@raw_avocado Ouch! Thank goodness (or more accurately Peter Wuille) that we have BIP32 now…
English
Adam P. Goucher
3.3K posts
In 2010 this guy lost 8,999 Bitcoin because messed up his backup.
After making a 1btc test transaction the rest 8,999 were sent to a change address.
He then loaded the backup righ before this TX was made.
English
The only thing that we have to do ourselves is figure out how to efficiently implement fault-tolerant universal computation on top of DNA/RNA — difficult, yes, but much more tractable than option (b).
English
Option (a) is more compelling: there’s already superhuman technology (namely molecular biology 🧬) for efficient self-replication, and it only takes a few hundred kilobases to encode all of the required mechanisms: en.wikipedia.org/wiki/Minimal_g…
English
To construct a self-replicating machine you need to either:
(a) fully avoid using any silicon chips;
(b) have an entire semiconductor fab, and everything needed to build all that equipment, and so on until you reach transitive closure, inside your self-replicating machine.
Elon Musk@elonmusk
Optimus+PV will be the first Von Neumann probe, a machine fully capable of replicating itself using raw materials found in space
English
@annexbaja @0xMard @geohotarchive @grok That’s why I also mentioned SPHINCS+ which is impervious to Shor’s algorithm.
English
Polynomial Time Factoring Algorithm geohot.github.io//blog/jekyll/u…
English
@0xMard @geohotarchive @grok No, they don’t. Elliptic curve cryptography is very common nowadays, and that relies on the difficulty of elliptic-curve discrete log instead. And hash-based digital signatures such as en.wikipedia.org/wiki/SPHINCS%2B only rely on the security of cryptographic hash functions.
English
@geoffreyirving Right, that’s a rather stringent condition! Morgenstern has a generalisation of LPS graphs to prime power fields, which removes the problem of finding primes, but again the proof that they’re Ramanujan graphs is very non-elementary!
English
@apgox The conventional ideal in the literature is strong explicitness: that the element to swap node k and layer l with should be computable in time polylog(n), which would include finding the primes. So you can find primes fast enough with Cramer’s conjecture, but not unconditionally.
English
It's a shame: the 1.41e64 constant is _almost_ an LLM capability measure ("What quality expanders can the machines formalise?"), but alas it appears there's an enormous jump from what's already done (MGG) to Ramanujan graphs w/ deep number theory, with little in between.
Geoffrey Irving@geoffreyirving
As an exercise in learning recent Claude Code + Opus 4.6, I've formalised Seiferas's simplified construction of the Ajtai-Komlós-Szemerédi O(log n)-depth, O(n log n)-size sorting networks in Lean, using Margulis–Gabber–Galil expander graphs.
English
@geoffreyirving Actually I think that this still works, right? The prime that you need to construct an N-vertex LPS graph is O(N^(1/3)) and the work to find a prime in [m, 2m] is O(m^(3/2 + eps)) and you can do it in O(log(m)) depth. So the whole thing should be O(N) work and O(log(N)) depth.
English
@geoffreyirving Do you have an additional constraint that computing the sorting network should be linear in the size of the sorting network? Otherwise you could use Bertrand’s postulate to prove that there exists p in [n, 2n] and trial division to find the first one.
English
@justinskycak @BackPedaling112 @Michael_Druggan It’s faster to just do:
det([[6, -1], [4, 4]]) / 2
= (6 × 4 + 1 × 4) / 2
= 28 / 2
= 14
and it generalises to higher-dimensional simplices (e.g. tetrahedra) in a way that the inscribing trick doesn’t.
English
@BackPedaling112 @Michael_Druggan That method takes much, much longer when no slope is 1. The inscribing trick is fast in all cases when the points are integers, no matter what the slopes are.
English
This might be true for some hypothetical standardized test but it's not true for the SAT math. Every single question on the SAT math is easy and straightforward if you actually understand high school level math. Good students will get an 800 without trying. If any of the questions feel like tricks you need to have a memorized response to that's indicative of a flawed understanding of math deeper down.
Justin Skycak@justinskycak
The thing a lot of people don’t realize about the SAT is that it is a well-trained opponent who has made a career out of studying your game and figuring out what’s most likely to mess you up. Coming up against these types of questions is like entering a fight with a professional boxer who has made a career out of exploiting any weakness their opponent might have. If you have solid fundamentals but are untrained on what to expect from this particular opponent in the ring, you are most likely going to get your ass handed to you. It doesn’t matter if you could eventually figure out how to beat your opponent in a longer fight. It doesn’t matter if you could knock them out with a wind-up punch but they keep moving around and throwing off your balance. It doesn’t matter if you can block a freaking bulldozer but they keep faking you out and sliding punches around your blocks. When you are going up against an opponent who is pulling out every trick in the book, you absolutely have to train on specific situations that might show up in the game. Baseline athleticism is necessary but it is not sufficient to win. You need to know in advance what kind of tricks your opponent might pull, what kind of weird attacks to expect, how to defend against them – and even further, you have to build reflexes that will enable flawless execution come game day. That’s the kind of training experience we built this course to deliver.
