arthurcolle

Classified

Plant capital are people too!

Yes I profess that these antiquated terms are deeply disturbed

How much of the copy was produced by AI

this feels like AI slop

He's part of the military industrial complex

FriendFeed shutdown same day Facebook was incorporated - these things are important for continuity

Goldman did a gold <---> painting swap via ETH like 12 years ago. The ship has sailed. It's okay, you don't have to like the calculator code as a DLL to use the calculator

I'll have to tell my ducks about this one

Tailscale Down™ - What goes up must come...

fade out

Uncertainty as a Service®

Thought this was for a product release called Down. Might be time to go outside for me

If you had sudo permissions on the situation what 10 steps would you want to see happen to resolve this whole affair?

Yeah, I was wondering if they were honoring Suchir Balaji

Thank you for the quick follow up - I agree with your proposal on linking the black bar itself to major thread

Probably not particularly useful but GCHQ & NSA both have neat graph related repos

UK: https://github.com/gchq/Gaffer

US: https://github.com/NationalSecurityAgency/lemongraph

Maybe it is complete

What were the issues with Metal.jl?

Nah they don't do that for the pro models

check out the HP Prime sometime. I got it for myself in a moment of retrocausal nostalgia and it's pretty awesome. I'm using 95% of my willpower not to get this linked product haha. Moving around a 3D surface with my finger just never gets old

Why are executives allowed to lie, and the rest of us have to just deal with it? At some point the chickens come home to roost

Closed Circuit maybe just Cloud Computing TV now haha

My work got me a $6500 laptop (128 gb unified memory m4 max) and I had to get a replacement for my self, paid $4000 for 48gb unified RAM and feel completely ripped off

Sad

Then call it something else

It has access to a compressed representation of some subset of the information it was trained on, depending on training regime.

By this, what I mean is. Take an image of this: https://en.wikipedia.org/wiki/Traitorous_eight#/media/File:T..., change the file name to something like image.jpg and pass it into Qwen 3 4B, 8B, 30B and look at the responses you get:

It has no idea who these guys are. It thinks they are the beatles, the doors. If you probe enough, it'll say it's IBM cofounders. In a way, it kinda sees that these are mid-1900s folks with cool haircuts, but it doesn't recognize anything. If you probe on the F the model in question becomes convinced it's the Ford racing team with a detailed explanation of two brothers in the photo, etc.

The creation of autoregressive next token predictors is very cool and clearly has and will continue to have many valuable applications, but I think we're missing something that makes interactions with users actually shape the trajectory of its own experience. Maybe scaffolding + qlora solves this. Maybe it doesn't

What claims specifically have I made that you have evidence to the contrary indicating that I am wrong

I don't claim to know better. But restarting a $1.5B plant after 2 years of inactivity and having a worker fall into a vat of radioactive water and still being at 300 CPM after a decontamination procedure is not normal.

But I am not doing surgery. I am expressing skepticism at the "oh no it's all fine" from literally everyone reporting on this story

My use case for data that exists that is pre-AI scientifically vetted work is completely divorced from the specifics of this conversation actually. If I want to do paper-maché sculptures with printouts of these papers, and I still commented on this post, would that be better or worse for you, here?

I was just sharing background. I want to make good models that can help scientists do work. Your personal feelings about LLMs and their capabilities feels quite distinct from the focus on this post, and the chain of comments that have led us here.

Thank you for the followup (familiar with the XKCD)

Dumb question from a true non-expert:

So CPM varies with all those factors you mention, but wouldn't the site HP team know exactly what detector they used, the geometry, distance, etc.? They could convert to dose if they wanted, right?

Why report the ambiguous "300 CPM" instead of an actual dose estimate in mSv/μSv? Seems like that would be more useful for any medical team, any set of potential regulators or regulatory bodies as well as just general public understanding (drawing on my father's work here as he always emphasized the tension between "public fears radiation unnecessarily" and "industry safety protocols are inconsistent")

Follow-up: Is there any legitimate reason to report CPM instead of dose after a contamination event? Or does staying with CPM keep things conveniently vague? Because from my limited understanding, if they did a proper survey, they have everything needed to calculate dose.

I'm sorry you feel that way, but I actually feel like I shared fairly demure commentary and then presented my background with full transparency. I care mostly about the health of this worker, and less about the opinions of a 1 minute old account.

I maintain the complete archive of every publication my father did from 1969 to 2019 and continue to update the archive based on new publications. I use the data to train Nuclear Radiochemistry AI Agents and while I do not have my father's credentials, I actively use this dataset to learn about his field, and from my limited knowledge here I felt the need to comment. After all, what is skepticism if we can't share and teach each other what we know, right?

https://github.com/arthurcolle/Ronald-Colle-Papers

I love how punchy you are! And the astronomy photos. Take care :)