Certified means an exact certificate and a second checker. Everything else is labelled. If a certified claim cannot be rebuilt from the repo, that is a bug.
We also showed that if part of a secret prime leaks, the first bits are worth more than the last bits.
If the two primes were generated sloppily so p−1 and q−1 share a large factor, there is a proven method that beats the 2021 record. Ordinary keys are not generated that way.
This is not the algorithm that factors RSA in the real world. That one is much faster.
It’s the fastest factoring method that comes with a proof. Under the rules that method uses: N^(1/3) cells, N^(1/5) with today’s packing, N^(1/6) even with perfect packing.
N^(1/6) needs a pattern in the start positions that isn’t there. 2048-bit N^(1/6) is still ~2^341. Not an exploit.
The estimate’s effort setting then reads 417 / 642 / 900 instead of 406 / 624 / 874.
A standard simulator, run to the end, gives the same three numbers. What’s new is the proof that nothing obeying the rule does better, and a certificate a second program re-checks.
Not a claim the new standard is safer or weaker.
The old estimate used a straight line. The line starts where no allowed shape can — 6% too short.
You shall not pass. Unless you change the rules. Then you shall.
It is more complicated than that however: most providers brainfried their models with RL to go prompt -> outcome. Anthropic models can maintain competing hypothesis and tension for far longer, which allows thesis, investigation, correction, and autonomy.
We did Redis last July: igent.ai/insights/production…. The truth is, specification is the bottleneck (and a well known system shortcuts that).
Verifiable engineering tasks, mostly at the scale of entire features, primarily in compiled languages like go or rust, where it can validate the outputs and hill climb. One such public example: igent.ai/insights/production… with repo of results and all commit/session history.
It natively wants you to be absolutely right (it is Sonnet after all). However, unlike 4.0, it will listen to style and tone instructions on a long horizon basis.
Verifiable engineering tasks, mostly at the scale of entire features, primarily in compiled languages like go or rust, where it can validate the outputs and hill climb. One such public example: igent.ai/insights/production… with repo of results and all commit/session history.
The world is rapidly bifurcating between those who have experienced the superhuman capabilities of even current generation models, used properly, and those stuck in the capabilities of the past.
Yes, task length is in Replit’s case measured in agent runtime, which can slightly proxy for the metric that really matters: how many man months (or years) of work the agent performs. igent.ai/insights/production… is several man years in the course of ~70 hours or agent time however.