Impressive work! For comparison, an H100 can do roughly 100-200 tok/s of Muse 30B for raw inference single-thread, going up to low thousands of tok/s with a large number of threads - and I am sure that for the massively-multi-threaded case they can optimize the prover further. So we roughly, sort of, have single-digit (<10x) overhead for LLM proving! Next step is getting single-digit overheads for FHE, and then ultimately vFHE (aka STARK * FHE). A crazy ambitious milestone given present FHE overheads, but because of how highly structured and almost-linear LLM inference is, it's closer to the realm of possibility than you might think. Single-digit-overhead all the things.

Important notice regarding Neyro Network incident. Our team has reviewed all findings within our scope of engagement. In May 2026, Hacken conducted a smart contract security audit of two contracts on BNB Smart Chain, an initial review and a subsequent remediation review. Nothing beyond those contracts was in scope. Both contracts show no signs of on-chain exploit. The first has never been used since deployment. The second has processed many transactions, all invoking a privileged loop() function restricted to OPERATOR_ROLE — it takes the contract's USDT balance and calls increaseLiquidity() to add liquidity to an existing Uniswap position, with no swaps or arbitrary external calls. To date, no one from the Neyro team has reached out to us regarding this audit or the incident. Full report:

Import-reliant nations in Southeast Asia are rolling out emergency decrees and rushing to subsidize rooftop solar panels as conflict in the Middle East continues to drive up the price of fossil fuels.

Importing ≠ portability The big AI models will now let you export your chat history and import it somewhere else. But there’s two catches. 1. It only works between the big models. The smaller AI tools in your stack never see any of it. 2. It's a file dump, not a sync. Every time your context changes, you're updating it by hand. Context should be portable everywhere you work, instead of being held together by a million little workarounds.

important distinction on "vibecobing" vs engineers using AI: to non-mathematicians, all mathematicians are all the same. they're just elite at math. to the normal person, there are no bounds to even understand the gaps between the #1# ranked mathematician vs #25# vs #1000# only when you are an elite mathematician can you easily spot what level other mathematicians are. #1# can easily place every single one of his peers. #25# can do the same, but mostly only for those that are worse than him. similarly, the normal person using AI to vibecode cannot comprehend how far the gap is between them using AI on ChatGPT/Claude/etc. vs an actual elite engineer using AI. the app that you end up vibecoding can look very similar to a real app after some prompting, sure. but that would be just like looking at two basketballs going into the same hoop and ignoring the form, the shooting accuracy, the distance, the move into the actual shot, etc... (the guy at the very end of the NBA bench is closer to LeBron James than you are to him) maybe one day, this all gets democratized and you can "vibecode" the next Meta or the next Google but we are many orders of magnitude away from that even being comprehensible.