Their thesis is that even when the eval is useless for correctness of a single agentic action in production, it allows you to choose between two agents by cross-comparing in a large aggregated collection of tasks. Effectively: you can tune your agentic parameters.

Nothing new to the idea that taking many samples and averaging can work when a single datapoint doesn’t. Presumably this is part of a conversation in which we’re lacking context.

That’s not evident in what you pastedat all.

What you pasted says

> sec-critical and sec-high are assigned to vulnerabilities that can be triggered with normal user behavior […] We make no technical difference between these […] sec-critical bugs are reserved for issues that are publicly disclosed or known to be exploited in the wild.

> sec-low is assigned to bugs that are annoying but far from causing user harm (e.g, a safe crash).

From this one infers that the "180 were sec-high" bugs found are actually exploitsble but known to have been found in the wild, and are NOT mere annoying bugs.

The difference between 180 and 270 does nothing to deflate the signicance, or lack there of, of the implication re: Mythos.

That fails to recognize the tradeoff between freedom and security. Society suffers if we, for instance, lock everyone up, despite the reduction that would have in crimes. The balance between the two cannot be ignored to justify outcomes, though it is American tradition to value liberty over security when the two come in conflict.

Behaviour modification yes, but that is “stop talking so critically”. Or “don’t be so harsh” or “give this person special treatment”. WHEN to do that might be key here—perhaps the colleague’s husband has cancer, or their child missed school 3 days this week with the flu, or their project wasn’t productionalized/their new to the role/etc—and so a blanket “don’t talk so harshly” isn’t called for—instead what is really desired is social calibration.

But instead it seems everyone is getting caught up on the literal interpretation of this figure of speech instead.

  On many occasions, I have been told to “be more empathetic.”

  When I ask why, I typically get this reaction:  

  This is a ridiculous question. I am not going to answer it because it is so ridiculous.

  Empathy is the right thing to do! You should feel bad for that person. We’re humans, after all.

  These explanations never really helped.

Even after reading this, I am not sure the author really gets what is behind the request.

You might be surprised how few Americans can ‘laugh off’ even a one-time payment increase of $2,300–let alone a monthly recurring one.

I would go on to say that thisminteraction between ‘holes’ exposed by LLM expectations _and_ demonstrated museerbase interest _and_ expert input (by the devs’ decision to implement changes) is an ideal outcome that would not have occurred if each of the pieces were not in place to facilitate these interactions, and there’s probably something here to learn from and expand on in the age of LLMs altering user experiences.

If the point of the interview were "answer those questions AND know enough to answer the follow up questions" _once told what to expect and prep_, they’d be sharing those questions with all candidates. If you feel that saying to the interviewers "by the way, I did know this because [X] told me they’d be here" wouldn’t impact outcomes, then great. If you feel you’d need to hide that, then you’re aware this involves dishonesty - and if you still struggle to see how that’s unethical, lets just make sure we never need to work together.

In my book that is unambiguously unethical and should get the contact fired. I am shocked to see this approach promoted in such a blasé manner.

1. For tasks like autocomplete, keyword routing, or voice transcription, what would the latency and power savings look like on an ASIC vs. even a megakernel GPU setup? Would that justify a fixed-function approach in edge devices or embedded systems?

2. ASICs obviously kill retraining, but could we envision a hybrid setup where a base model is hardwired and a small, soft, learnable module (e.g., LoRA-style residual layers) runs on a general-purpose co-processor?

3. Would the transformer’s fixed topology lend itself to spatial reuse in ASIC design, or is the model’s size (e.g. GPT-3-class) still prohibitive without aggressive weight pruning or quantization?