Anthropic and AI alignment research isn't about making AI that are DnD-style "good alignment", but making AI that have outcomes that are aligned with the goals that the designers intended for them. The chatbot AI model and goals are not the same model and goals for a defense AI.
The goals for a chatbot assistant are to be useful, correct, and not insult people. The goals for a defense AI are to extract correct features, provide useful guidance, and not kill the wrong people. If you are working in defense you already have a belief that your work is morally correct: most of those justifications are either that your work will kill bad people more effectively, and so save friendly lives, or will pick who to kill most correctly, and so save innocent lives. Having an AI that is better aligned towards those goals are better.
You may disagree that working in defense is ever morally justified! But Palantir dont't share those beliefs, and want to do as good of a job as they can, and so want the most aligned AI model they can.
They train us to drop fire on people but won't let us write "fuck" on the side of an airplane because it is obscene. (Col. Kurtz - Apocalypse Now)
Which, when you unpack it, is even more interesting. If you do embrace the emotional aspect of war you end up with situations like the my lai massacre. Does AI have the ability to prevent war crimes while engaging in "legal" killings feels like an interesting philosophical question.
Going by recent events I think the convention is to drone strike them and their entire family anyway, and then tally them up as a confirmed dead terrorist.
That's very misleading. Terrible war keeps on happening all the time, just not so much in the US and Europe for the last 70 years.
Yes, since WWII things have been relatively peaceful, but the key term there is relatively. As we speak a pretty awful war is happening in Gaza, in the last twenty years we've had multiple wars with pretty severe casualties, and if you go a little farther back you get to things like Vietnam.
It's true that specifically atomic bombs haven't been used.
> "Acts like firebombing of Tokyo or bombing of Dresden or atomic bombing don't happen now."
For the time being... Humanity's leaders are increasingly as insane (sometimes more-so) as their worshipers. I feel it's only a matter of time before atrocities and crimes against humanity skyrocket again. :(
> Acts like firebombing of Tokyo or bombing of Dresden or atomic bombing don't happen now
We still raze cities and drop incendiaries. America hasn’t gone to war with a near-peer nonnuclear power like Japan since WWII. To the extent we were faced with the prospect in the Cold War, both we and the Soviets were committed to MAD, i.e. using nukes. (Do you think unilateral disarmament in the Cold War would have lead to peace?)
There has been no militarily useful technology that was voluntarily abandoned. Just constrained. You can’t constrain a technology you don’t bother understanding.
> during the firebombing of Tokyo the US murdered 100,000 civilians
Are you arguing there was a war in which firebombing would have been useful but someone decided it was too mean?
Since WWII we invented better high explosives and stand-off precision weapons. If there were a strategic case for firebombing in a future war, have no delusions: it will happen. (Last year, incendiary weapons were used in “ in the Gaza Strip, Lebanon, Ukraine, and Syria” [1].)
What? Who argued war hasn’t changed with technology?
> civilians weren't murdered on the same scale
War wasn’t conducted on the same scale.
> why you are conflating the use of certain types of weapons and willingly allowing enormous collateral damage
I’m not. Nobody in this thread is. The point is the weapons are still stockpiled and used. We have never agreed to ban a useful military technology. Just contained or surpassed it.
AI will be used by militaries as long as it’s useful, even if it causes collateral damage. We will obviously try to reduce collateral damage. But in part because that makes the weapon more useful.
I know what you mean, but I don't have an answer myself.
Really the collateral damage in Ukraine is still ongoing, not in Tokyo for quite some time.
So it's tragically possible that Ukraine could end up worse than Tokyo by the time hostilities finally cease.
Maybe with Tokyo a closer equivalent might be if Ukraine attacked Moscow using a comparable approach, with a degree of disregard for collateral damage figured in. Although Russian strategy already seems to target any part of Kiev that can be hit, civilian or not.
Plus no two things like this are really on the same scale and it's never a direct comparison, but there's some common undercurrent that is either predatory or vengeful which sometimes can grow until it can't get much worse.
So what about prehistoric tribes, even pre-humans, who surely had occasionally completely massacred victim tribes from time to time, not much differently than pack animals have always been known to do.
Total extermination like that could be rapidly completed with no weapons of mass destruction or even gunpowder.
Isn't there some possibility that this tendency has been retained evolutionarily or culturally to some extent today, even though most people would say that's just the opposite of "humanity".
Passed down in an unbroken chain in some way?
Disclaimer: when I was a teenager I worked one summer with a German machinist who had survived the bombing of Dresden. Ironically the project we were on was components for the the most advanced projectile of its caliber, yet to come. Both of us would have liked to build something else, but most opportunities across-the-board affecting all ages had already evaporated due to inflation of the 1970's, and the runaway years hadn't even gotten there yet.
See the big picture at the top? Clearly it's some kind of mall damaged by senseless and cruel Russian strike that cannot have any other purpose but to terrorize population of Kharkiv into submission.
The place should look familiar, only now you can see destroyed MLRS vehicle (there were two, but the second one got evaporated: https://t.me/aleksandr_skif/3150)
>Isn't there some possibility that this tendency has been retained evolutionarily or culturally to some extent today
Sure, but there is an opposite tendency too and it's not going anywhere barring catastrophic changes like famine due to global warming.
One thing that changed is that everything is instantly reported through numerous channels, and globally: traditional broadcast media as well as independent reporters using Internet channels.
The cynic in me finds this quite naive - there are territories of the Earth where if you are a male of a certain age and you are killed in a drone strike you are automatically classified as a "military-aged male", i.e. a non-civilian, regardless of the existence of any other evidence.
