Comment Most likely the stupid kids (Score 1) 34
Smart kids who want to learn don't destroy their tools
Submission + - Anthropic can now track the bizarre inner workings of a large language model
tomatoguy writes: Having psychology-adjacent interests (and perhaps because it's a Friday afternoon), I found this fascinating.
What the firm found challenges some basic assumptions about how this technology really works.
MIT Technology Review
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.technologyreview.com%2F2025%2F03%2F27%2F1113916%2Fanthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model%2F (paywalled)
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Farchive.is%2F4mujU (free)
= This caught my eye first: studying something that claims to be brainy using brain-investigation tools and approaches.
Anthropic says it was inspired by brain-scan techniques used in neuroscience to build what the firm describes as a kind of microscope that can be pointed at different parts of a model while it runs. The technique highlights components that are active at different times. Researchers can then zoom in on different components and record when they are and are not active.
= Secondly, LLMs get "better" when they know when to shut up.
The latest generation of large language models, like Claude 3.5 and Gemini and GPT-4o, hallucinate far less than previous versions, thanks to extensive post-training (the steps that take an LLM trained on text scraped from most of the internet and turn it into a usable chatbot). But Batson’s team was surprised to find that this post-training seems to have made Claude refuse to speculate as a default behavior. When it did respond with false information, it was because some other component had overridden the “don’t speculate” component.
What the firm found challenges some basic assumptions about how this technology really works.
MIT Technology Review
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Fwww.technologyreview.com%2F2025%2F03%2F27%2F1113916%2Fanthropic-can-now-track-the-bizarre-inner-workings-of-a-large-language-model%2F (paywalled)
https://ancillary-proxy.atarimworker.io?url=https%3A%2F%2Farchive.is%2F4mujU (free)
= This caught my eye first: studying something that claims to be brainy using brain-investigation tools and approaches.
Anthropic says it was inspired by brain-scan techniques used in neuroscience to build what the firm describes as a kind of microscope that can be pointed at different parts of a model while it runs. The technique highlights components that are active at different times. Researchers can then zoom in on different components and record when they are and are not active.
= Secondly, LLMs get "better" when they know when to shut up.
The latest generation of large language models, like Claude 3.5 and Gemini and GPT-4o, hallucinate far less than previous versions, thanks to extensive post-training (the steps that take an LLM trained on text scraped from most of the internet and turn it into a usable chatbot). But Batson’s team was surprised to find that this post-training seems to have made Claude refuse to speculate as a default behavior. When it did respond with false information, it was because some other component had overridden the “don’t speculate” component.
Comment Re:In other news... (Score 1) 48
The technology for verifying the location of chips would rely on the chips communicating with a secured computer server that would use the length of time it takes for the signal to reach the server to verify where chips are, a concept that relies on knowing that computer signals move at the speed of light."
My ping time suggests that computer signals do not in fact move at the speed of light. This sounds like a massive pain in the ass false positive waiting to happen. If the US wants to make sure nobody wants to buy its chips, this is probably a great way to do it.
This is true.
A rule of thumb is that wired devices run around 60-70 percent of the speed of light. And that in free space, Light moves around a foot per nanosecond.
IT would be pretty complex anyhow. And then there are VPNs, which can be based in the US, but might be anywhere else.
Comment Re: Makes sense (Score 1) 28
Why do you think you should come later, buy the same, and have more meaning as a customer than them?
The first customer is getting several hours (or more) of comfy coffee shop sitting time, and the second customer is getting none. So why do you think that by coming earlier and buying the same, you have more meaning as a customer than them?
He might be angry that his "I will never come to an office again" outlook is being jeopardized, and he can't do anything about it. I'm not even being a smartass. This is going to be a problem for some people.
Comment Re:We Hate Customers! (Score 1) 28
A coffee shop is not a rented office space that can be acquired for the cost of one latte. I can see where they're coming from.
But...But... The coffee shops owe these people. Having the best of the best programmers at your shop is nothing to sneeze at.
Comment missed opportunities (Score 3, Insightful) 33
SHOULD have pointed them to 'suzyqable' adapters on ebay, and mrchromebox.
Then we'd have kids posting fun videos of chromebooks playing videogames instead.
But alas, we get to point out that 'maybe its not smart to skimp on short isolation circuitry on the usb port', and that 'maybe regulations are needed' instead.
The message nobody in power wants to have!
What a shame.
Comment How I would solve this (Score 2) 25
1) Secretly buy a game studio
2) Have them make an Air Traffic Controller simulator, where the user can choose from several airports
3) Hope nobody has read Ender's Game
4) Profit!
2) Have them make an Air Traffic Controller simulator, where the user can choose from several airports
3) Hope nobody has read Ender's Game
4) Profit!
