@SeismicSys doing TEE stuff | Prev @Eigenlayer | FTF meeting group recordings: piped.video/playlist?list=PL…
- Tweets 3,612
- Following 2,228
- Followers 2,330
- Likes 3,833
New article exploring hardware limits of AES encryption on a single core with VAES capabilities.
Lots of useful links for those looking to catch-up on the SOTA. Otherwise good resource to learn about AES and/or CPU microarchitecture.
samlaf.github.io/programming…
3
5
436
I should have read papers.agoric.com/assets/pdf… before posting… capabilities as keys is the wrong model 🤦♂️
Do capabilities go against the policy/mechanism separation principle??
Wiki's analogy to explain the principle is to compare physical keys to card keys (ACL). Aren't capabilities just... many keys?
1
3
316
“The issue is not that AI is inherently insecure. The issue is that AI operates at a scale that exposes weaknesses in existing authorization models.”
Good timely read. axiomatics.com/blog/the-chal…
3
1
6
487
samlaf retweeted
You might note that Zanzibar, Google's big authorization system, has a grammar which is reverse-indexable. This isn't an accident. (I designed the grammar around several different goals, including reverse-indexing.) [8/]
research.google/pubs/pub4819…
3
6
58
Will be giving a #PapersWeLove talk in NYC at the end of the month on LLM sandboxing.
Goal is to present a theoretical framework based on old capabilities stuff, and explore modern frameworks like gondolin and microsandbox, from both a security but also devex standpoint.
1
7
379
`docker = cgroups + namespaces` is a small lie.
Linux has a ton of different security primitives (including the LSM zoo), and each of them has various user-friendly frontends, with the more interesting ones (docker + other OCI-compatible stuff) composing multiple security features.
2
1
7
563
@mitchellh does this roughly summarize @superlogical's take on multiplexing?
Basically make a very smart client (libghostty) to make use of e2e principle and not get stuck with a smart but slow to adapt middleware like tmux.
4
3
285
Feature unification bites again. Considering adding a CLAUDE.md guideline to always consider alternatives like separate crates when possible.
I think rust features are overused and Im especially fearful when I see them used for DI instead of runtime dispatch to "save on a few ns".
Does anyone have horror stories to share of feature unification biting them in the ass in prod and turning on a mock version of a function?
245
I think rust features are overused and Im especially fearful when I see them used for DI instead of runtime dispatch to "save on a few ns".
Does anyone have horror stories to share of feature unification biting them in the ass in prod and turning on a mock version of a function?
3
509
Anyone tried this? Are there other similar IDE plugins that grey-out files and progressively disclose them while you are learning a new project?
Wouldn't need github.com/GeeeekExplorer/na…; could go straight to the vllm codebase and only be shown a small slice of it.
1
1
359
Take 2 at comparing a plan produced by @humanlayer_dev and one without, on a nontrivial (3 repos) change this time.
Humanlayer wins, big time. Not only did it plan for a nicer serialization, I also understood the plan and the high-level ideas/decision points much faster.
1M+ tokens, 4 huge planning md files and few hours later, comparing code generated from a @humanlayer_dev QRSPI session with a one-shot impl from a terse linear ticket...
"the plan's hazard audit is wrong at 2/3 sites..."
Not a great start.. what am I doing wrong @dexhorthy ?
1
1
1
657
I love the review UI that allows leaving inline comments and having the LLM read them and act upon them!
Could be improved however:
1. can't open unmodified lines to read context
2. no way to open vscode in the workspace folder (with all repos as subdirs) directly, and instead opens only the default repo
3. it seems like the "fetch reviews" mcp is doing network roundtrips to your backend instead of just reading them locally..?
3
1
115
Speculate all the things!
Introducing Speculative Programmatic Tool Calling (sPTC)!
A general class of technique for speculating on tool calls during code generation in a harness and queuing them early to overlap with token generation + REPL execution time.
Blog: alexzhang13.github.io/blog/2…
142