FWIW, early in my career I was a *nix admin and I actively loathe Microsoft software, but real and recent numbers tell a different story about administration costs.

Perhaps this was quite unlike Windows in the 2000s, but today with automation like MS InTune and usage of powershell scripts the gap isn't there anymore. MSPs managing Mac, Windows and Linux endpoints see little difference in the headcount required to manage them, many of the popular MSP toolkits like NinjaOne for example support all three OSes, and AI can port custom automation scripts easily across platforms. Much like the term UNIX admin is no longer in use (Cloud engineers/DevOps/SRE etc are the titles now), your description is historical and dates back some decades.

Harvard is an educational institution with a massive endowment that makes them independent and have to answer to no one. Apple is a publicly traded company whose one objective is to make money. The board is compelled to fire a CEO who fails at this and replace them with someone who can deliver more money to shareholders. They did this to Steve Jobs in 1985. If they fail to do this corporate raiders can sue and even conduct a hostile takeover. Harvard's board is concerned with their brand reputation, so graduates of their schools can proudly weave "I went to Harvard" into every professional conversation they have for the rest of their lives. Harvard has incentive to weather short-term monetary losses to preserve their pretentious brand image. There is no risk of a shareholder revolt or hostile takeover. No matter what happens, there will always be more students willing to pay to go to Harvard. Problems in higher ed will affect and lead to the closure of lower ranked schools first. A school like Harvard that counts founding fathers like John Adams as graduates works like the Catholic church, in decades and centuries, not in presidential terms or daily stock value changes.