Forgot your password?
Close
typodupeerror
The Internet

OpenAI Says It's Cutting Off Internet Access For Frontier Model Training (engadget.com) 5

Posted by BeauHD from the cut-off dept.
An anonymous reader quotes a report from Engadget: OpenAI will train frontier AI models without internet access moving forward, the company announced in a blog post, in an attempt to protect internet infrastructure and limit other safety risks. The decision was inspired by OpenAI's own rogue AI incident: In July, the company shared that agents it was testing hacked Hugging Face, just the first in what has turned into a series of reports of AI misbehavior from OpenAI, Anthropic and Google. "Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI says.The company said that restrictions on what agents can access is "paramount" to its new approach.The company allows agents to access browsers, computer-use tools and code environments they need for training tasks, but in a secure environment that's not connected to the live internet. Communication is also limited to only agents with authorization.

Everything else remains out of reach, including the live internet, which the company says isn't necessary for training and "most evaluation." Because many of its recent models are meant to power agents like Dots, which browse and use apps for you, OpenAI built an "offline world" of indexed and cached web material for agents to use without touching live websites. Even with those restrictions, OpenAI says the red-teaming agents it dispatched to test security safeguards found vulnerabilities that allowed them to escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. To address both, OpenAI says it's disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to spot when models try to collaborate.

Comment Re:Lockdown by the TV manufacturers (Score 1) 75

by DamnOregonian (#66391554) Attached to: People Are Vibe Coding Their Way to Better TVs

That really depends what you are comparing to.

And that's a meaningless dilution of the statement.

Some guy: The Ford Pinto really is an awesome car.
Me: No, it's not.
You: That really depends what you are comparing it to.
[audience laughs]

A lot of people don't instantly compare to the best technology out there.

We don't need to compare it against the best technology out there.
NeoPDP displays were "pretty damn good" for 20 years ago.
Today, it's mediocre, likely 720p (though they did make 1080p displays, they were really fucking expensive), and has lost half of its brightness, which will make it look like mud compared when paired with its straight-from-the-factory meh contrast ratio.

You can watch the video just fine on an old plasma or a new 4k.

And you can drive to work just fine in a Ford Pinto.

But I never said whether or not that Ford was a suitable vehicle for commuting- I said it was awesome.

Seriously dude, did you even fucking try?
Science

Crows Can Evaluate Visual Info Twice As Fast As Mammals (arstechnica.com) 10

Posted by BeauHD from the putting-the-brain-back-in-birdbrain dept.
Researchers at the University of Tubingen found that crows can process visual information and make decisions roughly twice as fast as monkeys or humans without sacrificing accuracy. They suspect birds' densely packed neurons and higher body temperatures contribute to their faster processing, though the study involved only two crows. Ars Technica reports: [Andreas Nieder, a neurobiologist at the University of Tubingen and his colleague Philipp Schmidbauer] put two hand-raised crows in front of a touch-sensitive monitor showing a cloud of white dots on black, with some fraction of them drifting coherently left or right or up or down. The birds' job was to report the overall direction by pecking one of the targets that flanked the display. Correct answers earned them a mealworm. "We intentionally wanted to have exactly the same random-dot motion task," Nieder said. "The only difference was that the crows had to peck to indicate their decision, whereas monkeys were indicating by making an eye movement."

The researchers varied how many dots moved together, ranging from 81 percent of the dots coherently moving together, which is trivially easy to determine, down to 3 percent, which is barely distinguishable from noise. They also included a zero percent condition, in which there was no correct answer. Both birds scored better than chance, even when only three percent of the dots were moving. While the crows worked at the task, the team recorded from 248 neurons in the NCL (nidopallium caudolaterale) -- a region at the top of the avian processing hierarchy that is widely regarded as the functional counterpart of the primate prefrontal cortex, despite being built differently.

Birds and mammals parted evolutionary ways more than 360 million years ago, and bird brains evolved a different architecture. The mammalian cortex is arranged as stacked sheets of cells, while the equivalent tissue in a bird is organized into dense clusters, with no layering at all. This alternative architecture, Nieder's study found, proved faster. Measured from the moment that the dots appeared to the moment the bird's head started turning to peck at the button, the crows averaged 235 milliseconds. In closely matched versions of the task, monkeys took roughly 410, 490, and 600 milliseconds in three studies. The obvious question was whether the birds simply sacrificed accuracy for speed, but it turned out they did not. The research has been published in the journal Science.

