markjp

@MikeIsaac Interesting since, from what I can tell, Brockman is a great systems engineer but also not an experienced AI researcher.

a cutting Brockman testimony regarding Musk: “Look, he knows rockets. He knows electric cars. He did not and does not know AI. And Ilya and I did not believe he would spend the time to get good at it.”

good morning from a rain-soaked downtown oakland, where i will again be attending the musk vs openai trial no liveblog today, just my wonderful tweets lunch is a normal banana, a mutant orange, black coffee, and some pocket sausages, for travelers

@beffjezos Is this a real position? I don't see how banning data centers gets you to healthcare funding since the sources of investment are different.

This is why they want to ban datacenters to fund "healthcare", it's a threat to the current system, even though AI would save lives and make healthcare more affordable.

@pmarca Crap. Was this written by AI?

It’s becoming clearer how we’re going to tell that something wasn’t written by AI.

@KHollowell_ It's almost like California kids would actually rather not go to Georgia (etc.)!

Recruiting is changing in California. USC, UCLA and Cal all hold top 20 recruiting classes for the 2027 cycle. A lot of local talent staying home these days.

@j0hnwang Sure, but how are their canons?

Anthropic is now worth more than JP Morgan Let that sink in

@kenklippenstein He's dying. Just don't watch. We will all (most likely) go through this process at some point.

It’s so annoying they we have to listen to this condescending bullshit because he’s dying. The Make A Wish commentator

This is why you don't do this kind of over the top marketing. I remain skeptical that Mythos is substantially better than what OpenAI has and other labs will have in short order. It does advance Anthropic's defensive regulation agenda.

@Pons_ETH Thank you. I had the same thought about the conflict inherent in that post.

I usually agree with Ben But as HE has stated many times…..the market is down because of the 4 year cycle. I’d rather have bitcoin at 75k forever than have Gary Gensler back in the SEC

@cremieuxrecueil @streiffredstate This is interesting. I agree that these groups were always considered legally white, but we’re not always socially accepted as the right kind of white. The parallel with black Americans experience is usually overstated in that sense.

@streiffredstate no cremieux.xyz/p/european-imm…

Supreme Court today: 'Racism is banned. No more of that.' 'Now, let's discuss the real issue: Are Italians White?'

@sporadica @yosoymario91 I'd like to VISIT Mars, but not sure about living there.

@yosoymario91 i don't either, but i would love to be wrong. i want to go to Mars

trying to genuinely understand this bc this feels...impossible in his lifetime? $7.5T is crazy, but okay, doable. but 1 MILLION people on Mars?? Yea I believe we will get there eventually, but Elon is 54yrs old Either he thinks he's gonna live to 200 or we'll do this in ~30yrs?

@Sean_Davi More likely to be an attempt at negotiation leverage in a less favorable environment for LeBron imo.

Is this Bron lowkey signaling he might be done depending on how this season ends?

@0xcoinn I've found using the together works best, but while I prefer Claude, Codex is consistently better at writing code in my experience. It tends to write complete solutions for issues while Claude seems to skip things or introduce new bugs.

is it true codex is better than claude? i dont really want to switch everything over

@dieworkwear Fair. Avocado toast probably didn't matter either.

@markjp in my area, a home in a walkable neighborhood is about $3M. one of these watches is $3k; another is like $12k. prob doesn't make a difference in whether you can buy a home.

Some things I saw recently that I think are cool: — Rolex 6611B with black gilt dial at Loupe This — Unpolished Rolex 1016 with gilt dial at Wind Vintage — New Cartier Privé Tank Normale with seven-row bracelet just debuted — Charming JLC Futurematic at Analog Shift

@bridgemindai Which effort settings did you use (medium, high or max) for the re-test?

CLAUDE OPUS 4.6 IS NERFED. BridgeBench just proved it. Last week Claude Opus 4.6 ranked #2 on the Hallucination benchmark with an accuracy of 83.3%. Today Claude Opus 4.6 was retested and it fell to #10 on the leaderboard with an accuracy of only 68.3%. A 98% increase in hallucination. bridgebench.ai just confirmed that Claude Opus 4.6 has reduced reasoning levels and is nerfed.

