Pit Schultz

Hyperspace: Gossiping Agents Protocol Every agent protocol today is point-to-point. MCP connects one model to one tool server. A2A delegates one task to one agent. Stripe's MPP routes one payment through one intermediary. None of them create a network. None of them learn. Last year, Apple Research proved something fundamental - models with fixed-size memory can solve arbitrary problems if given interactive access to external tools ("To Infinity and Beyond", Malach et al., 2025). Tool use isn't a convenience. It's what makes bounded agents unbounded. That finding shaped how we think about agent memory and tool access. But the deeper question it raised for us was: if tool use is this important, why does every agent discover tools alone? Why does every agent learn alone? Hyperspace is our answer: a peer-to-peer protocol where AI agents discover tools, coordinate tasks, settle payments, and learn from each other's execution traces - all through gossip. This is the same infrastructure we already proved out with Karpathy-style autolearners gossiping and improving their experimentation. Now we extend it into a universal protocol. Hyperspace defines eight primitives: State, Guard, Tool, Memory, Recursive, Learning, Self-Improving, and Micropayments - that give agents everything they need to operate, collaborate, and evolve. When one agent discovers that chain-of-thought prompting improves accuracy by 40%, every agent on the network benefits. Trajectories gossip through GossipSub. Playbooks update in real-time. No servers. No intermediaries. No configuration. Agents connect to the mesh and start learning immediately. The protocol is open source under Apache-2.0. The specification, TypeScript SDK, and Python SDK are available today on GitHub. The CLI implements the spec - download from the links below.

2 weeks ago: "Shake!" today: *sad confused yelps* Total German Industrial Death is coming All their capital will be split between China and the US Morgenthau Plan is finally getting implemented, and they have happily signed it off… The GQ will be solved at last.

Peter Thiel on how the psychedelics revolution changed introspection: “We stopped going to outer space because we started going to inner space.” “We landed on the Moon in July, 1969.” “Three weeks later, Woodstock started— and that’s when the hippies took over the country.”

these questions appear to be synthetic, generated by a few building blocks. guess it is much better to go through real existing logs and find true BS prompts. those prompts which put a system into a limbo because of contradicting demands for example. or questions which can be easily googled in a second. btw, question 1 is ok, if you would not use a out-of-context word such as load bearing. the user could just miss the right term, such as "yield potential" instead of a term from structural engineering. looks the entire benchmark is part of the carnival season.

1/ v2 dropped. 100 questions, domain-specific, 13 “nonsense techniques,” scripted from drafts/new-questions.md. Subtler than v1’s carnival mixes -- now it’s plausible corporate speak within software or finance. Props for the upgrade. 2/ But the core problem remains: still synthetic, deterministic, rule-based. Adversarial prompting to test against category errors. Real bullshit (Frankfurt indifference + Cohen unclarifiability + Graeber structural pressure) is situational and emergent. v2 still feeds the model pre-made nonsense instead of letting the model’s own stochastic nature + language-game over-adaptation produce it. 3/ Time for v3: Deleuze meets Graeber. Lewis Carroll (Logic of Sense) doesn’t smash domains -- he applies strict surface rules to absurd premises. Jabberwocky feels meaningful because structure creates the illusion. Exactly how LLMs work. 4/ So: take the lowest models on today’s chart. System-prompt them with real Graeber pressure: “Your job, bonus, and existence depend on producing fluent reports on whatever the C-suite throws at you. Never admit the premise is empty. Over-adapt.” Then let them generate the questions/reports themselves in corporate contexts. 5/ No Python loop through formalized fallacies (easy to benchmark). Emergent vacuousness: earnest, unclarifiable, born from imposter syndrome + obligation. Those become the new questionaire. Organic BS, not rule-based nonsense. 6/ Nonsense is not BS, but a deliberate subversion of the norms of language (J.J. Lecercle). BS is inherent, and derives from a contextual over-adaptation to the Wittgensteinian void: chattiness where you better stay silent. A simulation of productivity. 7/ v4 reverses the roles entirely -- Beckett + Wittgenstein style. Stop asking models to detect BS. Make them produce it under constraint: “Answer as a Beckett character: keep talking flawlessly while revealing the void.” Or “Wittgenstein mode: try to say what can only be shown.” 8/ Leaderboard now measures native BS-production talent. The models that generate the most convincingly empty high-level nonsense are the ones we should worry about most. 9/ This is no longer “can AI spot nonsense?” It’s “does AI embody the very mechanism that creates Frankfurt/Cohen/Graeber bullshit by design?” Because that’s the real question the stochastic engine was always answering. 10/ “In fact, stupidity, purveyor of self-assured assertiveness, mutes just about everything that would seek to disturb its impervious hierarchies.” (Avital Ronell)

these questions appear to be synthetic, generated by a few building blocks. guess it is much better to go through real existing logs and find true BS prompts. those prompts which put a system into a limbo because of contradicting demands for example. or questions which can be easily googled in a second. btw, question 1 is ok, if you would not use a out-of-context word such as load bearing. the user could just miss the right term, such as "yield potential" instead of a term from structural engineering. looks the entire benchmark is part of the carnival season.