sudosysgen
sudosysgen on Hacker News
DeepSeek R1 and DeepSeekMath were groundbreaking models which were substantially stronger than any models available for distillation at the time they were developed. The Chinese labs have advantages as well despite the lack of compute, too.
Also, this is exactly what one would do if they wanted to resell cheap tokens in places where US AI companies don't do business, whether or not there is distillation as an added bonus revenue stream.
It's true if you pay API prices overall, it's just that the OpenAI/Anthropic subscriptions provide way better value than their respective APIs, and that isn't true with the Kimi sub.
There is no deterrent to a motivated and well resources organization. Post training a model in general, let alone just for offensive cybersecurity, is well within the technical and financial capabilities of many cyber criminals. There's just no way out of that.
Yes, in general Israeli intelligence and hacking groups tend to enjoy far fewer consequences when they screw up than their Western counterparts, which then allow them to be far more aggressive.
How can you simultaneously decide you do not need to care about the lives of a group of people on the basis of a difference of ethnicity as you prioritize the formation of an ethnic state, and simultaneously refuse the premise that the formation of that state is to blame for ethnic strife? Your first and last paragraphs cannot be reconciled.
Of course it is scalable. We've scaled discovery massively without tool-agents, so why can't we scale more with them?
This is not entirely true. There are plenty of people of that opinion but there were plenty of people that thought that the US would be able to achieve air supremacy, and then destroy the missile cities, which would eventually mean Iran would lose its ability to close the strait and attack US bases and the Gulf states in a matter of weeks and months. Those people were wrong, and the US failed to achieve these tactical goals against a conventional military power, which is not something many US strategists were used to.
Of course plenty of people expected it would happen the way it did, but it was not unanimous at all. I have no doubt many in the general staff were rudely surprised.
This is not clear at all. The MMA program, which is the MQ9 replacement, is mostly the same mission set, but has a slightly lower per unit price tag and lower capability - there is a good chance that the MQ9B ends up being the MQ9 replacement.
Also, the US military is still procuring broadly similar drones, and will be for a while.
The lack of survivability is not the fault of the drone design. The design set it has is to have high unrefueled endurance, a medium payload capacity, and a reasonably low cost. The nature of technology means that to do so it is going to be relatively slow and have a large wingspan.
Additionally, they were lost in missions which require loitering in predictable patterns near known targets, in which they were shot down mostly by optically guided missiles. There is no indication that there is an alternative design that would be more survivable. The only hope is that there would be one that is significantly cheaper, but that is not obviously true.
So all in all, these are valuable losses. When they are replaced, they are very likely to be replaced by a substantially similar design at a significant fraction of the cost of acquisition.
I'm really no fan of the US MIC, but the MALE/HALE drones are actually a decent design. While they are more expensive than they should be, they are and have to be pretty big aircraft with sophisticated sensors, and so a decent chunk of the cost really is justified. Chinese/Russian/Iranian equivalents are broadly similar in design and cost 25-50% as much, which is about what you'd expect.
All in all, the attrition is not necessarily unexpected but it is a real problem, will have real and expensive consequences, and ultimately reflects a tactical failure in how this war was prosecuted which is a material factor in the current bad outcome. Many people really did think they would be able to survive enough to loiter around the missile cities as a backup in case they failed to be drstroyed, and that turned out to be wrong.
MDMA is far, far easier to make than a virus from nothing. People can make it in their basement.
I am aware, I personally knew someone who did so a decade ago. This is very different from synthesizing and culturing a virus from a sequence. Also, you wouldn't use CRISPR for this, you'd make a plasmid.
The best defense for these is upstream, so that the plasmid provider refuses to make it for you if it looks like a pathogen or you're not affiliated.
Why would they need help from an LLM to design a sequence when there are sequences for many existing highly pathogenic viruses and well understood ways to make them more harmful?
Besides, biology grad students rarely or never have access to a lab with zero oversight.
The genetic sequence for smallpox is publicly available. The bottleneck in causing maximum harm is that you need a large, expensive, well stocked lab and a lot of expertise. LLMs helping people generate a sequence just isn't the bottleneck to bioterrorism, it's entirely about abilities and equipment to deal with the physical world.
It's still worth thinking about it, of course, but by far the way to mitigate risks are things like wastewater surveillance, tracing purchases of physical supplies, etc...
Of course, an LLM could provide advice that makes it easier for a virologist to do something like this, but there is little indication the effect size there is notable.
The life sciences are very hard. A motivated undergrad requires extensive in-person teaching to become useful in far simpler projects. It's not clear how an LLM can in the foreseeable future bridge that gap when human experts can't teach someone without a lot of time and in person interaction, and that is with access to a staffed, resourced lab.
They didn't go straight, though. Chinese companies have been working for years on ICE. In fact, the largest companies were focused on ICE cars, and many were outcompeted by domestic EV upstarts.
Chinese cars have been sold worldwide for decades, they just weren't competitive with western cars even in very poor African countries.
We do have a pretty good idea of the quality of the current crop of Chinese ICE engines. They aren't great - I can go into more detail if you want.
Sure, none of that is inherent. But do note that China never caught up in internal combustion engines. They had to wait for a fundamental paradigm change to catch up, and were not able to copy ICE technology to a satisfactory level. Similarly, China has been able to make serviceable airframes for decades, it's just until now that they're getting close to price competition - and likely Comac will be more expensive for decades to come.
