synarchefriend
synarchefriend on Hacker News
I would expect code written by the latest frontier models with a "find all security issues make no mistakes" pass to be more secure than average human code written in the pre-AI era (not a high bar).
The OP shows that it was, in fact, solved in a jiffy (a few hours work done largely by an LLM) with a magic (generally applicable) exploit. Given that, I suspect the reputable security experts in the thread who claimed they could solve it were telling the truth.
Now look through this thread and be amazed at how most posters are apparently satisfied by this level of evidence and some are even calling for China-style Internet censorship in response. A good reminder that even enlightened progressives are helpless against cognitive bias when they want to believe the worst about their enemies.
You are confusing Scott Alexander with Eliezer Yudkowsky.
Do you even believe this cope in your heart? With each month that passes you're going to keep seeing increasingly advanced bugs and mathematical proofs found by AI. And you'll have to keep coming up with increasingly silly excuses why that isn't real progress. Maybe save yourself the cognitive dissonance and face reality now. It's not just a parrot.
We have open-weight LLMs like DeepSeek that prove the cost of running inference with near-frontier models can be very cheap.
The model is just taking the user's claim that it deleted the D drive at face value. Where is the actual command that would result in deleting the entire D drive?
The command it supposedly ran is not provided and the spaces explanation is obvious nonsense. It is possible the user deleted their own files accidentally or they disappeared for some other reason.
What I think is happening here based on the few details: Facebook regularly suspends accounts for posting pictures of naked kids, particularly in countries which do not have strong taboos against that. These get reported and sit in a government database with little more information than "Maria Santos from the Philippines posted CP." Then any time a Maria Santos from the Philippines arrives, they get flagged. If it's a common name, lots of people get flagged. This incompetence is the only way I can explain 10% of crew members getting flagged without intentional fraud by the government.
You would need to be severely intellectually challenged to think Musk actually intended to give a Nazi salute. The idea is so preposterous I refuse to believe you actually believe that. Or perhaps I am underestimating how detached from reality political fervor can make people.
You must be using a peculiar definition of expert because even generally conservative AI experts like LeCun now expect we could have human-level intelligence within 5 to 10 years.
The model was trained on games in PGN notation. It would be shocking if it found ASCII art easier to understand than what it was actually trained on.
o1 can also solve many arbitrary math problems that it could not have possibly seen in its training data. And it shows the steps that it uses to do so. How do you explain this without reasoning?
"Senior employees leaving due to its powers being so extreme"
This never happened. No one said it happened.
"the model some media outlet reported recently that is so powerful OAI is considering charging $2k/month for"
The Information reported someone at a meeting suggested this for future models, not specifically Strawberry, and that it would probably not actually be that high.
Incredible that so many tech types are now cheering on the government blocking websites.
Humans also need to break up the problem and think step-by-step to solve problems like 234878 * 452.
This happened once before and the driver still wasn't paying attention at as obvious a hazard as a train crossing? Not sure I trust them to distinguish between autopilot and FSD. Although of course FSD is not actually capable of driving without supervision so it could easily be true.
Another description:
https://www.gutenberg.org/files/66685/66685-h/66685-h.htm
Geological conditions existing in the middle of the Strait were, up to that time, almost entirely a matter of surmise, based on observations made on the British and French sides of the Channel, and in the process of finding out more about them, [18]Thomé de Gamond decided to descend in person to the bottom of the Channel to collect geological specimens. In 1855, at the age of forty-eight, he had the hardihood to make a number of such descents, unencumbered by diving equipment, in the middle of the Strait. Naked except for wrappings that he wound about his head to keep in place pads of buttered lint he had plastered over his ears, to protect them from high water pressure, he would plunge to the bottom of the Channel, weighted down by bags of flints and trailing a long safety line attached to his body, and a red distress line attached to his left arm, from a rowboat occupied also by a Channel pilot, a young assistant, and his own daughter, who went along to keep watch over him. On the deepest of these descents, at a point off Folkestone, Thomé de Gamond, having put a spoonful of olive oil into his mouth as a lubricant that would allow him to expel air from his lungs without permitting water at high pressure to force its way in, dived down weighted by four bags of flints weighing a total of 180 pounds. About his waist he wore a belt of ten inflated pig's bladders, which were to pull him rapidly to the surface after he had scooped up his geological specimen from the Channel bed and released his ballast, and, using this system, he actually touched bottom at a depth of between 99 and 108 feet. His ascent from this particular dive was not unremarkable, either; in an account of it, he wrote that just after he had left the bottom of the Channel with a sample of clay
... I was attacked by voracious fish, which seized me by the legs and arms. One of them bit me on the chin, and would at the same time have attacked my throat if it had not been preserved by a thick handkerchief.... I was fortunate enough not to open my mouth, and I reappeared on top of the water after being immersed [19]fifty-two seconds. My men saw one of the monsters which had assailed me, and which did not leave me until I had reached the surface. They were conger eels.
They probably want to emphasize that it's not another llama derivative.
Having an explanation doesn't mean it isn't a weird anomaly.
I believe the SmartCut feature reencodes just the frames needed to reach the closest keyframe.
Providing no access to the underlying data is far more common in other fields of science than AI.