FORKFORK DOG retweeted
9/ where it breaks down: proofs are blind where they abstract. I planted a bug in 4 lines the spec summarizes away — writes routed to the wrong order book. every proof stayed green. unit tests caught it instead.
so no, you don't get to delete your tests.
FORKFORK DOG retweeted
真实的知识空间是一个高维稀疏矩阵。
这是我学习时遵从的法则之一。
任何一个知识点,如果只在二维或三维的经典坐标系里定位,都会丢失绝大部分信息。
比如光速不变,不仅仅属于物理学,它同时延伸至哲学、因果律的上限,工程学、GPS校准,信息论、传输极限,甚至美学与认知论。
概念的每一个特征方向都是一个维度。真实的知识空间拥有近乎无限的维度,任何两个看似无关的维度之间,都可能存在隐藏的正交或非线性相关性。
虽然潜藏的维度近乎无限,但矩阵中的有效非零元(这里我定义其为实际被观测、证明和连接的知识)极度稀疏。
人类的认知框架和实验手段只能探测到这个巨大张量中极小的一部分坐标点。
在绝大多数坐标位置上,充斥着未知的未知,即零元素,甚至很多看似连续的学科体系,本质上也是建立在几个极其孤立的突破性发现上的插值。
既然空间是高维且稀疏的,我,以及人工智能又是如何理解它的?
在这个过程中,我并不需要硬记整个高维矩阵,我会在稀疏的非零元之间寻找低维流形。
对我来说,做到洞察或理解基本原理,其标准就是找到一个最佳的压缩基底,把极其稀疏的高维矩阵投影到几个核心的特征向量上。比如,牛顿将繁复的自然现象归结为几条基本守恒定律,或者把复杂的系统行为压缩为几组微分方程。这都是绝佳案例。
我为何而学习?
就本质而言,原因很简单,就是在高维空间里建立稀疏向量之间的有效连接而已。
而所谓的创新,就是人类需要在一个极其稀疏的矩阵中,寻找两个跨越极远维度、此前从未被连接过的非零坐标,并在它们之间建起一条连线。
在十年前,我很容易把知识空间想象成一个挂在那里等待人类去填补的超巨型 Excel 表格,但真实情况其实更加诡异。知识矩阵的许多坐标点,在被观测或被关联之前,处于未定义的叠加态。
因为很多知识知是高维要素碰撞时的涌现。
例如,算法博弈论或量子信息论并不是原本就写在矩阵某个固定坐标上的点,它们在计算机科学、博弈论与量子物理、信息论这两个远距离高维向量被拉通的瞬间,瞬间塌缩出的一整片新的非零子空间。
所以经过一些时间的学习时,我思考后发现矩阵表达高度依赖于我选择的Basis。在欧几里得坐标系下极度稀疏的数据,换到傅里叶频域或波粒二象性基底上,可能立刻变得集中且连续。
历史上,大部分科学思想革命,如相对论、量子力学、深度学习的表征学习,就是人类找到了一组全新的正交基,将原本混乱稀疏的高维投影,瞬间坍缩成极简的低维流形。
关于上面这个事情,我就想到了2020年疫情时期我思考的一个事:跨学科的极远正交。
在前沿领域,两个分支学科的特征向量可能完全正交,点积为零。
这意味着一个顶级物理学家和一个顶级经济学家,在彼此的坐标系里看对方,看到的都是零元、噪音或不存在。
这就是为什么我发现跨学科交流极其痛苦。
这是因为人类大脑的注意力机制很难跨越极高维度的稀疏鸿沟。
在稀疏数据中,插值,在已知知识点之间连线是安全的,但一旦外,当人类推向未知深空探索时,是极度危险的。
高维空间的诅咒在于稍微偏离已知样本流形一小步,体积就会呈指数级爆炸,人类落入的绝大多数地方都是绝对的空白与虚无。
FORKFORK DOG retweeted
If it may be relevant to you, $1,000,000 in Smart Contract Security audit grant funding is live now, sponsored by Guardian.
Apply for up to a full audit cost grant here: audit.guardianaudits.com/gua…
FORKFORK DOG retweeted
Replying to @asymmetric_re
thanks, we have been loving it
regarding issues or feature requests
cc: @dirtymic26 @forkfork
our fuzzing chads 🙏
FORKFORK DOG retweeted
crucible by @asymmetric_re is good work
we are already a power user 😄
github.com/asymmetric-resear…
FORKFORK DOG retweeted
The reality of the situation:
SRs are not going to be replaced by AI, even by Mythos.
Intractable to get even the best models to get exactly the right impact, find everything, interpret the protocol intention exactly right (what if the docs are just off) and do so with 100% consistency so there are 0 mistakes every time at scale.
Realistically LLMs are far from deterministic, and security is not verifiable, so this alone means it will always make sense to have a subject matter security expert at the very least using and verifying agentic systems, even in the limit of LLMs.
When there are vast sums of money on the line, it doesn’t make sense to cut the human from the loop.
FORKFORK DOG retweeted
if fable 5 agrees to do your work instead of downgrading you to opus, you’re in the permanent underclass
FORKFORK DOG retweeted
mind boggling that this rainbow of availability is at the foundation layer of modern software engineering
FORKFORK DOG retweeted
🚨 The Guardian x @LimitBreak AMM Defender Contest is now live!
