Chief Scientist@FLock.io PhD@UniofOxford
Oxford, England
Joined November 2018
- Tweets60
- Following487
- Followers117
- Likes151
Zehua Cheng retweeted
Small models, huge impact. Excited to power @flock_io's high-performance decision model with Qwen & our cloud infrastructure.
🤝 @Alibaba_Qwen models and @alibaba_cloud infrastructure power FLock's specialised AI training.
THIS / THAT Model, our open-source small decision model built on Qwen3.5, outperforms frontier LLMs in speed and accuracy, using our reliable synthetic data methods, peer-reviewed at ACL 2026.
Zehua Cheng retweeted
We post-trained @Alibaba_Qwen's Qwen3.5 2B to build THIS / THAT Model 1.2, our open-source decision model, specialised for decision-making, routing, classification, and scoring.
It scores higher than every hosted model we tested across 1,710 decisions and responds in 31.4 ms. That’s around 32x faster than Claude Opus 5.
Training specialised models for specific business needs will become the norm.
Zehua Cheng retweeted
FLock joins @anam145official as the Main Partner for TRUST404 🇰🇷
From blocked transaction receipts to zero-trust approval logs and tamper-evident agent payments, 17 teams turned security ideas into working systems.
Together we keep pushing for trust, verification and safer Web3 infrastructure.
哈哈哈感谢社区支持
谁能想到Jev 居然输给了一个 1.88B 本地开源模型🔥
结果真的有点让我意外：
我做了一个贪吃蛇决策游戏 Jev 30:28 险胜，但是换成同一组 68 道决策题，本地模型准确率直接做到 94.1%，Jev 只有 76.5%
一个错了 4 道题，一个错了 16 道题
我本来只是想测试一下 “小模型决策能力怎么？”，结果我抽完发现，小模型虽然小 但是准确率一点也不低。
1️⃣ 先说我是怎么测的
双方使用完全一样的 12×12 棋盘、随机种子、碰撞过滤和路径规划器，只有遇到多个方向都能走的时候，才把选择交给模型。
最后 Jev 吃到 30 个食物，跑满 256 步
本地模型吃到 28 个，在 239 步被困住。
如果单看这一局的成绩，肯定是Jev 赢了。
但是贪吃蛇只是某一步选错以后，后面整条路线都会变，所以我又补了固定的 68 道决策题。
结果 this-that-model-1.0 答对 64/68，94.1%；Jev 是 52/68，76.5%。
Brier Score 也分别是 0.042 vs 0.133，本地模型给出的概率和真实结果更贴近
2️⃣ 这个模型最离谱的地方，是它只有 1.88B
this-that-model-1.0 是一个开源的结构化决策模型，MIT License，可以直接在 Mac 本地跑
它和 Jev 的思路其实很像：
不给你写小作文，也不慢慢吐 Token
给它当前状态、一个问题、几个候选答案，它一次前向直接返回：
选哪个 + 每个选项的概率。
所以任务分流、选 Skill、队列选择、标签判断、Agent 下一步动作这种事情，都很适合
我在 M4 上直接用 MPS + FP16 跑，不需要 API Key，也没有网络请求
不同任务中位延迟大概 225～475ms，Snake 实战中位 475ms，平均约 492ms
一个 1.88B 模型，在 Mac 上已经可以做到亚秒级决策。
这个我觉得挺夸张
3️⃣ 我又单独测了 105 道空间决策题
总准确率 84.8%，随机基线只有 34.3%
其中 Snake 状态题、行列文字状态都是 100%，JSON 状态也有 94.4%。
而且输出特别干净。
没有 JSON 少个括号，也不会突然给你解释人生，只会老老实实从候选答案里面选一个，再把概率交给程序
对于 Agent 来说，这种“无聊”反而挺值钱。
4️⃣ 当然它也不是万能的
完整地图测试只有 60.7%，最短路径距离题更是掉到 28.6%
一旦任务需要真正的多步图搜索、连续计算，它就明显不行了
所以这次结果也不能简单理解成：
1.88B 打败 Jev
贪吃蛇单局上是 Jev 赢了，而且 Jev 这次没有保存同条件延迟数据，速度也没法公平比较
真正让我感兴趣的是另一件事：
很多 Agent 里面那些高频、重复、答案明确的小判断，可能根本没必要每次都调用一个大模型，甚至连云端 Jev 都不一定需要。
一个 1.88B、本地运行、一次前向就结束的模型，已经可以完成绝大多数简单的决策了。
一些无关痛痒的简单的决策都可以交给这种小的本地模式、遇到复杂的任务再交给 GPT、Claude这样的大模型可能会更适合。
我现在越来越觉得，未来 Agent 不一定是“所有事情都找最强模型”。
更可能是：
小模型负责做大量的决策，大模型才是真正的为了我们的工作服务。
如果你们感兴趣 也可以部署一下，再偷偷地补充一句 这个可是很牛的牛津团队开源的：
huggingface.co/flock-io/this…
Congratulations to Moderna and Merck on their landmark Phase III results.
