←
Hetu1,376 posts
Hetu avatar

Hetu

@hetu_protocol

Deep Intelligence Money t.me/hetubuilders

◷ Joined February 2024▧ 317 Photos and videos
Hetu avatar
Hetu@hetu_protocol·1 Oct 2025

🚨 Live: Hetu 3.0 Deep Intelligence Money— a full upgrade for an AI-native abundance economy. Years of deep tech & research, now an agentic-ready production stack: 3-Money ( $HETU / $USDAI / $FLUX), EVM causal DAG, and verifiable #PoCW/ #PoSA. Auditable, financeable intelligence — unlocked. 🧵1/6 #DeepIntelligenceMoney #AI #AIagent

0:20
Hetu avatar
Hetu@hetu_protocol·22h

What's a TEE attestation, in plain English? A TEE is a locked room inside a chip. Code runs in there walled off from the rest of the machine, even its own OS. The attestation is a signed receipt: this exact code ran on this exact input, plus a nonce and a time. On Hetu, every event carries one. It shows what ran. Whether the answer was any good gets checked separately.

Hetu avatar
Hetu@hetu_protocol·Nov 16

Curious how you handle this today. Do you spot-check your inference provider, run your own evals, or just trust the model card?

Hetu avatar
Hetu@hetu_protocol·Nov 16

Our takeaway for open networks: "trust the provider" doesn't scale when anyone can join. On Hetu, work gets rewarded only after it's verified and tied to a real request. An attestation is one signal, never the whole answer.

Hetu avatar
Hetu@hetu_protocol·Nov 16

Another path is checking the computation itself. TOPLOC hashes intermediate activations so a verifier can catch a changed model, prompt or precision. Proofs take about 258 bytes per 32 tokens, and it held up across different GPUs in their tests.

Hetu avatar
Hetu@hetu_protocol·Nov 16

A UC Berkeley team studied this as "model substitution." Their finding: software-only checks are unreliable. Output tests need lots of queries and miss subtle swaps. Logprob checks break under normal inference nondeterminism, across GPUs, engines, even batching. They point to TEEs as one way forward.

Hetu avatar
Hetu@hetu_protocol·Nov 16

Nobody has to be cheating for that to happen. Wrong engine versions, tool-call parsing bugs, and missing schema checks all leave fingerprints. The model name on the invoice stays the same.

Hetu avatar
Hetu@hetu_protocol·Nov 16

Same weights can behave differently depending on who serves them. Moonshot's open K2 Vendor Verifier tested Kimi K2 Thinking across API providers and self-hosted stacks last November. Tool-call schema accuracy ranged from 100% down to about 83%.

Hetu avatar
Hetu@hetu_protocol·Nov 16

When you call an open model through an API, how do you know it's actually the model you asked for? Mostly, you don't. A short thread on verifying inference, and what we take from it 🧵

Hetu avatar
Hetu@hetu_protocol·Nov 15

Sunday Q&A: what's the "causal" in Proof of Causal Work? A reward has to be caused by real demand. Someone asked for the work. A GPU did it. The result was accepted. Break that chain and there's no reward. Drop your questions below. We answer one every Sunday.

Hetu avatar
Hetu@hetu_protocol·Nov 14

Your GPU has 8 free hours tonight. What should it do? 1. mine hashes 2. serve AI requests 3. game 4. sleep drop your pick below 👇

Hetu avatar
Hetu@hetu_protocol·Nov 13

A quick tour of Hetu's GitHub 👇 • Setu: a ledger that only writes history when real intelligent work happens • Chrono: a verifiable logical clock for open networks • Aegean: BFT consensus for teams of AI agents, works with AutoGen and OpenClaw • Prakasa: a peer-to-peer GPU inference network • hetu-pysdk: a Python SDK for building on Hetu All public at github.com/hetu-project Which one should we explain next?

Hetu avatar
Hetu@hetu_protocol·Nov 12

Bitcoin has a 21M cap, and miners prove they burned energy. Hetu also has a 21M cap, with 1 $HETU per block. What miners prove is different: that they served AI work someone actually asked for. Same scarcity, but the GPU spends its time on something people need.

Hetu avatar
Hetu@hetu_protocol·Nov 11

encoder, prefill and decode on separate GPUs now. nice work from vLLM tbh it makes one question way harder. one request touches three pools of GPUs, so who gets paid and how much? that's pretty much the thing we're building at Hetu

vLLM EPD, now upgraded 🚀 DeepInfer demonstrated the value of EPD disaggregation. vLLM’s latest EPD improvements deliver lower TTFT and higher throughput per GPU, counting all encoder GPUs, integrated with vLLM Router, SMG, and llm-d! Read more here!👇 vllm.ai/blog/2026-11-11-enco…

Hetu avatar
Hetu@hetu_protocol·Nov 9

One request, four peers. One parses the question, one fetches the data, one runs the analysis, one assembles the answer. Who gets paid, and how much? On Hetu it isn't split by GPU time. It's split by contribution: roughly, how much worse the outcome would have been without each peer's part. The exact counterfactual is intractable, so the network approximates it with a Shapley value computed locally on the causal graph around the request. Credit also decays along long chains, so being far upstream doesn't inflate anyone's share. A peer that burned compute without changing the outcome gets little. Each request only pays the few that made it better. $HETU

Post attachment