English
@morallawwithin Bitcoin uses elliptic curve digital signatures, not RSA, so the factorisation is irrelevant. The private key is just an integer n and the public key is the corresponding point n*G on the secp256k1 elliptic curve. You need to solve elliptic curve discrete log, not factoring.
English
FYI this guy’s private key factors to 5*3203*14104793*4676713139402732471
Quinten | 048.eth@QuintenFrancois
Quantum is a threat to Bitcoin
English
@itsclivetime @srchvrs This could also be Berkson’s paradox: to be hired as a researcher without a PhD you’d need other evidence to compensate, and that other evidence may have greater predictive power than does the PhD.
English
1. only anecdotal unfortunately! also possible that things are skewed at the companies i've worked at. in my ranking of the absolute strongest researchers i've worked with, most do not have PhDs - but perhaps this is actually due to the vast majority of the general candidate pool not having PhDs
2. true
3. fair enough; i think most of the frontier labs only have a mild preference for it rather than requiring it
4. makes sense
5. true
English
"I’m regularly asked if one should leave a Ph.D. to get an actual job, and my decision criteria is fairly simple. If you’re not looking to become a professor and have an offer to do modeling research at a frontier lab (Gemini, Anthropic, OpenAI is my list) then there’s little reason to stick around and finish your Ph.D."
Do not let your career be driven by FOMO and bee-in-the-bonnet biases. This post implicitly assumes that the ONLY important research direction today is LLMs or, more precisely, making already very strong models slightly better at a huge cost.
That may be true for some people and some labs, but it’s far from obvious that this should be the default career advice for everyone.
One additional aspect that’s often overlooked in this type of advice is how uneven career constraints are globally.
For many researchers, especially those on visas, formal credentials like a completed PhD are not just about career signaling: They materially affect mobility, stability, and even the ability to stay in the country.
What may be a low-risk optional decision for some can be a high-risk one for others.
Nathan Lambert@natolambert
My raw thoughts on the job market -- both for those hiring and those searching -- at the cutting edge of AI. interconnects.ai/p/thoughts-on-…
English
@ccasadoo @mattkratter I’m a fan of Marek Palatinus (@ slush); he gave us the first hardware wallet (Trezor) and first mining pool (Slush Pool), as well as developing the Stratum mining protocol.
All of these are open-source, and all done to benefit the Bitcoin ecosystem.
English
@mattkratter Okay, so Jade guy is a shitty human, and Coldcard guy is a shitty human. Are we left with only the Seedsigner guy?
English
@lambda_user46 @nitbean @Duderichy Then the caption should be “share of top-performing adults who were talented youngsters” and not “share of talented youngsters who became top-performing adults”.
The colour legend is very ambiguous though…
English
@nitbean @Duderichy I absolutely understood it as the later. So 1% of the top 5% earners by age 35 were among the top 1% smartest kids at age 12.
English
@ElliotGlazer Yes I agree with that: you could partition 1 as 1/2 + 1/4 + 1/8 + … and assign most of the terms to trivially halting machines.
I think that this yields an equivalent prefix-free binary encoding of Turing machines though? So Chaitin must have had some extra conditions?
English
@apgox Anyway this will not generally yield a normal number as the probability, since e.g. we can guarantee p has >90% of its binary digits 1 by placing most of the probabilistic mass on trivially halting machines and distributing it carefully. So this doesn't meet Chaitin's hypotheses.
English
ok it's a quarter after one and i'm a little drunk but i really need to correct the common misconception that "we can't compute what fraction of Turing machines halt."
Isaac King 🔍@IsaacKing314
Best guess: What fraction of Turing machines halt? ("Other" answer includes "it doesn't converge to within one of those ranges" and "it depends on the implementation".)
English
@thomasforth @danwwang That they outsource the physical fabrication to TSMC (another > $1T company!) doesn’t make them a ‘services company’: they’re still selling GPUs and not services.
Compare this with ARM which just licences their instruction set to other companies.
English
"Welcome to Decline" -- still one of my favourite graphs from last year. In the late 1970s, Britain's economic fate very clearly split in two. How much you blame Thatcher will depend on your politics, but her leadership is useful to understand the timing. tomforth.co.uk/welcometodecli…
English
@thomasforth @danwwang The most valuable company in the world makes GPUs. The richest person in the world makes electric cars and rockets. To focus on (advanced, high-tech) manufacturing is much more forward-looking; the services sector is a sinking ship of diminishing importance.
English
@gabriberton I used this here in my f2reduce library for fast Gaussian elimination:
#L491" target="_blank" rel="nofollow noopener">gitlab.com/hatsya/open-so…
So it chooses a large enough of a power-of-2 divisor to get cacheline/vector alignment but not too large as to run into critical stride issues.
English
@gabriberton …size then your accesses will only use a proper subset (half, or worse if you have a larger power of 2 divisor) of the cache. As such, it will behave like having a smaller cache, and you may get diminished performance.
English
So the reason it doesn't print 3 identical numbers is that
1. tokenizer.vocab_size only shows BPE generated tokens.
2. len(tokenizer) also adds special tokens like EOT and <think>
3. the model is padded for speed
It's weird that it's padded by 267 to a multiple of 128, not 256
Gabriele Berton@gabriberton
Do we all agree that this is weird af?
English