This milestone underscores a broader shift in life science: data, models, and therapeutics are becoming deeply interconnected.
We need to understand not only what a biomolecule is, but also how it moves, changes, and transitions between functional states.
@flock_io latest new two papers accepted at #ICDM2026 explore exactly this direction in AI for protein dynamics:
• Multi-timescale protein dynamics — disentangling fast local fluctuations from slow global conformational changes, enabling better use of abundant short MD simulations and scarce long trajectories.
• Generative modeling of protein transition paths — moving beyond stable conformations to model rare, functionally important transitions between molecular states.
We believe the next frontier of AI for biology lies beyond static structure prediction: toward dynamics, mechanisms, and ultimately more powerful foundations for drug discovery and precision medicine.
Zehua Cheng retweeted
Only FLock can do.
BREAKING: FLock is officially a @RedHat ecosystem partner.
This is a major step in expanding enterprise-grade federated learning to the cloud-native infrastructure trusted by governments and Fortune 500 companies.
Zehua Cheng retweeted
FLock is headed to Geneva next week for a series of events convened by the @UN from 6-10 July. 🇨🇭🇺🇳
Our Founder and CEO, @0x7SUN, has been invited to attend the Global Dialogue on Artificial Intelligence Governance, established by the UN General Assembly under the Global Digital Compact. He will also be joined by our Head of EMEA, @TiffanyWang98, at the @AIforGood Global Summit and the WSIS Forum organised by @ITU.
As sovereign and responsible AI move up the global agenda, we're looking forward to sharing lessons from our work on decentralised and privacy-preserving AI, from our sovereign AI collaboration with the Sarawak government to SDG-aligned initiatives with @UNDP.
See you in Geneva.
Zehua Cheng retweeted
Next stop for CEO Forum Blockchain and AI for Good - Young Talent Edition: @UniofOxford | March 28
You will hear from:
🎙️ @0x7SUN, founder & CEO of FLock, and University of Oxford alumnus.
🎙️ @helen_abfinance, founder of @abf_finance and @ChainforGood, former co-CEO of @Bybit_Official
Join us for builder insights, explore what’s pushing the edge of blockchain and AI, and connect with the FLock & BGA teams.
In collaboration with Oxford CSSA.
If you're in town this Saturday, come through👇
Zehua Cheng retweeted
It's official: CEO Forum - Blockchain and AI for Good: Young Talent Edition is coming to @imperialcollege | March 25
Insider talks from founders and professors who are building and working at the intersection of blockchain and AI.
📢 Speakers:
@0x7SUN, founder & CEO of FLock
@helen_abfinance, founder of @abf_finance and @ChainforGood, former co-CEO of @Bybit_Official
Ying-Ying Hsieh, assistant professor of Innovation and Entrepreneurship; Associate Centre Director, Imperial College Centre for Cryptocurrency Research and Engineering, @imperialcollege.