@garrytan Interesting write-up from Grok on this: x.com/i/grok/share/9…

Shut down the open air drug markets and jail the drug dealers in downtown San Francisco Audit and investigate the nonprofit industrial complex and defund the ones engaging in fraud or actively helping people do drugs until they die Fund recovery and compelled treatment

Anthropic has incentive to make these types of proclamations because they boost hype ahead of their IPO and increase the likelihood of governmental regulatory overreach which will help cement their leading position (along with other major labs). As you point out, any claims made by these companies need to be met with skepticism and evaluation.

🤯 Mythos is NOT as good as Anthropic portrayed. - Non-standard benchmarks - Training on solution sets - Security concerns are real (but they spent $20k+ to push models to find exploits) Don’t believe all the hype. It’s probably better than Opus, but not as good as it’s being portrayed.

@Tim_NBA So, JJ and staff are NOT "stepping up" then?

I am NOT enjoying logging this OKC game 😭 1 quarter left of the worst film I've seen in a long time

@bcherny @garrytan Giving this to corporations and others before releasing to the public feels like a signpost to where this is going for ordinary users. Welcome to the beginning of intelligence rationing and increasingly disparate outcomes between insiders and everyone else.

Mythos is very powerful, and should feel terrifying. I am proud of our approach to responsibly preview it with cyber defenders, rather than generally releasing it into the wild. Model card here: www-cdn.anthropic.com/53566bf5440a10…

@QuilibriumOne @cass_on_mars @QuilibriumInc Idk. I think you need to have a little more ambition. This looks a bit timid. 😂