45 days to hunt for $150,000 in rewards across Critical, High, and Medium severity findings.
Details on how to participate available below 👇
FORKFORK DOG retweeted
As a Solana Engineer,
Please slap yourself if you cannot clearly explain at least 10 of the following:
Accounts model vs account-based chains
Program Derived Addresses (PDAs)
Seeds & bump mechanics
Account ownership rules
Rent exemption mechanics
Lamports vs SOL accounting
System Program responsibilities
Upgradeable program loader internals
Instruction serialization (Borsh vs Anchor)
Transaction message structure
Versioned transactions (v0)
Address Lookup Tables (ALTs)
Compute budget & compute units
Priority fees & fee markets
Transaction simulation vs execution
Preflight checks
Atomic transaction guarantees
Cross Program Invocation (CPI) mechanics
Reentrancy in Solana (why it differs from EVM)
Signer authorization model
Writable vs readonly accounts
Account locking & parallel execution
Sealevel runtime scheduling
Optimistic concurrency model
Deterministic execution constraints
Transaction size limits (1232 bytes problem)
Account data layout design
Zero-copy accounts
Account realloc patterns
Serialization cost vs compute tradeoffs
Anchor discriminator logic
Anchor account validation pipeline
Constraint macros under the hood
IDL generation & client binding
Anchor events vs logs
Program logs & debugging strategy
LiteSVM vs local validator testing
Test validator architecture
Sysvars (Clock, Rent, Instructions, SlotHashes)
Slot vs block vs epoch
Proof of History (PoH) mechanics
Leader schedule & block production
Turbine block propagation
Gulf Stream transaction forwarding
QUIC networking in Solana
Replay stage & fork choice
Commitment levels (processed/confirmed/finalized)
RPC consistency guarantees
State compression basics
Merkle tree accounts
Concurrent Merkle Trees
Compressed NFTs architecture
Token Program vs Token-2022 extensions
Associated Token Accounts derivation
Mint authority vs freeze authority
Delegation mechanics
Metaplex metadata accounts
Candy Machine architecture
On-chain randomness limitations
Oracle design patterns (Pyth/Switchboard)
Price update verification
Upgradeable program security risks
Program authority management
Multisig program patterns
Escrow design patterns
Time-locked accounts
Vesting contract mechanics
Replay attack prevention
Signature verification flow
Ed25519 program usage
Secp256k1 verification program
Compute exhaustion attacks
Account inflation attacks
Rent draining vectors
CPI depth limits
Stack frame limits
Heap allocation constraints
Deterministic randomness strategies
Idempotent instruction design
State machine modeling on-chain
Instruction replay safety
Optimistic UI vs finalized state
Indexer architecture (Helius/Triton style)
Websocket subscriptions vs polling
Event indexing from logs
RPC rate limiting strategies
Transaction confirmation UX design
Wallet adapter lifecycle
Phantom signing flow internals
Partial signing transactions
Durable nonce accounts
Offline transaction signing
MEV on Solana
Jito bundles & block engines
Priority fee auctions
Transaction landing optimization
Failed transaction debugging workflow
Compute profiling techniques
Program size optimization
BPF execution model
LLVM → SBF compilation pipeline
Stack vs heap limits in SBF
Program deployment lifecycle
Upgradeable buffer accounts
Devnet vs mainnet behavioral differences
Forks & rollback handling
State migration strategies
Backward-compatible account upgrades
FORKFORK DOG retweeted
~7,000+ nSLOC, $150k reward pool
Come find stuff and get paid
🚨 A new $150,000 Guardian Defender contest is coming soon.
Hunt for bugs in @limitbreak’s AMM — Critical, High, and Medium severities will all be eligible for payouts.
Full details and how to participate below 👇
takopi v0.5.0 is out 🐙
our third release in a day!
- new agents: opencode, pi
first external contribution by @snfernandez who has added @opencode runner
$ uv tool install -U takopi
github.com/banteg/takopi
now supports codex, claude code, opencode, and pi
FORKFORK DOG retweeted
By 2028:
* 50% of current auditors will be fully replaced by AI
* 30% will be training AIs to find bugs they used to find manually
* 20% will still be doing manual audits
The future is increased automation & Human-AI Symbiosis, embrace it or get left behind 👨🔧🤖🚀
FORKFORK DOG retweeted
A cool insight into the review/auditing process w/ the folks at @GuardianAudits. This example was not from the Baseline code, but we got similar depth of review.
One maybe undervalued part of their audits is that their teams write complete invariant test suites with state of the art tools (echidna + medusa atm) which you can bring into your repo & integrate with your development process.
Not only that, but studying their invariant tests can improve your skills as a smart contract/protocol dev and teach you new perspectives on how to test + what needs to be tested. By running their invariants in our development suites, we have found bugs that we never would have found otherwise.
I have no doubt in my mind @0xOwenThurm, @giraffe0x, and the rest of the team at Guardian will play a key role in the ultimate success of DeFi and crypto. They are pioneering a standard in testing/auditing that we will one day look back and take for granted as the "obvious" way to do things.
My utmost respect to these devs, and auditors/security researchers in general, for helping secure our space and giving us the tools, knowledge, and confidence to pioneer innovative protocols. Thank you.


















