ilc

What harness/quant did you use for testing?

Not the parts which matter to trust. Which is my point.

You can state the math, but not why it won't discuss various topics, etc. Once you see the models waffling on subject with objective truths. You wonder what else is wrong.

I do not exempt US models from this. They do it too, ask anything about politics, elections etc. And they can get... weird.

It doesn't take much to create a systemic error class in a model at these scales. And history has shown nation states are willing to do these things.

Just be wary.

Training data, training methodology. All NOT OPEN.

Until we know what a model is trained on, and how it is trained in high detail, I hesitate to call them "Open Source" in any way. They are free. But, we don't know what their priorities are etc. Witness the censorship we see in all models in one form or another. I'm not absolving any side of this.

Just saying: Don't be blind.

Opus: Yeah, it can be a little fuzzy at 600-700k, I prefer it 400K and under in general. But, I've had usable sessions to 850k.

It really depends on the session, and what you are doing how far you can push it.

There will always be abusers. The issue I see is day 1 asks. Not day N. If I'm out a week asking why is reasonable. If I'm out a day, it's more annoying for everyone.

And I'm not advocating for no limits with no responsibility to the employer. That'd be stupid. Clearly there's a "more sane" middle ground.

Even on a menial job, they are dealing with more value than they make daily, by definition.

A middle road exists, and it is solid, why it isn't being taken, I don't know.

Misinformation. Incorrect sorting selected to prove a point.

Nice try OP.

Disclaimer: I'm a US Citizen / Worker, white collar.

My first job had: Unlimited sick time. Take it when you need it. No doctor's note. I think you may have needed one at the week mark. Never hit at that employer. Most employers have been limited sick, with a note at a week.

I can't imagine note on day 1. That's just... nuts. The lack of trust shown there is massive. I'm making calls on stuff far larger than my sick days. If you can't trust me on sick days, how can you trust me to do my job?

Given that they directly compare to GPT-5.5 in their documentation. This comes off as puppy kicking to me. They state it is not SOTA, even IN its domain!

Honestly: Think twice before dragging your firm into what you say.

Disclaimer: I speak for myself. Not any firm I am associated with.

I could, I hit it in some situations. But I'm running some fair size repos where I do multi-worktree work.

That said, it hasn't been enough of an issue for me to want to fix.

Built my own orchestration layer, to allow me to optimize the one thing that is most precious when working with AI. My attention and my time.

It's dead Jim.

When you realize that being a great "programmer", isn't about writing the most code, but getting the job done...

AI will click as another tool in the toolbox.

I can't summarize 40~ YOE of programming easily. (30+ professional)

I can tell you: Your problems are a layer higher than you think.

Coding, Architecture, etc. Those get the face time. Process, and Discipline, and where the money is made and lost in AI.

To give a minor example: My first attempt at a major project with AI failed HORRIBLY. But I stepped back and figured out why. What short-comings did my approach have, what short-comings did the AI have. Root Cause Analysis.

Next day I sat down with the AI and developed a PLAN of what to do. Yes, a day spent on a plan.

Then we executed the plan. (or it did and I kept it on track, and fixed problems in the plan as things happened.) On the third day I'd completed a VERY complex task. I mean STUPIDLY complex, something I knew WHAT I wanted to do, and roughly how, but not the exact details, and not at the level to implement it. I'm sure 1-2 weeks of research could have taught me. Or I could let the AI do it.

... And that formed my style of working with AI.

If you need a mentor pop in the Svalboard discord, and join #sval-dev. You should be able to figure out who I am.

As a Principal SWE, who has done his fair share of big stuff.

I'm excited to work with AI. Why? Because it magnifies the thing I do well: Make technical decisions. Coding is ONE place I do that, but architecture, debugging etc. All use that same skill. Making good technical decisions.

And if you can make good choices, AI is a MEGA force multiplier. You just have to be willing to let go of the reins a hair.

I think you over estimate how valuable really good Principal level talent is when you have AIs that can take over for entire teams.