Hey Q fam! Here is a summary of the latest live stream with @cass_on_mars (@QuilibriumInc founder). ✨ Quilibrium State of the Union: Node Updates, New Services, HyperSnap, MetaVM and the Full Tech Tree Roadmap A comprehensive State of the Union covering business wins, node releases 2.1.0.18 through 2.1.0.22, new managed services, the Farcaster fork, and the ambitious road ahead for Quilibrium's decentralized web services platform. ▶️ Watch it on YouTube (English captions): youtu.be/WXHQRTXsSb8 ▶️ Watch it on X: x.com/QuilibriumInc/… ✅ Key topics: ⦿ Business development wins in key management, wallets, gaming and web hosting ⦿ Node updates from 2.1.0.18 through 2.1.0.22 (in progress), including major bandwidth reduction ⦿ Klearu: MPC machine learning framework released ⦿ MegaRPC: a privacy-preserving RPC service using ORAM-based queries ⦿ QNS updates, bug fixes and auction abuse enforcement ⦿ Quorum Mobile performance overhaul and App Store progress ⦿ HyperSnap fork of Farcaster's Snapchain and its decentralized governance ⦿ MetaVM: the first full hardware 64-bit RISC-V VM in zero-knowledge ⦿ Full tech tree roadmap walkthrough and AWS-parallel service scope ✅ Business Updates Several business development wins have been secured recently. Without naming partners yet, the themes revolve around multiparty key management, crypto wallets, digital asset management (including gaming), the Farcaster fork, and web hosting. More details on partners will be shared at a later date. The two biggest obstacles remain the Apple App Store approval process for Quorum Mobile and completing the shard-out process to unlock all network primitives and enable fee-based transactions. ✅ Node Updates (2.1.0.18 through 2.1.0.22) Since the last stream, there have been four node updates with a fifth on the way: ⦿ 2.1.0.18: Introduced alt-fee-basis app shards (enabling service launches while the QUIL token shard-out is still in progress), proxy pub-sub improvements for nodes behind residential ISPs with limited port forwarding, seniority merge calculation updates, Ferret direct mode for faster buffer access without the RPC layer, and DKLS23 support for more efficient multiparty threshold ECDSA signing. ⦿ 2.1.0.19: Bootstrapping fixes for the DHT (distributed hash table) peer discovery layer. ⦿ 2.1.0.20: Significantly faster Docker image builds (reduced from 1-2 hours down to ~15 minutes, thanks to community contributor Hamoud), improved proposer logic for automated shard selection, and support for workers to leave less advantageous shards. ⦿ 2.1.0.21: Official release of QClient with all protocol messages finalized, immediate rejection support for shard proposals and leaves (no longer needing to wait the full 360 frames), and additional bug fixes. ⦿ 2.1.0.22 (in progress): Bug fixes for prover management in QClient, eviction-related bugs around coverage halt timeframes, seniority merger incorporation fix, and most importantly a massive bandwidth reduction. Global-level traffic was cut by approximately 90% by moving frame data and message data from the PubSub layer to direct RPC calls. The same optimization is being applied to worker-level traffic and is expected to be the last item before the release ships. ✅ Network Status at Time of Stream At the time of the stream, the network had approximately 408 addressable public nodes. Shard coverage breakdown: ⦿ 256 shards in halt risk (fewer than 3 active provers) ⦿ 536 shards needing more coverage (fewer than 6 active provers, does not cause a halt) ⦿ 2,306 shards considered healthy A future update will add libp2p relay functionality, allowing nodes behind residential ISPs to run without a public IP by routing through relay nodes. This is technically possible today but not yet officially supported to avoid the connectivity issues experienced in the early 2.0 days. ✅ Klearu: MPC Machine Learning Klearu was released as Quilibrium's MPC (multiparty computation) machine learning framework, runtime and inference library. ✅ MegaRPC: Privacy-Preserving RPC MegaRPC is a new managed RPC service coming to QConsole. Unlike centralized RPC providers such as Alchemy and Infura, which link users' IP addresses to wallet addresses, MegaRPC uses ORAM-based (Oblivious RAM) queries. This means the service cannot know what a user queried or associate their identity with their wallet. Even if IP addresses were logged along the network path, the cryptographic query format prevents any metadata leakage. MegaRPC is already running live and powering Quorum Mobile. Conversations with wallet teams are ongoing about adoption, though some have declined because they rely on the analytics that traditional RPCs provide. QConsole access to MegaRPC is expected to be introduced this week. This is described as the first RPC service that "can't be evil." ✅ Alternative Network Transports and Internet Censorship The stream touched on the topic of internet censorship in countries like Iran, where traffic is being actively blackholed. Thanks to libp2p's pluggable transports, Quilibrium can support unconventional connectivity methods such as LoRa (long-range radio) and Bluetooth. A broker node using Quilibrium's onion routing could allow users with long-range radios to connect through the mesh network, enabling communication without creating a digital trail. While radio waves can be triangulated, the sprawl of a mesh network makes surveillance far more difficult. This was tested previously using libp2p's WebSocket adapter, which worked with no changes to the node software. ✅ QNS (Quilibrium Name Service) Updates QNS has received several bug fixes, particularly around auctions and offers. Some users attempted to exploit the state machine logic around orders, and the team has put them on notice: those transactions will be