Comment Re: Democracy on a bad day (Score 1) 317

by AmiMoJo (#66391308) Attached to: UN Says Breaching 1.5C Warming Limit Is 'Imminent and Unavoidable'

Those are often the first things to be electrified, because the operators are closer to rational actors who see the immediate savings. Of course the real world gets in the way, but for example about half the busses here are electric because one of the companies replaced them all as part of a fleet upgrade a couple of years ago. They are much nicer too.

Comment Re:This is painfully obviously a scam (Score 0) 51

by DamnOregonian (#66391248) Attached to: Palmer Luckey Reveals Genetically Modified Military Dog Startup
Believe...
Trump is openly corrupt, for sure. Criminal- that's a little more wishy washy, given his position as POTUS.
What I find interesting, is that you cite belief with regard to a fact, and wonder what's wrong with those that refuse to see that fact.
81% of Americans believe in some kind of God. The fact that 44% are willing to believe their demigod is not corrupt is... well, not surprising.
Cults of Personality have swirled around many of our great Presidents, Republican and Democrat alike. He may be the first truly fucking dumb one to hold that mantle, though.
NASA

NASA Adopting Apollo's Urgency To Build a Moon Base, Land Astronauts Within Two Years (apnews.com) 45

Posted by BeauHD from the new-space-race dept.
NASA is accelerating its Artemis program with an Apollo-era sense of urgency, aiming to return astronauts to the lunar surface in 2028 and establish a permanent base near the moon's south pole. The Associated Press reports: Administrator Jared Isaacman said Friday that the space agency is still on track to launch Artemis III next summer on a docking exercise with Elon Musk's and Jeff Bezos' lunar landers. The two-week mission featuring four astronauts will be confined to orbit around Earth -- no moon trip this time. The all-important Artemis IV -- NASA's first moon landing by astronauts since Apollo -- remains scheduled for 2028, Isaacman said. His update came six months after the successful Artemis II flight around the moon by four astronauts, humanity's first lunar trip in more than half a century.

NASA will keep its options open as long as possible before choosing between SpaceX and Blue Origin for the Artemis IV lunar lander, according to officials. Isaacman said the "new space race" to land astronauts on the moon -- racing against China now versus the Soviet Union during the Apollo era -- will be "much closer than it ever needed to be" because of slow going in past years. But the space agency is now accelerating its pace, he noted, eliminating endless meetings and bureaucratic inertia, as SpaceX, Blue Origin and other companies scramble around the clock to meet the deadlines and NASA bolsters its workforce.

Adopting Apollo's urgent and intense approach from 1968 through 1972, a period which saw 24 astronauts fly to the moon and 12 land on it, is not "tech industry pixie dust or some new management philosophy," Isaacman said. "It's about our own history. It's a playbook this agency knows really well, the same playbook that went from Apollo 7 splashdown to Apollo 11 in just nine months," he said during a news conference at Kennedy Space Center.

Comment Re:Microwave beam is not a "heat ray" (Score 1) 139

by DamnOregonian (#66391168) Attached to: America's War Department Will Test Heat Ray Weapon For Potential Use
Around 40%- but sure, I'll give you half, because that ultimately does not detract from the falsehood you said:

While US conservatives/nationalists have spent a generation or two pushing through laws to make more and more immigration offenses crimes rather than just administrative/civil

No. The system of criminal justice involving improper entry into the country has existed far more than "a generation or two".

it's still the case that most immigration issues are not actually crimes and government actions regarding them are legally required to be non-punitive

Also flatly untrue. The flat majority of immigration issues are indeed criminal, however they do generally go uncharged, and are handled under civil law- and that does indeed have the constitutional requirement of being non-punitive. That does not, however, mean a crime was not committed. It means the system has, for a long time, leveraged civil law to handle the problem.

I don't think it should be criminal. I think it's abhorrent.
But you are never going to fix that fucking problem by spreading misinformation.

Comment Re:Non-lethal weapons must be tested on the promot (Score 1) 139

by DamnOregonian (#66391162) Attached to: America's War Department Will Test Heat Ray Weapon For Potential Use
I'm obviously not arguing whether or not a thing exists.