My point is moreso that they can succeed in many domains, but it's mechanically unlikely they'll succeed on some domains unless we force their hand. Semis were one of them.
There is definitely comparative advantage in certain technologies which depend on a niche supply chain with deep materials science experience, simply due to experience and network economies. China has many such advantages too.
Lithography is very hard to dominate because if you want to do so, you have to convince your ecosystems to switch to domestic suppliers. But in the short term, they don't want to as that will hurt their competitiveness. That's why, for many sectors still and for lithography until the bans, there was no domestic program with realistic prospects to ever truly outcompete.
The cost is not just pure development. There are very high switching costs, and downstream opportunity costs affecting the entire ecosystem. As an example, Nikon was competitive with ASML in DUV for a while, but no one was willing to switch because the cost/risk was way too high. For lithography, that was essentially the issue - no serious foundry was willing to take the massive risk to move away from ASML - and, worse, to skip EUV.
A second reason is that Chinese foundries still had to compete with foreign foundries. The cost of doing lithography with new Chinese DUV machines will be significantly higher than with ASML DUV and then significantly higher than with ASML EUV. So to build an ecosystem that can one day truly compete with ASML, you'd need the entire Chinese semi industry to take significant losses to get to the point where they can compete with ASML and pose a threat in the market.
There is still many sectors like this where China has some kind of parity, but at a total cost of ownership so much higher than western competitors than they are not a serious market threat in the near future. In fact, the US, EU, etc... are in similar positions. Unless the Chinese economy becomes much stronger than the US/EU across the board, one would expect it to be case due to comparative advantage - the cost of directing industry to ignore superior alternatives will slow them down in other fields, unless the Chinese economy is just that much more effective. So without sanctions, it's difficult for the Chinese state to force enough cooperation to solve the issue without spending amounts of money that are a significant drag even to China.
The steelman of the argument is that, due to legacy in technology and comparative advantage, unless they expect a serious supply chain tech, many core technologies will be too expensive to duplicate for China and so there will be areas where they will not advance. There is some amount of evidence that was true for lithography, they did not seem to be making much progress until they were forced to, and there didn't seem to be enough capital available in China to catch up.
Of course, the extent to which they was true and for how long it would be true is up for debate, but it's not a prima facie foregone conclusion
Mythos is Fable without safeguards. Open weight models are already essentially at the Mythos level, and with extra post training a well resourced actor could deploy would almost certainly be significantly better than Mythos at hacking. Whatever their threshold is, it's much higher than that.
As the private/working reserves are set to reach operational minimums in a month or less at the current pace, the SPR will have to increase draw rate or there will have to be further demand destruction. If it goes to 7mn a week, below the 9.9mn peak draw recorded, that's 4.5 months. If it has to exceed peak rates as private/working reserves running out bites more than the optimistic numbers I'm using, or if China returns closer to pre-war oil imports, it will run out in 2-3 months. Before it runs out, the maximum draw rate will likely start falling - there are multiple facilities and some are already at or close to operational minimums - which will force demand destruction even before operational minimums.
The operational minimums for commercial crude inventories are estimated at 300-380mn barrels, with current inventories at . Weekly draw rate this week was 7.2mn just for the commercial inventories, and levels were ~400mn barrels. So there's 20-100mn barrels remaining in the commercial inventory. Even moreso than the SPR, this is uneven, and maximum draw rates will decrease as individual storage reaches minimums.
It changes how many attempts you get until the attack surfaces changes to react to a failed attack, and it does so in a way that is not predictable to the attacker.
> "defenders have to be right 100% of the time, while attackers only have to be right once"
If you have an adaptive system that can react to attacks flexible (say, your own AI agent), then no, that's not correct. It is correct in the classical conception of cybersecurity where the defender is basically static.
No, he refused the DoW demand to use Anthropic models without limits. But Anthropic still agreed to military usage of their models, including for strike planning.
There are plenty of comparable systems in US bases that nevertheless were frequently hit.
Iran has been getting consistent impacts on US bases in Jordan essentially every day for the past few weeks, and they are well defended.
At the end of the day Iran has more missiles than the US has interceptors. So when you get to the point when you're talking about running out, they just don't need to launch big salvos anymore.
Iran has recently demonstrated the ability to execute accurate strikes in Jordan. If it can do so in Jordan, it can do so in Israel once Israeli air defences are similarly degraded. Doubly so because many missiles meant for Israel were intercepted by Jordan.
Iran knows Israel won't stop itself from attacking for too long. So they are going to wait until that point and they are likely going to try the same targeting methodology to paralyze Israeli airbases. If they manage to do so, they will be able to significantly reduce the US/Israeli strike tempo and to be more successful using drones, not even missiles.
It's not, but it's not coincidence either. The reason for the US entanglements in the Persian Gulf intensifying are because it's a choke point for oil, and the reason the US feels freer to mess around is because the US became an oil exporter. So the fact it's hurting China more than the US (in some ways) is related to the factors that led the war to start, even though it's not 27D hyperchess.
Anthropic had Mythos-Preview for many months internally, but from available sources was an active work in progress, and it seems they started releasing it via Project Glasswing to partners before the final checkpoint was available.
Many of these techniques haven't been published very long ago - it often takes a good 6-8 months for techniques to percolate. But also, they come at a complexity cost and, seemingly, also at a stability cost.