Ian Kronquist

Operating Systems & Security. Works on iOS & MacOS. Likes hiking & cooking. Also @[email protected]

[Halvar Flake @halvarflake]

I'm German, a culture so worried about alignment that the french resistance had booby traps triggered by hanging pictures on the wall slightly misaligned, only to explode a charge when someone aligned them.

[Ian Kronquist @IanKronquist]

@never_released @mweinbach What’s so bad in your opinion?

[third-party evaluation partner @mycoliza]

i wanted “a bicycle for the mind” but what we got was “Doordash for Thinking”

[Ian Kronquist @IanKronquist]

@AlexHollings52 Does this range limit change in an in-orbit scenario where presumably thermal blooming doesn’t apply?

[Ian Kronquist @IanKronquist]

@landaire @heliumbrowser I don’t think there’s even a proper UI to enable it, you just have to go to the preferences key value store UI and check a box. I can’t check right now because I don’t have a copy of Firefox handy but just go to about:config search for tabs and read the options.

[Ian Kronquist @IanKronquist]

@thelcjones What paper is this image from?

[domas @xoreaxeaxeax]

Everyone's trying to make CPUs faster.

I’m trying to make them worse.

New project: CPU deoptimization.
Searching for the slowest possible machine instructions

x86 single instruction record:
198,002,498,236 cycles, 62 seconds

The assembly hall of shame: https://t.co/G4QnFQjsZx

[Ian Kronquist @IanKronquist]

@magikarp_tokens Are you just solving for what the tokens are or going deeper? I would love to know the science behind why those tokens were chosen, I suspect it’s more complicated than minimizing tokens spent over the train set.

[Ian Kronquist @IanKronquist]

@gamozolabs They are all invariably a PITA to figure out and set up but I think it’s important to know eg whether your attention kernel is in the compute bound regime or the memory bound regime, and whether you’re saturating all of the compute ports or caches you expect.

[Ian Kronquist @IanKronquist]

@gamozolabs Besides top level tok/s how are you measuring perf? I know there are some nvidia tools I’ve had a hit or miss experience with but it can be very interesting to know what the performance limiters of each module/fused kernel are.

[Ian Kronquist @IanKronquist]

@Niko_sk2x @mycoliza If the implementation matches the spec and the spec covers the necessary security properties. But I’m a fan.

[Ian Kronquist @IanKronquist]

@mycoliza 1. If bug density falls but the total amount of software increases the total number of vulnerabilities may stay similar. 2. In a world where attackers have larger budgets than defenders, we may wind up in a similar equilibrium without a stark change in development process.

[Ian Kronquist @IanKronquist]

@kerckhove_ts If you’re tired, tipsy, hungry, or in a distracting environment and still want to play chess, you probably won’t perform at your usual level. Playing unranked in those situations relieves the worry of tanking your normal rating.

[Mateusz Krzywicki @krzywix]

Did any1 do an experiment holding on to a vuln found with Opus/Mythos or GPT for a while to see if they snitch?

[Ian Kronquist @IanKronquist]

@bztree It might be worth benchmarking the matmul kernels MLX ships as a point of comparison. They are open source and seem pretty well optimized on the M1 anecdotally and have code paths for newer hw.

[Ian Kronquist @IanKronquist]

@neomewtwo I ordered a DGX spark from aliexpress for the same price or so. Got a postal tracking number and eagerly kept checking as the presumably completely unrelated package got delivered to the next state over. Got my money back.

[Ian Kronquist @IanKronquist]

@MarkKronquist @OfficialJoel4_ 36

[Ian Kronquist @IanKronquist]

@dfsec_com Pics or it didn’t happen