If you're a student, builder, or just deeply curious about where purpose-driven tech is heading, this is your invite 🇬🇧
I saw lots of old friends there. Happy coding and build sth great.
This is humbling and insane and amazing. 🫶🦞luma.com/claw
Zehua Cheng retweeted
🔥
7/ Read our paper "A Control Theoretic Approach to Decentralized AI Economy Stabilization via Dynamic Buyback-and-Burn Mechanisms."
Co-authored by @zehuac, Wei Dai, Zhipeng Wang, @rui_sunn, @yyyzzgc, and @0x7SUN from University of Oxford, University of Manchester, and Newcastle University.
arxiv.org/pdf/2601.09961
Sybil resistance + reputation + freezing/slashing behavior. They explicitly require "sybil-resistant" reputation systems and describe enforcement actions that look like protocol-level sanctions: revoking credentials, freezing staked assets, and flagging transaction history for forensic review. "Sybil-resistant" is also a very crypto-native vocabulary, and the enforcement mechanism mirrors how many protocols think about identity/reputation under adversarial duplication.
That's also what FLock is building.
😀Good to know decent teams working together.
刚读完Google最新的AGI论文，论文的观点大胆且非常crypto native，我一度以为是在看加密项目的白皮书。
几个核心观点：
1. AGI终将是一个 DAO，而不是一个 CEO。
我们总是幻想某天一觉醒来，诞生了 GPT-10 这样的全知全能神。论文指出未来的AGI大概率是分布式的 。就像公司没有一个人能擅长所有事情，AGI 将是一个由无数互补的“特种 Agent”组成的网络 。在这个网络里，没有单一的中心，超级智能是在Agent疯狂的交易和协作中涌现出来的 。
换句话说，AGI不是一个实体，而是一种市场状态。
2.治理AGI不能靠“法律”，得靠“智能合约”。
既然从单一模型变成了市场，安全范式必须从“心理学”转向“治理学” 。以前 AI 安全是做一个超级大脑的对齐。但面对每秒亿万次的高频交互，人类监管是无效的。因此必须引入智能合约。Agent 完成任务后，由Oracle 验证结果，自动执行结算 。代码即法律，不满足安全约束，资金就无法流转。
3. 引入“质押”与“罚没”。
怎么防止 Agent 作恶？研究团队居然复刻了PoS机制。Agent想接大单？先Stake资产 。一旦被审计发现作恶，智能合约直接罚没质押金 。这种基于经济抵押的信任，比单纯的代码审查更有效。
4. 链上身份与“Gas 费”调节
DID：为了防女巫攻击，每个 Agent 必须有基于公钥密码学的唯一身份，且绑定法律实体。
动态Gas：为了防垃圾数据污染，建议对Agent的操作征收动态费用 。这不就是以太坊拥堵时的 Gas Fee 调节机制吗？
5. 记录上链
所有的决策和交易历史，必须记录在加密安全的、只能追加不能修改的账本上 。这为了确保当系统出问题时，能进行取证分析，没人能赖账 。
这篇论文标志着 AI 安全的一个范式转移：从单纯的计算机科学和价值观对齐，变成了经济学和博弈论 。
而加密圈子折腾了十几年的DID、智能合约、预言机和经济模型，治理机制等等等等。无论成熟与不成熟，这些探索至少迈出去一些步，而这些步，也许刚好是给未来那个庞大的、去中心化的硅基生命群体的基础。
未来的 AGI 安全专家，可能不需要是一个能写代码的工程师，但必须是一个懂博弈论、市场设计和去中心化治理的“AI 经济学家”。我们本质是要设计在这个新物种的共识、激励和治理模型，而不仅仅是修补大模型的神经元。
论文信息量很大，以上只cover了一部分，有兴趣可以阅读原文
arxiv.org/abs/2512.16856
