Benjamin Chan

@sarahdingwang @benchmark @lightspeedvp @ycombinator @ScottWu46 @ibab @talbroda And lastly, thank you to all Exa users and friends who've been with us these past 5 years. We will not stop until the world’s information is fully organized for you, and for every intelligence across the world 🫡

Introducing Exa Instant: the first sub-200ms search engine. Faster than Google, it's custom built to power realtime AI products like chat and voice.

@_patrickogrady @Thomas_S_Kim Congrats! it looks awesome :)

@Thomas_S_Kim (And obviously huge thanks to @Vervious for providing feedback during those original 23 attempts 😅)

[SPOILER 🚨]: Yes, yes it can. Embedded VRFs are a tricky business. If you're designing your own, make sure to check out our latest blog (co-written with @LayerZero_Core) first:

@kartik1507 @soubhik_deb @ittaia btw will you be turning this into a paper? It seems neat to make catching-up in simplex more communication-efficient. philosophically it's a tradeoff between simplicity of liveness<->complexity of dealing with lost messages.

@soubhik_deb @ittaia 5. In that sense, PBFT, SBFT, Tendermint, HotStuff, Hydrangea, Hydrangea++ follow the former approach whereas Simplex, Kudzu, Alpenglow follow the latter. A post describing one aspect of how they differ is here: decentralizedthoughts.github.io/2025-12-19-cc-…

there has been insane progress in BFT consensus since the hotstuff / tendermint era while much of CT has been focused on app-layer and revenue narratives. over the break i got a chance to dug into recent protocols + genealogies. will write short summaries over next few weeks, starting with the following. n ≥ 3f+1: SOTA moved from HotStuff to Simplex, making the protocol much simpler and cutting confirmation delay (7δ -> 3δ). n ≥ 3f+2p+1: SBFT (dual-mode, fast path only under crash faults) -> Banyan (simultaneous 2-round fast path and 3-round slow path, conf. delay of 2δ under p crash faults) -> Kudzu (fixes unbounded msg complexity from Banyan + leader bandwidth bottlenecks) -> Solana alpenglow (speculative block production for δ block time, production-grade voting + dissemination, byz + crash fault tolerance for fast path under stronger than PS model) -> Hydrangea (fast+slow path under byz+crash faults under PS model with n ≥ 3f+2c+ k + 1 and p=ceil((c+k)/2)) -> Hydrangea++ (speculative block production that cuts proposal delay in Hydrangea to δ) -> Minimmit (dropped slow path entirely, byz-only with n ≥ 5f+1, extremely easy to understand) feels like there is a race going on to approach the speed of light in latency while maxing out on throughput probably missed a lot more of recent developments.

The protocol is dear to me; it's so elegant, fast, easy. I maintain that one day Simplex Consensus will be the byzantine protocol of choice everywhere (industry, in distributed systems classes). simplex.blog

Excited to see Simplex powering practical onchain payments!

Introducing Exa 2.1 We scaled our pre-training and test-time compute by an order of magnitude, unlocking frontier search API performance for both fast and agentic search. Deep dive below:

what a time to be here

@qrazhan_ultd @B_Chou hello prashan! :)

@B_Chou omg worlds colliding, i interned with him at fitbit

Legend. I basically took Alpenglow Votor and just yelled "why cant you be more like Simplex?!" at it until something popped out

Decentralized thoughts on 2 round BFT in Simplex style With @ElodinStorm and Ling Ren decentralizedthoughts.github.io/2025-07-18-two…

Eshan Chattopadhyay is the latest Godel prize winner!

The underlying consensus is based on Simplex Consensus! It's pretty cool to see. simplex.blog

@yupi_asa @_patrickogrady @AndrewLewisPye @Tim_Roughgarden We can also deploy Simplex in an Algorand-like setting, it is essentially a drop-in replacement for their underlying permissioned protocol (eprint.iacr.org/2018/377.pdf)

@_patrickogrady @Vervious @AndrewLewisPye @Tim_Roughgarden Algorand is permissionless by their work. Quasi-permissionless is permissionless. Quasi coz of what's known about the parties and in this setting all nice properties are achieved compared to fully permissionless & dynamically available settings which have impossibility results.

Over the last few months, we built a new type of stack designed from the ground up for those that aren’t getting what they need from the frameworks available today. github.com/commonwarexyz/…

@_patrickogrady @samlafer This is so awesome 🙇‍♂️

@samlafer @Vervious I but only live in the shadow of @Vervious 🙇

Benjamin Chan the GOAT. @Vervious @_patrickogrady would love to host you guys if you want to present this at our Fault Tolerant Frenz distributed systems reading group. Perhaps we can split into a theory part with @Vervious and later a more practical session with @_patrickogrady

Unlike other “off-the-shelf” consensus implementations, consensus::simplex is refreshingly minimal. It delivers simple agreement over opaque “blobs”. Functions that have historically motivated developers to create forks, like block broadcast, are abstracted as interfaces:

Today, we are excited to share our biggest release yet: consensus::simplex (our first of many upcoming consensus dialects). consensus::simplex provides BFT agreement via a new, optimistically responsive construction inspired by Simplex Consensus: eprint.iacr.org/2023/463

@nickthehabsfan how much of the increased offense is due to increased time on ice? He's seems like he's always on the ice (and do the OHL and NCAA even track TOI?)

Many people have asked about Sam Dickinson, and given his scoring this year, I can see why. It's crazy to think that San Jose got Celebrini and Dickinson out of this draft. Add that to the other top picks they'll soon get, and that's a promising core. #TheFutureIsTeal

@mumen_rider99 @nickthehabsfan I second this request

The Winnipeg Jets are 15-1-0 and are 1st in the NHL by SIX points already! A big part of this start is the emergence of Cole Perfetti as a future star in Winnipeg. Perfetti up front makes the Jets' forwards dangerous not only now but also in the future... #GoJetsGo

@yoavgo @BlancheMinerva Not super sure I understand, because we can always run a randomized verifier many times to amplify its success probability; or are you looking for problems where, with some probability over some distribution of instances, there is no “good” verifier at all?

@BlancheMinerva what i am after is (i think) a somewhat different setup, where the verifier does not *use* probability for its advantage, but rather can fail stochastically on some instances.

CS theory puts a lot of weight on the ability to recognize/verify solutions in polynomial time. you can also do quite a bit with adding randomization to the input or the algorithm. but are there works on cases where the verifier is stochastic and non uniform (can make mistakes)?

@yoavgo @discobordism Concretely this most reminds me of verifiers for zero-knowledge for 3-coloring/etc, or as another example the pcp theorem via gap amplification

@discobordism i was thinking more of, say, 3SAT solving where your verifier may make ocassional random mistakes.