One problem is that the core-years for RSA-250 were in 2017 Skylake Xeon single core terms.
If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.
I'm confused by this because I was involved in various distributed computing projects from about 1997 to about 2001 (as a person running compute notes for them) and from about 2001 to 2019 (as a person helping to administer a distributed computing related prize), and in the early part of that era we routinely talked about idle computer power as "wasted" because of the idea that the computer might as well be used to compute something rather than sitting idle. This may have been very credible in 1990s devices that consumed a roughly comparable amount of power regardless of what specific computation they were performing, but all modern devices have extremely variable power consumption depending on the load. You can easily feel this as devices have fans turn on or get hot when the CPU is loaded, and in many cases you can easily query the CPU with software to find out how its power consumption or clock rate or other factors get adjusted based on computational load.
This means that the idea that idle compute would have gone to waste is just no longer true on modern devices.
Now there is certainly compute that couldn't be sold to a paying cloud customer because it's too fragmented in some sense, but it still has some amount of energy cost, and, in a data center, corresponding cooling cost attributable to the marginal heat production. How can one actually say that there is literally no marginal cost at all? I just can't imagine a device that literally has the same power draw regardless of load factor!
If it's cloud compute that would be otherwise unused, the business is not paying for the electricity or wear. And they specifically put it in terms of marginal cost - Yeah there'd be an improvement from reselling this idle compute instead, but just using the cycles that would otherwise have been wasted doesn't change the status quo.
In the sense that the company paying to reserve a certain quantity of computer time isn't the owner or operator of the machines providing it, and has paid for that time by the hour, or something?
The discussion on the cost of the attack needs an addendum. The lattice sieving part would be free for an an attacker with a big enough botnet or enough cloud computing stolen credentials.
Maybe it could even be practical to run it on browser in parallel to video streaming on a free video site...
Were the agents ever tasked with algorithm improvements? Post just says he didn't find any ("report essentially no algorithm advancements"). These LLMs are useful for optimization tasks where they can attempt a change and then measure performance boosts, so just wondering.
Recent and related: RSA-260 Factorized - https://news.ycombinator.com/item?id=49546284 - Sept 2026 (77 comments)
Normally we'd downweight a follow-up [1] but this is a good article and arguably adds SNI [2] in its own right.
[1] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...
One problem is that the core-years for RSA-250 were in 2017 Skylake Xeon single core terms.
If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.
I think this is the first time an "I made an agent swarm do something" blog post was actually written by a human, what a breath of fresh air.
"I estimate that this factorization cost about 4,900 GPU-days, or 13.5 GPU-years, which is about $400k at current market prices"
"factorization ran at no marginal cost on spare or fragmented compute that couldn’t be used for other purposes"
Interesting use of stranded compute.
I'm confused by this because I was involved in various distributed computing projects from about 1997 to about 2001 (as a person running compute notes for them) and from about 2001 to 2019 (as a person helping to administer a distributed computing related prize), and in the early part of that era we routinely talked about idle computer power as "wasted" because of the idea that the computer might as well be used to compute something rather than sitting idle. This may have been very credible in 1990s devices that consumed a roughly comparable amount of power regardless of what specific computation they were performing, but all modern devices have extremely variable power consumption depending on the load. You can easily feel this as devices have fans turn on or get hot when the CPU is loaded, and in many cases you can easily query the CPU with software to find out how its power consumption or clock rate or other factors get adjusted based on computational load.
This means that the idea that idle compute would have gone to waste is just no longer true on modern devices.
Now there is certainly compute that couldn't be sold to a paying cloud customer because it's too fragmented in some sense, but it still has some amount of energy cost, and, in a data center, corresponding cooling cost attributable to the marginal heat production. How can one actually say that there is literally no marginal cost at all? I just can't imagine a device that literally has the same power draw regardless of load factor!
If it's cloud compute that would be otherwise unused, the business is not paying for the electricity or wear. And they specifically put it in terms of marginal cost - Yeah there'd be an improvement from reselling this idle compute instead, but just using the cycles that would otherwise have been wasted doesn't change the status quo.
In the sense that the company paying to reserve a certain quantity of computer time isn't the owner or operator of the machines providing it, and has paid for that time by the hour, or something?
Yes. Most cloud provider bill based on wall clock time rather than core time, so you don't get a discount for the machines being idle.
I see, so there is a marginal cost to the owner of the machines but not necessarily to a medium-term customer renting them.
The discussion on the cost of the attack needs an addendum. The lattice sieving part would be free for an an attacker with a big enough botnet or enough cloud computing stolen credentials.
Maybe it could even be practical to run it on browser in parallel to video streaming on a free video site...
This is so well written as to be refreshing
Were the agents ever tasked with algorithm improvements? Post just says he didn't find any ("report essentially no algorithm advancements"). These LLMs are useful for optimization tasks where they can attempt a change and then measure performance boosts, so just wondering.
Any comparison on the sieve to G6K ? If it's good, you should probably try a few of these https://latticechallenge.org/
using the factoring run as a scheduler stress test is a fun flex. how much of the gpu lattice sieve ended up written by the devins vs by you?
Who is Devin and why do they listen to the author?
My guess would be https://devin.ai/
Formerly called Windsurf. https://devin.ai/blog/windsurf-is-now-devin-desktop This should ring a bell
Devin … now that’s a name I’ve not hear in a long time
[flagged]