So what will happen if AI suggests someone is a terrorist when they are not? Well, in the worst scenario they'll be killed, and it'll be very close to what we have today, except somewhere in a private military database there might be an automatically generated record from an LLM ok-ing the target.
That's been done over a decade ago with random forests. There's no need to apply anything as advanced as generative AI 'hallucinations' for such a trivial problem. /s
I wrote (some) of a Lua interpreter in Rust, which runs PUC Lua 5.1 dumped bytecodes. Now I'm rewrite it all because I want to do type specialization: I'm implementing "Simple and Effective Type Check Removal through Lazy Basic Block Versioning", which is used as the basis of Shopify's YJIT for Ruby. I have a decent idea of how it needs to be implemented and have a prototype kinda working, although some of the jump targets are still wrong right now. Once I get it all sketched out then I'll have to scale it back up to a full Lua implementation again, and then I'll add table shape specialization afterwards (which is a bit trickier, and outlined in a sequel paper).
You can mount volumes in your Docker containers: what I usually do is mount my host source directory over the container's source directory, so that changes are immediately visible inside the container, and then hot reloading in e.g. Flask works exactly like you expect.
Years ago I had a similar train of thought: Zendesk is used by a ton of companies for their support site, and back then HTTPOnly cookies and javascript site isolation were much less of a thing. I found an XSS bug on Zendesk, which also translates into XSS on any site that used it as `support.fortune500.com` subdomain (which was a lot). You could then use it to exploit the main site, either by leaking user cookies or reading CSRF tokens from page contents because it was a subdomain.
Zendesk gave me a tshirt but not any money for it. C'est la vie.
It was one of the worst movies I've ever seen, IMO. It came across like a combination of a high school shakespeare parody and someone retelling The Fountainhead from memory but if they read it five years ago. The story was nonsensical, the message and themes were a muddy mess, the dialog was both stilted and overwrought. Basically the only positive thing I can say about the movie was the costume design was good. I didn't walk out of the theater because I thought I might as well watch to see how they finish up the story: that was a mistake. I should have walked out. You should not see Megalopolis.
Your comment is a bit snarky and I understand what you're saying but my understanding was that everyone thought Moore's Law was over when the computers stopped getting 2x faster every 18 months but that's not what Moore said. He said the transistor count would double amd it just so happened that that meant computers got a lot faster too because they benefited from all the xtra transistors. But GPUs do still get faster with more transistors because of the embarrassingly parallel nature of many matrix operations. So doubling the transistor count could still effectively significantly improve the speed and reduce time of AI training which is dominated by matrix math. Am I wrong? Did the transistor count doubling also slow down?
There's always someone out there claiming Moore's law to be dead. It never is. Chiplets and chip stacking in particular have enabled big increases in transistor count without associated downsides like decreased yield. So long as there is innovation and competition in the semiconductor industry, density / $, which you correctly point out was Moore's original metric, will continue to increase.
The potential for increasing performance and decreasing cost of LLM inference is even higher than Moore's law suggests, as the algorithms improve and are pushed closer to ideal hardware to run them.
This was always a strange point of contention - Intel denied using memristors. I assume there were some sort of patent or trademark issues.
WP:
"Development of 3D XPoint began around 2012.[8] Intel and Micron had developed other non-volatile phase-change memory (PCM) technologies previously;[note 1] Mark Durcan of Micron said 3D XPoint architecture differs from previous offerings of PCM, and uses chalcogenide materials for both selector and storage parts of the memory cell that are faster and more stable than traditional PCM materials like GST.[10] But today, it is thought of as a subset of ReRAM.[11] According to patents a variety of materials can be used as the chalcogenide material.[12][13][14]
3D XPoint has been stated to use electrical resistance and to be bit addressable.[15] Similarities to the resistive random-access memory under development by Crossbar Inc. have been noted, but 3D XPoint uses different storage physics.[8] Specifically, transistors are replaced by threshold switches as selectors in the memory cells.[16] 3D XPoint developers indicate that it is based on changes in resistance of the bulk material.[2] Intel CEO Brian Krzanich responded to ongoing questions on the XPoint material that the switching was based on "bulk material properties".[3] Intel has stated that 3D XPoint does not use a phase-change or memristor technology,[17] although this is disputed by independent reviewers.[18]
According to reverse engineering firm TechInsights, 3D XPoint uses germanium-antimony-tellurium (GST) with low silicon content as the data storage material which is accessed by ovonic threshold switches (OTSes)[19][20] made of ternary phased selenium-germanium-silicon with arsenic doping.[21][22]"
IIRC, performance was fantastic, but they were never able/willing to match the data density and data cost improvements in stacked-NAND flash, and without forcing themselves into the market at competitive rates, nobody wanted to write applications or design hardware suited to their unique strengths as low-latency caches.
There is still, to this day, a numerical niche for these drives, which is being served imperfectly by either normal TLC drives of very large size, SLC cache drives, or DRAM expansion cards connecting to the CPU through a PCIE bus. Just not at the prices they wanted to charge.
But wasn't the potentially transformative market intended to be "persistent DRAM" for instant-on devices removing the distinction between memory and storage, requiring DRAM-like speed rather than NAND-like speed ?
I recall their early R/W speed performance projections being far faster than what they ever achieved with Optane drives.
The products that used a PCIe X4 interface with a block storage protocol layered on top were never intended to deliver the best performance the memory was capable of.
Interesting - I wasn't aware, but even avoiding the PCI bus the performance must have been lacking as that link talks of "memory tiering". I guess this was "mid tier" somewhere between SSD and DRAM, which is a bit of a no-mans land especially in terms of conventional systems architecture ... really just a fast type of storage, or storage cache (a bit like a hybrid SSD-HDD drive).
reply