Submission + - DOGE software engineer's computer infected by info-stealing malware (arstechnica.com)
gkelley writes: Login credentials belonging to an employee at both the Cybersecurity and Infrastructure Security Agency and the Department of Government Efficiency have appeared in multiple public leaks from info-stealer malware, a strong indication that devices belonging to him have been hacked in recent years.
Kyle Schutt is a 30-something-year-old software engineer who, according to Dropsite News, gained access in February to a “core financial management system” belonging to the Federal Emergency Management Agency. As an employee of DOGE, Schutt accessed FEMA’s proprietary software for managing both disaster and non-disaster funding grants. Under his role at CISA, he likely is privy to sensitive information regarding the security of civilian federal government networks and critical infrastructure throughout the US.
Kyle Schutt is a 30-something-year-old software engineer who, according to Dropsite News, gained access in February to a “core financial management system” belonging to the Federal Emergency Management Agency. As an employee of DOGE, Schutt accessed FEMA’s proprietary software for managing both disaster and non-disaster funding grants. Under his role at CISA, he likely is privy to sensitive information regarding the security of civilian federal government networks and critical infrastructure throughout the US.
Submission + - DOGE software engineer's computer infected by info-stealing malware (arstechnica.com)
gkelley writes: Login credentials belonging to an employee at both the Cybersecurity and Infrastructure Security Agency and the Department of Government Efficiency have appeared in multiple public leaks from info-stealer malware, a strong indication that devices belonging to him have been hacked in recent years.
Kyle Schutt is a 30-something-year-old software engineer who, according to Dropsite News, gained access in February to a “core financial management system” belonging to the Federal Emergency Management Agency. As an employee of DOGE, Schutt accessed FEMA’s proprietary software for managing both disaster and non-disaster funding grants. Under his role at CISA, he likely is privy to sensitive information regarding the security of civilian federal government networks and critical infrastructure throughout the US.
Kyle Schutt is a 30-something-year-old software engineer who, according to Dropsite News, gained access in February to a “core financial management system” belonging to the Federal Emergency Management Agency. As an employee of DOGE, Schutt accessed FEMA’s proprietary software for managing both disaster and non-disaster funding grants. Under his role at CISA, he likely is privy to sensitive information regarding the security of civilian federal government networks and critical infrastructure throughout the US.
Comment Re:First World Problems (Score 4, Informative) 25
Is anybody going to mention the $25B set aside under Biden to upgrade ATC? and then cancelled by trump?
THIS is not a first or third world problem, this is s trump problem
Comment Re:That's the definition (Score 1) 12
Actually, I think the legal term might be "latches", but I haven't seen the original promise.
Still, if anyone feels like suing Whoop, they've probably got a good case.
Comment Re:Didn't they remove restrictions on AI chip expo (Score 1) 48
Maybe this is part of the "replace".
Comment Re:Politicians fail at understanding technology (Score 1) 48
IIUC, the plan is that if the chip doesn't receive the appropriate acknowledgement from a remote server it will "refuse to boot". Like many games.
It's doable, but I can't imagine anyone buying it. (Well, I've got a weak imagination in some directions.) This seems pretty much a guarantee to put the vendor out of the business. Even within the US one probably couldn't rely on getting acknowledgement to proceed. (And it would require not only the chip to be designed, but also the motherboard would need to have a special trace running to each of the "protected" chips. And you couldn't run it off-line.
Comment Re:Public/Private key solves all this (Score 1) 17
PKI infrastructure is more than just DNS.
For example, your addressbook SHOULD have an entry for "public key[s]" (or encryption key or whatever). Then you can import the public key from a text message, or from a twitter message, or a QR code on a business card, or any other way that you get people's email address and phone number.
OSX has some infrastructure for handling encryption keys that is worth mentioning because it is featureful, but it's not quite user friendly yet, and hasn't been integrated into the personal contacts system.
For example, your addressbook SHOULD have an entry for "public key[s]" (or encryption key or whatever). Then you can import the public key from a text message, or from a twitter message, or a QR code on a business card, or any other way that you get people's email address and phone number.
OSX has some infrastructure for handling encryption keys that is worth mentioning because it is featureful, but it's not quite user friendly yet, and hasn't been integrated into the personal contacts system.
Comment Re:What could go wrong? (Score 1) 48
The estimate I've heard is "about 5 years and a tiny bit more". Of course, this depends on exactly which chips you are talking about. IIUC, Holland is still the only source for the best photolithography machines. That may slow them down a bit...unless the US pisses off the EU.
Slashdot Top Deals
"Gotcha, you snot-necked weenies!" -- Post Bros. Comics
