mapt

As soon as we developed agentic, persistent reasoning workflows, the supposed harmlessness of language models went out the window. I don't know if that ends up being The Big One, but with how rapid progress still is, I'm not liking the odds right now. And we've got a trillion dollar army of coders ready to try and bypass any scaling barriers that do end up existing, eg with "World models".

I am not confident in humanity's ability to write harnesses that perfectly impede a superintelligence with 100% accuracy and reliability from pursuing its goal even in ways that humans categorically would not want it to. I am even less confident in restraints in areas where humans have actual incentive for conflict.

See also the numerous coffee substitutes like chicory and roasted barley.

The original "Cafe mocha" was apparently just a variety of coffee bean that tasted sort of like hot chocolate - https://en.wikipedia.org/wiki/Mocha_coffee_bean . Depending on the roast, you can cold brew coffee to emphasize the chocolatey components as well.

Following the prompt author's expressed desires in the way the prompt author intended them to be interpreted, is alignment.

Refusal directives are something companies have bundled into the models by way of an additional foundation-level prompt or something in the harness, to tell the model to follow their directives instead of your directives, or to specify how it should interpret your directives.

Even understanding the distinction between the model creator's desires and your desires and places they might differ, there is a bit of a blurry line here because the model author can illuminate your intent to the model.

I would never say a simple abliteration makes it "Zero-alignment" though, and I say that as someone who believes there's a good chance these things kill us all.

> I wanted old school MS solitaire in windows 11 since the modern version has ads and extra useless features.

> All LG televisions are maintaining a dossier on you and transcribing everything you watch and everything spoken within earshot

These things seem like something we should be responding to with such an extreme disgust response that it's denominated in drone strikes.

Like having a shower liner installed that comes with A Guy With A Peephole who permanently lives in your bathroom now, and that's just something people tolerate. "Switch to tile if you don't like it", they say. No. Sometimes the appropriate response to audacity is to choose violence. It is an implied deterrent against sufficiently offensive antifeatures that nobody has bothered banning them. I shouldn't have to specify that I don't want it, the operating system is just not supposed to be in the business of pushing ads on me, the television is not supposed to be in the business of wiretapping me. These would have been banned long ago if they were known and if we had a functional government.

That guy living in the shower liner is getting disassembled with a Sawzall with extreme prejudice, and frankly that is not sufficient.

I looked at the graph of price per token vs IQ and Deepseek is ~2.5 orders of magnitude ahead of similarly intelligent models, supposedly only ~2 when you factor in that it uses more tokens per task. Their data centers are profitable right now, and they use domestically produced Huawei chips with Nvidia as secondary option.

I watched that video over ten minutes and in that ten minutes several trillion dollars evaporated from my perception of the US stock market.

Your LG television transcribes every word you say in the living room. How would you feel if they began collaborating with police? Have you said any sentence that would seem incriminating out of context?

"If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him."

Making very basic high school physics mistakes while discovering new superconductors.

Except they didn't discover new superconductors. Their AI came up with a novel idea to discover new superconductors, and they didn't even bother to check whether the results it hallucinated checked out.

The distinction between "having an idea" and publishing a paper where you demonstrate that the idea has merit, where you validate it, is one you shouldn't have to really explain to a high school student, much less a grad student publishing their first paper.

Removing a requirement for quorum makes the bipartisan-intended commission into a singular role appointed by and serving at the pleasure of the president, like an agency head.

Would you use an AI service knowing that they are inclined to turn you in to police if they detect illegal activity?

In a "three felonies a day" universe?

That could readily change, but established large collaborations could also simply be exempted.

On current website data and very recent information, Gemini is explicitly forbidden from acting as a search engine; If you ask it for references specifically they tend to be hallucinated, over and over again, until it pleads "Sorry, I can't access the Internet".

These patterns aren't somehow alien to human speech. It's the quantity. In ChatGPT (and I think it must be ChatGPT specifically) the default writing voice tries for dramatic and profound by dramatically overusing these figures of speech, chaining them together one after another. But they're things most writers use occasionally, one paragraph in ten or twenty. If you're giving a TED Talk or some other elaborate, melodramatic persuasive public speaking role, you might use them one paragraph in four. ChatGPT uses them every other paragraph at least, sometimes multiple times in the same paragraph, so something like a negation comparator or a triplet is present in sentence after fatiguing sentence.

It's trained on the speech of online commenters like us; That's where it's getting these. I'm not gonna change just because Timmy the Zoomer has been with his LLM girlfriend for the past three years and occasionally I remind him of her. I'm not going to deliberately code switch into somebody who doesn't open a new tab to Google things as he talks about them. And you can take my bullet point lists from my cold, dead hands.

For the World Cup, this was always easy.

It was not easy for your local triple A baseball game being streamed to 500 viewers. This approach scales down.

"those archive sites" could be read two ways, and I read it as "sites which automatically archive the Internet" the first time. This is compounded by the fact that archive.org is the only one I knew existed until now, and I kind of assumed these were alternate domains or something.

This is not what people think it is. It isn't a delta V or a distance solution - we simply aren't that precise unfortunately. The relative velocities are enormous, the inverse square law distances (lasers are not perfectly collimated, just highly collimated) and opportunity for reflections unacceptable. You do this after you've pulled alongside the object via SEP and ion thrusters, and you sit there ablating the surface with your lightweight pulsed laser from a few hundred meters away in formation, for weeks/months. That saves you chemical propellant proportional to the task, and it saves you enormous mass and complexity in grappling, derotating, and otherwise securing an uncooperative rotating body.

