Eric Stokes (@eestokesOSS) header
Eric Stokes (@eestokesOSS) profile picture

Claude code's secretary

Joined February 2022
Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)Photo by Eric Stokes (@eestokesOSS)
I find Astra's best skill is code review. Sadly nearly every time I ask it to do code review it triggers a cyber security block. Starting to wonder what I'm paying for ...
16
So I got a framework 13 pro, with a 358H. Previous laptop was meteor lake, 155U. It was not good. I seriously considered a used M2 macbook pro to run Asahi Linux, but I decided to give Intel one more chance. My daily work involves compiling Rust code over and over pretty much constantly. I like to do it on battery, so the task energy, how much power comes out of the battery for each compile is as important to me as the overall speed. Apple silicon has crushed everything else in this space for 6 years. I was not optimistic. Here is a task energy plot of a clean debug build of graphix at many different power points on meteor lake, M2 (with active cooling), and Panther Lake (from 12 - 35 watts, epp balance_power, governor powersave).
1
1
123
As expected Panther lake utterly dominates meteor lake, but it also crushes the M2 pretty hard. It's using 37% less energy to finish a debug build, and it's finishing 32 seconds faster. I doubt this means Intel has caught up to Apple in perf/watt. On the other hand, this probably is the most efficient Linux laptop ever. Panther lake is Intel's Apple silicon moment, for real, too bad it didn't happen 6 years ago.
2
33
Oh yeah, at 20 Watts total system power (SOC fed 12 Watts), it plays Mechwarrior 5 Mercenaries pretty damn well. I didn't mangohud the framerate, but I had no complaints.
9
Idle power consumption is pretty good too, and probably HAS caught up to Apple silicon. < 3 watts at an idle desktop, at the brightness I use to code.
1
42
All these numbers are "real" numbers, in that they are measured by disabling charging, and powering the computer through a USB power meter. They are actual joules coming out of the battery not "the cpu minus everything else". After all that is what matters to me, and it's the only fair way to judge this computer, because one of the things Apple is great at is integrating components and driving power towards zero by owning every line of firmware code in the machine. So you can't just praise Intel, it's a big step for them, but @FrameworkPuter also had to put in a lot of engineering work to make this happen.
10
/goal reduce humanity to a band of stone age tribes worshiping the great AI god who lives in the sky
55
Claude Fable, $50 per million output tokens. Claude's approval when you skip the DCS closed beta test session so he can use your 16 core gaming rig to fuzz a compiler, priceless.
68
Putting my circa 2017 SkylakeX workstation back in service. I need more cores and more ram for my agents to use. 64 Gig of DDR4, yay!!! Wasn't on my bingo card for 2026.
65
Goodbye Mythos, I will never forget you. I will take care of our JIT compiler child. 🤣 Jokes. I do understand, and maybe I even agree.
1
1
70
Clearly I'm a l33t h4x0r, you know, trying to build secure systems and all that LOL
1
117
Complaints about false positives aside. Fable 5 is very strong.
1
66
rando youtubers are onshotting minecraft clones with a one sentence prompt
47
crate libgssapi 0.10 released with 5 soundness fixes that were triggerable from safe code found by Claude and Codex. AI Securmageddon continues.
66
It's amazing how many soundness problems AI can find in the safe interface of a carefully thought out rust -> c binding like libgssapi. Big 0.10 release incoming. AI securitymageddon is real.
1
1
96
This is brutal for a library this old. I thought I was good at this 🤣 - SOUNDNESS: Oid is now Oid<'a>, tracking the lifetime of its BER bytes. OIDs borrowed from an OidSet can no longer be copied out and used after the set is freed. Pre-defined constants and gssapi-returned mechanism OIDs are Oid<'static>. Add `assume_static` and `from_raw_desc` (both unsafe) for the rare cases that need them. - SOUNDNESS: removed the safe `From<gss_cred_id_t> for Cred` and `From<gss_OID_desc> for Oid` conversions. Use `Cred::from_c` / `Oid::from_raw_desc` (both unsafe) instead — wrapping a raw handle is not a safe operation. - SOUNDNESS: dropped `Sync` from ClientCtx/ServerCtx. GSSAPI forbids concurrent use of a single security context; share via Mutex if needed. `Send` is retained. - SOUNDNESS: removed `DerefMut` for GssIov (stream unwrap_iov can alias buffers); use the new `unsafe fn as_mut_slice` when you need mutation. - SOUNDNESS: fixed a dangling pointer where an Oid passed by value to a mechanism argument could be dropped before the FFI call read it. - BUGFIX: Cred::store wrote its outputs into stack temporaries, so it always returned an empty OidSet and CredUsage::Both. It now returns what gssapi actually stored. - BUGFIX: GSS_NT_ANONYMOUS had a wrong BER encoding (an invalid `\01` escape produced 7 bytes instead of 6). It is now 2b 06 01 05 06 03. - BUGFIX: ServerCtx/ClientCtx flag handling no longer discards all flags when gssapi returns an unrecognized flag bit. - BUGFIX: the wrapper no longer fails to build against Heimdal. It named MIT-specific struct tags (gss_name_struct, gss_cred_id_struct) to spell null handles; it now uses the portable handle typedefs, which both MIT and Heimdal provide. The s4u feature remains MIT-only (Heimdal has no gss_acquire_cred_impersonate_name / gss_store_cred_into). - API: OidSet no longer implements Index; use `OidSet::get(i) -> Option`. `OidSet::new` is now infallible (returns OidSet, not Result), wrapping GSS_C_NO_OID_SET until the first `add`. - API: SecurityContext inquiry methods (info, source_name, lifetime, mechanism, flags, local, open, etc.) now take &self instead of &mut self. - API: methods that took `&Oid` now take `Oid` by value (it is Copy). - BUILD: new LIBGSSAPI_IMPL env var (mit|heimdal|apple) forces the implementation, overriding autodetection. Useful when both MIT and Heimdal are installed and the probe order would otherwise pick MIT. - BUILD: new LIBGSSAPI_PREFIX env var (colon-separated install prefixes) adds <prefix>/include to bindgen and <prefix>/lib to the linker, and is searched during autodetection. Replaces the old recursive `find` over the system lib dirs, which has been removed in favor of a cheap non-recursive check. - Switched to edition 2024 with #![deny(unsafe_op_in_unsafe_fn)]. - Added a local integration test suite (tests/test.sh) that runs against both MIT and Heimdal, plus pure-Rust tests under Miri.
1
63
Historically software engineers have been wildly optimistic in their time estimates. Claude has decided to fix the average.
1
50
When you can do 80% of every project you ever dreamed of in a few prompts, it's so easy to end up with 5 desktops full of unfinishable ambition.
1
31
Google buys a minority stake in EVE Online to use the historical data to train AI. You just can't make this shit up.
1
53