eutropia

It sounds more like she was arrested for not following an order given by the police officer to move out of the way of the robot.

I think the bit about "saying what you write" is what stands out.

I wonder if you took the pure speech to speech models and STT'd the output if it would remove the LLM voice? After all, can you even imagine hearing a human utter aloud some of the nonsense you see spewed by claude? Or give a longform presentation or speech in the form of the essays that are 100% LLM-written?

Maybe there's something in our human brains connecting oration and writing that influences the style of writing that pure-text engines just lack?

tandem OLED is one of those techniques that makes me think "why didn't we just do this in the first place?"

"emergent" is just a word people use when they can't explain how the properties of a system lead to its more complex behavior.

There's no magic involved, just as-yet-unexplained phenomena.

Everything in the chain of causality has some explanation or mechanism, understanding it step by step dissolves the illusion of "emergence".

wait does the math on this work?

  $1T  = 1,000,000,000,000
  
  $10m =        10,000,000
  
  *365 =     3,650,000,000
  
  *300 = 1,095,000,000,000

welp.

No this is more like having someone standing over your cubicle, and every other person's cubicle, taking notes of every time you go to the bathroom or get a drink of water or stretch.

Watching you.

All day.

every day.

Weird, I've never (that I can recall) had a comment flagged on HN before.

@dang (not sure that's really a thing) any idea what's wrong with my comment on this thread?

My mother used to work as an environmental consultant in her career as a chemical engineer; and one of the larger projects she helped direct was a water treatment facility for a pit-mine-turned-lake.

it turns out that all of the things leaching into the water made it quite acidic and relatively toxic!

so, y'know, probably good to check what kind of mine it was before going for a swim!

the real reward is the equality of dignity we earned along the way.

it's a real cost, and a real benefit. I guess values vary on which is more important.

at current prices it's above $1000/kg to orbit, and we're just conveniently ignoring all of the operational constraints of a completely unmanned datacenter upon which you can perform no maintenance.

What do you do when a micrometeorite causes pinprick leaks in your radiator loop?

is LEO sufficiently shielded from cosmic and solar radiation?

What's with those giant blobs of asteroids offset in jupiter's orbit by like 60deg on either side?

More applications need the ability to log in as a "read-only" mode, so you can more safely grant access to tools like this.

I can imagine something like Dots being utterly invaluable for running a traditional brick and mortar business, streamlining all of the admin work, but until they're really safe and well integrated we'll have to wait...

I mean it sucks but it kinda seems like their work training new models has vastly outpaced the ability to comprehensively secure those systems.

I don't think it's because they're incompetent or lazy, but that the nature of the problem is that security vulnerabilities seem to scale superlinearly with complexity but model training scales linearly or logarithmically with complexity.

But the organization isn't allocating the resources accordingly, which would likely be economically unsustainable for the competitive environment they're in. Put simply, if the security team was more than twice the size of the research team, they might have a chance at keeping up.

I am curious about the physical layout for moving the highly-concentrated heat in the GPUs out to the diffuse radiation panels required to keep them at operating temperature. Would it need something like a liquid cooling loop coupled with large radiators?

and also curious how it's cheaper to send all the expensive silicon to space and throw it away when its obsolete compared to just racking it in the midwest somewhere?

Right after hiring the CEO of character.ai as its new CTO.

standard israeli psyop tactic: use jewish people as human shields from criticism directed at the state of Israel by trying as hard as possible to conflate antisemitism with negative sentiment towards israel.

I'm super excited for all the people saying "lol just point it at a genai photo" to make a blog post later about how they successfully subverted this approach and fooled people into believing their image was verified without detection as a fraudulent image. because that'll be a lot more interesting than a reflexive snarkpost about the first idea that came to mind to hypothetically defeat it.

I think the timestamp attestation of the digital negative puts a real hamper on this, because it puts a bounded time window on the photo as part of the cryptographic chain of evidence.

So if you're the proud owner of "literally the only photo of a ridiculously unusual event in a highly public area" which is bounded to either a plausible 15-30 minute window or a sketchy March2026->Now window, people can do something like "hey, gee, did anyone else see that UFO over the golden gate bridge at 3pm?"

plus, you know, the confidence score from their secret neural network, which has an unknown scoring function.

The UI looks fun but reminds me of the hours and hours I spent tweaking my linux distro as a younger nerd.

It didn't help me get anything done, but it sure did look computery!

If pre-ipo marketing pushes them to train a model capable of resolving a millennium problem in mathematics in a weekend, then, to quote XKCD:

  "Mission. Fucking. Acccomplished."

https://xkcd.com/810/

The mat is the objective (or at least distinct from your own personal judgement) measure of your goal. If your goal is important to you, an objective measure of your progress towards it is more valuable than your own flawed and biased opinion.

The mat for the diary might be the times you consulted back to it for useful introspection and came out better for it, the mat for your health goal could be your blood test from the doctor, etc.

The point is that the mat doesn't care what you think of yourself. I would ask you: when you judged your progress from yesterday, was it with a mat - a brutally objective test - or was it with something less honest?

Whatever your goals are, if you really, truly care about them, and you understand what your true objective is, you'll value the mat. The mat is not about what other people think, it is about undeniably simple and real feedback on your objective.

Do his predictions about the datacenters being a risky and unprofitable business suffer the same fate as the others?

I sighed at the pleasant but claude-looking design of the page.

idk what's wrong with me. it's a good looking page but I just instantly didn't care the second I saw the close-to-solarized-light color scheme and the typography.

It's similar to my reaction of sites with the maliciously compliant cookie popups or the auto-play video scroller things.

All the parasitic grifters founding these companies seem to be characterized by a uniform lack of human decency.

if the best idea you can come up with for a company is is some variation exploiting humanity, you're not a visionary or an entrepreneur.

You're just an asshole.

This is a modest start on an important direction for AI Alignment work; which is, as the authors observe, commonly comprised of tasks which are not readily empirically verifiable and not easily mathematically modeled - so it's hard to get at with normal RL techniques.

I find the ACCoRD benchmark the most interesting, because you could theoretically scale it up from the baseline mode of testing two instances of the same model for their `P(A) ≥ P(A&B)` respectively, you could do `P(A)≥ P(A&B) && P(A) ≥ P(A&C) && P(A&B) ≥ P(A&B&C) && P(A&C) ≥ P(A&B&C) ...` etc

i.e. a swarm of model instances could be collectively measured for consistency for even more confidence, right?

At any rate, even the basic idea of measuring a model for consistency in beliefs improves our ability to bound the amount of trust we can put on it with introspection methods.

another firefox on linux user (cachyos), and it loaded up and worked flawlessly for me. anecdata!

as someone with ADHD who always finished exams first or was the designated spokesperson for group presentations...

I have no idea why you'd want to be taking a test a second longer than you already have to.

Either you know the material or you don't, anything taking a lot of time is borderline stuff where you're scrapping for points you might not deserve in the first place lol

Many people set their DNS to 1.1.1.1 -- would be easy for cloudflare to resolve any TLD it feels like

the person cooking it is Kenji Lopez-Alt, who is an MIT graduate turned professional chef, and his wife is the one who's colombian.

He's got several excellent cookbooks (like the Food Lab and Wok) and is a regular contributor to serious eats and has youtube etc...