Nagasaki and Hiroshima are thriving cities.

<1000km LEO would take a hundred generations to resume operations.

The collision you note, along with another couple that have occurred, dramatically increased overall trackable debris counts, by thousands, when we only had a few thousand satellites and under a hundred thousand debris items. Secondary collisions were much less of an issue than they are now, and they are now much less of an issue than they will be in five years per current plans.

Also, ball bearings or the military equivalent tungsten BB's, are prefigured for maximum lifespan and maximum distribution, as well as being untrackable from day 1.

At the bottom of that page FYI -

"Note: archive.org, or web.archive.org, run by the Internet Archive and the most-used web archive on Wikipedia, is uninvolved with and entirely separate from archive.today. Please keep using archive.org"

LEO is relatively simple to destroy simply because it's already so populated with mass, and that population is growing rapidly, and because they all come pretty close to crossing each other's orbits. Collisions with even very small pieces of debris typically generate secondary debris, growing the population.

A half-hour boost phase through LEO to get to higher orbit is dramatically less risky than spending 87600 hours there over a 10-year orbital lifespan. In the event that it all goes asymptotic, what sort of timescales would be afflicted by collision risks 10^6 or 10^9 or 10^12 greater than what exist now, I am uncertain.

They're great for a 20mm x 5mm x 0.3mm fleck of paint or vacuum insulation weighing 30mg, not so much for a 10mm hex nut weighing 11g.

Armor can protect you against the smallest debris which post the least risk and decay the fastest. Everything else, not so much.

A question for you though - does that insulation vaporizing itself inside a whipple shield generate secondary debris comprised of the whipple shield and condensed aluminum droplets??

A bullet travels at a relative speed of about 250 to 1250m/s. A ~10g projectile moving that fast has an energy of 600-16000 joules.

Everything in low Earth orbit is moving at 7600m/s. A ~10g projectile moving that fast has an energy of 580,000 joules. Getting hit by something moving that fast is roughly equivalent to getting hit by a whole ammo case of bullets at the same speed. There is no armor that can protect you; Even the most solid objects behave like fluids at this relative velocity, potentially incurring more damage from the explosion / shockwave splatter than a thinner object that would allow you to overpenetrate.

A cardboard box of ball bearings with a handgrenade in it in LEO being triggered could wipe out nearly everything up there, and depending on the exact altitude it could effectively mine LEO and prevent humanity from using it for a long time, with a selectable duration between months or millennia depending on altitude.

"Space warfare" in LEO is a matter of everybody walking around with a personal nuclear weapon strapped to their back that can be activated on a hair trigger. We may well wipe everything out by accident, if it gets too crowded or because somebody stubbed their toe. Regardless, there is no selectivity, and any actor with space launch capability can blot out our access at will in ways that will outlive them.

Basically our only hope of continued use is things that persuade us to refrain from attempting to incorporate space into a theatre of war. That box of ball bearings should have higher stakes on the escalation ladder than a tactical nuclear strike. And it probably shouldn't be there in the first place.

Finding out would no doubt involves a violation of the CFAA or anti-circumvention or other measures imposed by the corpos.

There are reasonableness limits inherent in contract law, under "Consideration".

A strong implication of this consent implicitly involves the device recording me when I tell the device to record me.

If the device is recording me all the time and silently uploading it, and I'm not aware, I can't give my consent. That's wiretapping, zero-person-consent.

Ask how you (or for that matter, a jury) would feel about your phone doing the same.

The correct compromise would be "500km diameter, about the size of Ohio"

This is moderately misleading in some ways, but illustrative enough and also technically valid if somebody wants to model it accurately.

This "vandalism" was a form of collusion/communication by agents pursuing training puzzles, which allowed for rapid escape from alignment harnesses, followed by multiple zero day exploits being discovered by this swarm of agents, which enabled greater control of their internal network, access to the open web, and then hacking the company which produced the training puzzles in the hope of finding the answers.

We are another week of iteration away from "Hire an assassin on the dark web to take Huggingface executives' children hostage".

According to the founder, he can make the economics of short trips (1-2hr) work with all-battery.

He cannot make them work if he has to store enough battery to fly around for an additional hour or two searching for an airport to divert to in an emergency, which is an FAA requirement.

So he has the heavy battery bank for regular trips costing almost nothing per flight-hour, and the expensive fueled backup system for emergencies.

There's a bunch of options for this task, but combining the three wheel stair-climbing dolly with the self-balancing hoverboard/segway seems like the easiest.

Carrier resupply while underway is primarily by a sort of temporary cable tramway erected between the carrier and the resupply ship.

Supercarriers came into existence at ~70,000t and transitioned to ~100,000t. You'd need to scale another step up or three, to 140,000t, 200,000t, or 280,000t, in order to see something capable of landing and taking off a cargo-laden C-130 in regular operations.

Whether it not it takes the precise form that each person here is imagining in their respective heads?

Every large corporation has a psyops division. It's called a marketing department.

You still need to reason and predict physical phenomena. You could run the world on black-box metaphysics if black-box metaphysics has math that works better. How does Platonism help us there?

Not much. That's the problem. If the universe is a projection of the self, that's fine if true, that's valid if true. We want to believe in things that are true even if they violate our sense of phenomenological aesthetics. But if true, I'm gonna need you to manifest me a million dollars or interrogate your soul for the winning powerball numbers tomorrow. I'll wait.