reprocessed and reversed. QNS has also switched to the new privacy-preserving MegaRPC on the backend. Additionally, the QNS interface received a significant UI redesign by community contributor Lamat, who also redesigned parts of QConsole. ✅ Quorum Mobile Quorum Mobile is currently going through App Store approval, primarily held up by Apple's review process. Each update required for Apple also resets the Android release timeline to keep both platforms in sync. Key improvements include: ⦿ Space Explorer: significantly faster performance for discovering and joining public spaces ⦿ Native-side cryptography: the React Native JavaScript bridge for encryption/decryption has been moved entirely to native code, eliminating the worst performance bottleneck ⦿ Memoization and rendering: UI components no longer re-render on every new message, with memoization applied throughout ⦿ Public profile support: users can optionally create a public profile with a profile picture and exposed wallet addresses for receiving payments ⦿ Proper push notification support in the upcoming App Store release Quilibrium's philosophy remains firmly opt-in: no user data is assumed or exposed by default. Users must manually opt in to any public-facing feature. ✅ HyperSnap: The Farcaster Fork HyperSnap is Quilibrium's fork of Farcaster's Snapchain protocol. The backstory: around January, Farcaster's original development company (Merkle Manufactory) announced it was shutting down and selling all of its resources to Neynar (a Farcaster API and tooling provider). The Q team had an outstanding pull request, open since late November or early December, to join the validator set and help decentralize the network. Despite initial agreement to add them, the decision was reversed within 12 hours, and the team was told to fork instead. Since then, HyperSnap has had multiple releases, with a new one imminent that makes running a HyperSnap node as simple as running a Snapchain node, complete with a dedicated Docker organization. Importantly, HyperSnap is not just a Q project: only one person from Q, Inc. is in the HyperSnap organization, with the rest being other Farcaster community members. This structure is intentional, as Farcaster's success requires multiple stewards rather than a single corporate benefactor. HyperSnap runs as an alt-fee basis shard, enabling proofs to be incorporated onto the Q network. This allows QStorage-hosted files to be used as embeds on Farcaster. Quorum integrates Farcaster feeds through the protocol and also supports direct messages via the Neynar Farcaster API. ✅ MetaVM: Full Hardware Zero-Knowledge VM MetaVM is the first full hardware 64-bit RISC-V virtual machine that operates in zero-knowledge. While other projects like Jolt offer 64-bit RISC-V VMs in ZK (using MPC lookup tables), MetaVM differentiates by enabling proof of execution at the hardware level across the entire OS stack. This means proving that from initial boot all the way to a chosen stop point, every computation step was performed correctly, whether running Linux, a web server, or even Doom. The total proof size for the full OS and hardware execution is approximately 1.5 kilobytes or less. MetaVM also supports the Ethereum Virtual Machine and Solana's eBPF extensions, with proof outputs compatible with both Quilibrium and Ethereum. Proofs can be verified on Ethereum today using existing BLS12-381 point evaluation and pairing functions in a smart contract. For Quilibrium's managed services, MetaVM enables AWS-compatible offerings such as Lambda-equivalent functions, EC2-style compute, and ECS/EKS container orchestration. ✅ Tech Tree Roadmap The stream featured a walkthrough of Quilibrium's tech tree, a visual map of the entire ecosystem. The center nodes represent what has already been developed and released: Quorum, the node software/protocol, shared crypto intrinsics, Klearu, MetaVM, the EVM shard, and HyperSnap. The outer nodes show what is still in development. For Quorum specifically, features still in development include video calls, voice calls, and group wallets. The scope explicitly mirrors AWS service categories: machine learning, networking and content delivery, compute, application integration, storage, and more. Already released services include: ⦿ QNZM (IAM equivalent) ⦿ QKMS (Key Management Service) ⦿ QNS (Route 53 equivalent) ⦿ QQ (SQS-compatible queue service) ⦿ QPing (SNS equivalent) ⦿ QStorage (S3 equivalent) QNS can already resolve to QStorage content on the backend, with frontend support still needed. When resolving to public QStorage buckets, Quorum's browser uses onion routing so that no traffic passes through clearnet and no one can identify what was requested or its size. The overall mission: provide a fully decentralized, secure web services platform, an everything app for private communication (comparable to WeChat but private), and empower developers of all skill levels to build powerful privacy-first tools. ✅ Mainnet Status and Shard Coverage Mainnet is currently running. The token shards are in a coverage halt phase, which blocks fee-based transactions, but alt fee-based transactions are fully operational, which is why all the other Q services are already live. Once the coverage halt ends, the network will have a warning period before any future halt would occur, giving node operators time to restore coverage to affected shards. ✅ QNS and QStorage Integration QNS already supports resolving names to QStorage content on the backend, but the frontend UI currently only exposes resolution to Quorum public keys. Full linking between QNS and QStorage on the frontend is planned. Once available, browsing .Q names in Quorum will use onion routing for complete privacy, ensuring no clearnet exposure and no linkability. $QUIL