avatar
Jay
24,451 Twitter followers
Follow
Co-Founder @Sei_Labs // Previously at @RobinhoodApp
Posts
avatar
Jay
03-21
Thread
finality is the most abused word in blockchain. our research team just dropped a very relevant series on this. core thesis: finality is not a boolean. it's a spectrum of credible commitments. the only honest question is — under what assumptions, and at what cost will it break? almost nobody waits for Ethereum's formal finality (~18 min). rollups claim 5-second finality but at that point you're trusting a sequencer and a security council of ~12 entities. not the full validator set. no general-purpose rollup has reached Stage 2. not one. things people get wrong: more stake = more security. not if 3 entities control it. $50B across 3 validators is easier to corrupt than $10B across 50 independent operators. instant finality = real finality. a 200ms commitment from 1 validator is instant and weak. 400ms from 40 validators with 2 rounds of voting is slightly slower and categorically stronger. faster = better. a system that hits 150ms by relaxing fault tolerance from 33% to 20% has made finality faster and weaker. this is why we built Sei the way we did, and why we're designing Sei Giga using the same framework. 400ms finality under classical BFT assumptions studied for decades = credibly fast. read the series below.
HNST
2.9%
avatar
Jay
03-18
finality is subjective and really depends on the observer. in this case 13s finality is dependent on full synchrony of ethereum's ~1 million validators. the network has never meaningfully fallen out of synchrony, so in principle this works, BUT there has never been an incentive for this attack vector before. people think that breaking synchrony requires you to partition the network (a catastrophic attack, but one which requires more coordination). in reality you can eclipse one single sequencer / single client using this new fast finality rule. as more value gets secured in this way, incentives for these attack vectors obviously increase. the assumptions here aren't strong enough to be securing millions for any of the major l2s. they are very conditional and come with risks. when you dig deep the assumptions and risks are made clear. that's great. but the top level marketing saying this is real 13 second finality is untrue. we have a research post in production showing the full breadth of what finality can actually mean. you will be shocked at how far the meaning can get stretched. twitter.com/jayendra_jog/statu...
OBSR
5.29%
avatar
Jay
03-06
shipping product with ai isn't as easy as everyone is saying our eng team ships daily with cursor, claude code and codex. after months of integrating ai into every workflow, here's the honest breakdown: what ai is genuinely great at: - scaffolding + boilerplate - refactoring - tests - documentation what ai still can't touch: - performance bottleneck root causes - complex consensus behavior - long-tail state growth & mitigation - validator coordination - production incident debugging the hard part nobody talks about: workflows are evolving weekly. tooling is fragmented. each model has different strengths. what works today is obsolete in two weeks. our playbook right now: keep the team lean, give engineers more scope, increase output per person. claude .md files are becoming essential. code review starts with cursor, escalates to a human. curious to hear about experiences from other founders.
HNST
2.9%
loading indicator
Loading..