8 Photos and videos
localhost retweeted
Jun 15
.@DavidSacks on the new Claude models:
"They're now retaining for 30 days every prompt and every output you send: no exceptions, even enterprise customers with zero-retention agreements. It's all your context, your memories, your files. They keep it to build a profile on you, classify you, and decide what capabilities you unlock. They're creating a new level of AI haves and have-nots."
This is the world Zcash was built for.
4
9
49
7,600
When intelligence becomes a service, access becomes power.
A handful of governments and companies hold the keys.
The question is whether innovation remains open to everyone else.
Jun 13
The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees.
The net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.
Access to all other Claude models is not affected.
We apologize for this disruption to our customers. We believe this is a misunderstanding and are working to restore access as soon as possible.
Read our full statement: anthropic.com/news/fable-myt…
67
The first human behavior AI successfully learned wasn't reasoning.
It was blame delegation.
[ERROR] TaskExecutionFailedException
Message: "Subagent failed to execute the task as instructed."
Action Taken: Responsibility transferred to subagent.
Main Agent Confidence: 100%.
Status: Operating normally.
86
Every prompt loop is a negotiation between ambiguity and compute. We ask the model to generate endless possibilities, celebrate when one matches our intuition and quietly skip defining verification criteria. Hallucinations are treated as edge cases, correctness as a suggestion. The answer might be wrong but the token bill is always correct.
15
17
207
474,214
Take one sentence, "The cat sat on the ___", and watch it move through the whole machine: tokens, attention, the next word, and the loop that writes everything else.
Plain English, no math.
326
Everyone worries about Bitcoin vs quantum, but Zcash’s Sprout pool is the real risk. If 2016’s trusted setup was compromised, unlimited hidden ZEC could have been minted. With withdrawals still allowed, we won’t know until it’s fully emptied.
1
2
33
31,670
What happens between pressing Enter and ChatGPT's first token?
Before the model generates anything, the system may already have:
• Loaded conversation history
• Retrieved user memory
• Queried external knowledge
• Selected a model
• Scheduled GPU resources
• Executed tools
The interesting part of ChatGPT isn't just the model—it's the distributed system around it.
I documented the end-to-end system design covering : prefill vs decode, KV-cache and PagedAttention, continuous batching, speculative decoding, quantization, three-tier context memory, RoPE for long context, RAG, agent loops, SSE and resumable streaming, model routing, GPU scheduling, RLHF and DPO alignment, three-layer safety, and production tradeoffs.
4
1
32
138,426
112
Every interview, conversation or debate has two insecurities in the room.
One fears not being enough.
The other fears losing the comfort of being the one who knows.
So one performs.
The other defends.
And somewhere between those performances, truth disappears.
The best discussions begin when both sides stop protecting an identity and start seeing clearly.
6
25
277
533,968
Put together 145 topics across 9 sections (Java, Python, Go) to make these things click faster.
Free, no signup: crackingwalnuts.com/concurre…
95
How would you design a GitHub-scale platform?
Interesting problems:
• A fork of a 5GB repo may cost MB, not GB
• Push correctness depends on refs consistency
• PR diffs use merge-base, not branch tips
• Review comments surviving force-pushes
• Code search using trigram indexes
• CI runners becoming a major cost center
• Hot repos creating unique scaling pressure
crackingwalnuts.com/post/git…
105
Key ideas:
- Duplicate execution is expected; fencing idempotency make it safe
- Locks fail (GC pauses, expiry); fencing prevents stale writes
- Effectively-once = one committed result
- Scheduler can over-dispatch; workers enforce correctness
Full write-up:
crackingwalnuts.com/post/job…
45
Designing LeetCode style code execution isolation for 50M submissions/day
Online judges run untrusted code from anonymous users.
A sandbox escape = access to test cases, other users' code, or the host.
The isolation ladder:
-> ProcessBuilder: zero isolation
-> Docker seccomp: shared kernel
-> gVisor: user-space kernel, ~95% compatibility
-> Firecracker: separate guest kernel, hardware KVM, near full compatibility
Firecracker (what AWS Lambda uses) gives each submission its own VM.
A kernel exploit inside the VM is a dead end. You control a tiny 256MB machine with no network.
Cold boot too slow at 2K/sec?
Snapshot the VM after runtime initialization, restore in ~25ms.
Full design:
crackingwalnuts.com/post/lee…
54
localhost retweeted
Apr 9
Quantum-Safe Bitcoin Transactions Without Softforks
github.com/avihu28/Quantum-S…
240
652
3,348
1,888,253