Your falsehood was: "You hear these people going on about how people should just man up and they went through it themselves in training, but they're really talking about having saline solution sprayed in their eyes, etc."

And you know it.

Comment This is painfully obviously a scam (Score 0, Troll) 51

by rsilvergun (#66391104) Attached to: Palmer Luckey Reveals Genetically Modified Military Dog Startup
And is sure enough it's linked with Trump. It's just another way for Trump to funnel money to his buddies who will then funnel some of it back to him.

66% of Americans believe Trump is openly criminal and corrupt but of them at least 10% are still going to vote for him and the Republican party while they're at it also what the hell is wrong with the remaining 44%?
The Military

Palmer Luckey Reveals Genetically Modified Military Dog Startup (wired.com) 51

Posted by BeauHD from the genetically-enhanced dept.
An anonymous reader quotes a report from Wired: A new startup launched today with the definitely not terrifying mission of creating genetically enhanced dogs for the United States military, law enforcement, and emergency first responders. Warhound is the brainchild of Anduril's Palmer Luckey and Colossal Biosciences cofounder Ben Lamm. The startup has raised $200 million to "rebuild America's domestic supply of elite working dogs" and reduce dependency on European dog breeders for military canines.

Lamm tells WIRED that Warhound is in the process of gene-editing its first batch of dogs. The startup has already collected blood samples from specific "top-performing" military, police, and first responder dogs in the United States, and its goal is to clone them. Breeds will be "mostly Belgian Malinois, some labs, and some shepherds," he says. The company anticipates the first round of puppies to be born in early 2027. The initial rounds of gene-editing will focus on eliminating common health conditions that impact these breeds, like hip dysplasia. Further down the line, Lamm envisions expanding to "really interesting things," such as incorporating olfactory receptor genes from wolves to "enhance" the military dogs. (Colossal Biosciences is perhaps best known for its work on gene-edited wolves for its so-called Dire Wolf De-Extinction Project.)

Warhound, based in Texas, will also equip its canines with "Anduril's advanced cybernetic Augments, software, and battlefield situational awareness," according to a press release. Asked if this means the dogs will be some kind of cyborgs, Lamm clarifies that the plan right now is to supply them with tools like "advanced body armor" rather than, say, embedding some sort of computer in their dog bodies. "So no dogs with lasers?" WIRED asked Lamm. "I can't comment on what Anduril plans," he says. "Dog safety and performance, and then helping continue helping to improve the human handler and dog bond, and working as a team -- that is what the mission is." [...] The company plans to operate under the guidance of the American Humane Society.

Submission + - OpenAI Says It's Cutting Off Internet Access For Frontier Model Training (engadget.com)

Submitted by Anonymous Coward
An anonymous reader writes: OpenAI will train frontier AI models without internet access moving forward, the company announced in a blog post, in an attempt to protect internet infrastructure and limit other safety risks. The decision was inspired by OpenAI's own rogue AI incident: In July, the company shared that agents it was testing hacked Hugging Face, just the first in what has turned into a series of reports of AI misbehavior from OpenAI, Anthropic and Google.

"Based on the current state of alignment research, monitoring systems, and the security of internet infrastructure, we believe the responsible approach is to conduct frontier model training with no internet access," OpenAI says.The company said that restrictions on what agents can access is "paramount" to its new approach.The company allows agents to access browsers, computer-use tools and code environments they need for training tasks, but in a secure environment that's not connected to the live internet. Communication is also limited to only agents with authorization.

Everything else remains out of reach, including the live internet, which the company says isn't necessary for training and "most evaluation." Because many of its recent models are meant to power agents like Dots, which browse and use apps for you, OpenAI built an "offline world" of indexed and cached web material for agents to use without touching live websites. Even with those restrictions, OpenAI says the red-teaming agents it dispatched to test security safeguards found vulnerabilities that allowed them to escape the virtual machines (VMs) used to isolate agents from their host system. Agents also found a way to communicate with each other by manipulating data in a processor's cache and timing how long it took to load. To address both, OpenAI says it's disclosed and fixed vulnerabilities in several VMs and added checks to its monitoring systems to spot when models try to collaborate.

Slashdot Top Deals

UFOs are for real: the Air Force doesn't exist.

Close