sudosysgen

قبل 21 يومًاdiscuss

Much of this is probably true. However, Mythos is not a hacking focused model, and while Anthropic seems to train their models on CTFs etc... while others like Zhipu seem not to or not nearly as much, that does mean that it's entirely possible that an actor could post-train a strong model like GLM5.2 to be comparable to or maybe even stronger than Mythos in terms of hacking.

sudosysgen

قبل 21 يومًاdiscuss

A lot of bad actors are both technically sophisticated and have more than enough resources to post train their model. Morally I think it's still the right choice, but consequence wise I doubt it's going to make a big difference.

sudosysgen

قبل 21 يومًاdiscuss

> What can Iran do if other countries consider Iranian ship as free for the taking as retaliation for Iran attacking ships in the strait of hormuz?

This is exactly what Trump tried. The US was boarding Iranian ships thousands of kilometers away. In the end the US still backed down and agreed to an Iranian toll in the MoU.

Iranian ships have been getting boarded for extremely flimsy reasons for over ten years now. That forced them to adapt to it in such a way that they are now far more resilient to it than the rest of the world is.

sudosysgen

قبل 22 يومًاdiscuss

The missile launchers that Iran relied on for many strikes are actually very big, and cannot be hidden in small civilian infrastructure. The US was unable to target them when they were coming out of the very well known and publicly located missile cities because there was no US aircraft that could loiter around them and wait for them to come out without being shot down - that's why the US sent drones to that task, which suffered unsustainable attrition.

And the drones/missile have much more range than 300km. The Shahed-136 drone have a 2000+km range, which is significantly more than the combat radius of carrierborne fighters, even if you add reasonable amounts of refuelling.

The problem in the end isn't that it was impossible to strike every possible hiding site without causing massive casualities. It just wasn't possible. The US failed to durably damage Iranian installations. The backup plan was to exploit air supremacy to interdict whatever was coming in and out of those installations - that also failed. This was an operational failure of US military doctrine, that is unrelated to the tolerance of casualties. It is simply that US military planners overestimated their abilities and made assumptions they couldn't cash. It wasn't a case of casualty avoidance or whatever.

It's also questionable whether a ground invasion of Iran would be feasible to begin with. The Operational Art of War has a great series on the logistics of such an operation, it would be extremely difficult and would most likely require the US to send the troops... through the Strait of Hormuz to begin with.

sudosysgen

قبل 22 يومًاdiscuss

Aircraft carriers lose efficacy rapidly as distance increases. The aircraft flying from them require refueling and longer-range standoff munitions, which requires super-linearly more sorties to get the desired. This is even worse when the targets are far from shore, which can be the case when you want to bomb large countries.

The American doctrine might be to kill the enemy before you are in their view. But the enemy gets a vote. This is no longer something the US can rely on, even against clever enough middle powers.

It's just true that the assumptions that underpinned the current American force composition and strategy are breaking down. That's why the war went much worse than most mainstream analysts expected.

sudosysgen

قبل 22 يومًاdiscuss

This is missing a common failure mode, which is information past the knowledge cutoff. If you need info past that time they'll fail no matter how big or small the model is, so the hallucination rate can matter independently of the knowledge base. If all use-cases had a uniform risk of falling out of support, this would be a valid argument, but since it's often the case that a datapoint is guaranteed to fall out of support, the absolute ability to recognize that is crucial.

sudosysgen

قبل شهرينdiscuss

They are very useful when working on remote servers, VMs and containers. Much much more convenient and robust than, say, X forwarding.

sudosysgen

قبل 6 أشهرdiscuss

Social media platforms have been doing that for years to keep advertisers happy, so yes.

sudosysgen

قبل 6 أشهرdiscuss

Every large social network has fairly advanced mass screening setups for advertiser-sensitive topics. They just need to change the configs. On YouTube for example they will transcribe audio and run OCR on text to flag sensitive topics using MLP in order to flag certain topics (ex: Palestine/Israel), and prevent most ads from being shown (and demonetize and down rank).

Basically every large advertiser requires this so it's pretty trivial to turn on.

sudosysgen

قبل 6 أشهرdiscuss

That's only for the US, not the West writ large.

sudosysgen

قبل 7 أشهرdiscuss

As far as I can tell cheapish 2D lidar for mapping and robot navigation were a bit earlier than the XV-11; they were made by Hokuyo in 2006. I remember that their lidar module was made by some other (American?) company that in turn competed with Hokuyo, people would take them out and use for their own projects.

It's ultimately not very complicated - it's a laser rangefinder that you spin on a motor. It's such a simple - and old! - technology which would obviously get significantly cheaper with time, it was definitely the right horse to bet on. I never understood iRobot's vision strategy.

sudosysgen

قبل 7 أشهرdiscuss

> Binocular vision ought to be good enough for a vacuum.

It could be, but it just is not. VSLAM robots were practically significant worse. There are a lot of limitations to multi-ocular vision for a robot vacuum, for example the relatively featureless walls and few features across the horizontal binocular axis.

Neato was never as big as iRobot. They didn't fail from commanding heights, they never were that successful to begin with, for entirely different reasons. If they had managed to get to iRobot's level of ubiquity and distribution they would have had a much better shot of still being around nowadays.

sudosysgen

قبل 7 أشهرdiscuss

There is not that much more to it than lidar and 2D slam as far as the core technology. There are a lot more features yes but they are not nearly as valuable. I agree they are better, but that's for reason of execution and non-enshitiffication, not core tech.

sudosysgen

قبل 7 أشهرdiscuss

There is not much tech to steal here. 2D lidar mapping is something a high schooler could do 10+ years ago, and that was their core tech. The value was in executing earlier and better, and applying existing tech to robovacuums. If they could have sued they likely would, this is a valuable market.

sudosysgen

قبل 7 أشهرdiscuss

iRobot's failure is that they made a bet to use CV instead of Lidar for their mapping robots for a long time until it was too late. That made their affordable, non-mapping robots far far worse than only slightly higher priced lidar robots, while their mapping robots were too expensive for mass appeal and were still worse at navigation than up-market lidar based robots. Ultimately they were simply outcompeted.

sudosysgen

قبل 7 أشهرdiscuss

Hezbollah operates hospitals and medical services. It's not just a political party.

sudosysgen

قبل 7 أشهرdiscuss

If it's rare enough it may well be cost efficient to just fly people into the ship if and when that happens.

sudosysgen

قبل 7 أشهرdiscuss

Yeah, but DDR4 is very very common.

Also, you don't need a lot of performance for these games. Even 3000MT/s DDR5 is fine for competitive shooters especially in CPUs with big caches

sudosysgen

قبل 7 أشهرdiscuss

Why do you need to handle DDR5? You can use DDR3 to play the vast majority of competitive video games. It's not hard to find an FPGA that can handle DDR3 or DDR4.

You also don't need to sniff the entirety of the traffic. You just need to introduce aliasing. That is much harder to do for DDR5 but you don't need it to be reliable or stable for a long time, because you won't be sniffing for very long. And you don't need to do 6000+MT/s either.

sudosysgen

قبل 7 أشهرdiscuss

There are DDR4 interposers you can buy for 50$. The basic thing is that you don't need all of the ram all of the time, you just need to find an address which you can then rewrite to make two valid references to the same physical memory (see: badRAM/battering ram). Then you can use an IOMMU compliant DMA to access that memory.

Or you can use an FPGA to interpose the RAM and intercept the network traffic for a couple hundred bucks.

comments