As an older and higher up engineer, I worry more for the youngsters than myself. I'll find a spot. I'm using AI, I'm doing things at rates that are pretty crazy.

That's all powered by decades of good decision making practice. Youngsters don't have that. They don't have the painful lessons hard earned.

Minor note to anyone from taalas:

The background on your site genuinely made me wonder what was wrong with my monitor.

LLMs are where you need the most tests.

You want them writing tests especially in critical sections, I'll push to 100% coverage. (Not all code goes there, but thing that MUST work or everything crumbles. Yeah I do it.)

There was one time I was doing the classic: Pull a bug find 2 more thing. And I just told the LLM. "100% test coverage on the thing giving me problems." it found 4 bugs, fixed them, and that functionality has been rock solid since.

100% coverage is not a normal tool. But when you need it. Man does it help.

I disagree.

Freedom of Speech guarantees the right to speak. Not the right to have no repercussions.

Elon has GREAT interest in Freedom of Speech, it enables him to have far more power than regulating the type of "speech" he showed in cancelling that customer's order.

It can. But I don't want to compete for my slot with others who can give the same talk, or a talk that is similar.

I want to make the conference committee choose between "Do we want ilc's talk on X." or "Do we want foo's talk on Y." If we are both discussing the same thing, if I'm unknown, I will lose. OTOH, if I have something interesting to talk about... I have 2 routes to "victory". "ilc gives great talks, he gets good grades and is working on his skills." and "Man that's a damn cool topic. We want that at our conference, even if ilc isn't the BEST speaker, the combo is better."

I didn't start out as the best presenter. I learned. But I always knew I had to have an interesting topic, something that made it worth them giving me a slot.

No, it means you have something unique to say.

The bar is there, but it is lower than you expect. If you have a truly unique point of view to express, that brings some value to the table, slots will open up.

And I've spoken at plenty of conferences. :) Not always in the glamour rooms/slots. But... I did have one talk fill a room out the door. That was a talk on a difficult/controversial topic, and by then... I was probably about as expert as they came on the issue.

I didn't start with that though. I just started with a simple point of view talk. And I'd argue the second version of that talk is still one of the best I've given in my life.

No, I put them with lmgtfy. You are being told that your question is easy to research and you didn't do the work, most of the time.

Also heaven forbid, AI can be right. I realize this is a shocker to many here. But AI has use, especially in easy cases.

I think AI clouds the real issues around Junior hiring. Defective companies.

Let's say you hire your great new engineer. Ok, great! Now their value is going to escalate RAPIDLY over the next 2-3 years. And by rapidly, it could be 50-100%. Because someone else will pay that to NOT train a person fresh out of college!

What company hands out raises aggressively enough to stay ahead of that truth? None of them, maybe a MANGA or some other thing. But most don't.

So, managers figure out fresh out of college == training employees for other people, so why bother? The company may not even break even!

That is the REAL catch 22. Not AI. It is how the value of people changes early in their career.

So you can get rooted by the security issues disclosed.

Isn't it a wonderful catch 22?

Thanks. :) And yeah, I skimmed it. I stand by my comment.

To Quote the Page:

Notes Revise and resubmit, Quarterly Journal of Economics

I'd wait for the revision.

The same way I would with a human:

If I thought the service should only be 1000 lines tops:

- Reject due to excess complexity.

If it is a proper solution:

- Use AI to review it, asking it to be VERY critical of the code, and look for spots where human review may be needed, architecture wise, design wise and implementation wise.

- Ask the AI again to do a security review etc.

- Tell the author to break the PR down into human size chunks using git.

Why those things? It's likely some manager is gonna tell me review it anyways. And if so, I want to have a head start, and if there's critical shoot down level issues I can find with an AI quickly. I'd just shut the PR down now.

As in any "security" situation, in this case the security of your codebase and sanity, defense in depth is the answer.

What you take if you use a bad one?

Was. They've been diverging.

https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct

Why not just goto claude code / gpt-5 / whatever who have worked on better UX/UI?

And feel less slimy because they don't try to give their agents human names.