I don’t know if the spec supports that on its own. Although, it’s a good feature request.
You’d have to update the WG configuration each time a new IPv6 address connected. So, you would probably need to connect through something like a client that could push a config update and restart the service.
Not impossible, but that’s another layer of complexity to maintain.
Yeah. Almost all of the "killer apps" for LLMs all revolve around generating content, images, or videos. My question is always the same, "Is there really such a massive market for mediocre content?"
Yep. I think this is right on. The anthropomorphization in their behavior and problem descriptions is flawed.
It's precisely that analogy we learned early in our study of neural networks: the layers analyze the curves, straight segments, edges, size, shape, etc. But, when we look at the activation patterns, we see they are not doing anything remotely like that. They look like stochastic correlations, and the activation pattern was almost entirely random.
The same thing is happening here, but at incomprehensible scales and with fortunes being sunk into hope.
I feel like this fails on the premise that the models can be improved to the point where they are reliable. I don't know that holds true. It is extremely uncommon that making a system more complex makes it more reliable.
In the rare cases where more complexity produces a more reliable system, that complexity is always incremental, not sudden.
With our current approach to deep neural networks and LLMs, we missed the incremental step and jumped to rodent brain levels of complexity. Now, we are hoping that we can improve our way to stability.
I don't know of any examples where that has happened - so I am not optimistic about the chances here.
My graduate research was in this area. My lab group developed swarm robots for various terrestrial and space exploration tasks. I spent a lot of time probing why our swarm robots developed pathological behavioral breakdowns - running away from construction projects, burying each other, etc... The issue was so fundamental to our machine learning methods that we never found a way to reliably address it—by the time I left, anyway. No matter how we reconfigured the neural networks, trained, punished, deprived, or implemented forced forgetting or fine-tuned, nothing seemed to eliminate the catastrophic behavioral edge cases—nothing except for dramatically simplifying the neural networks.
Once I started seeing these behaviors in our robots, their appearance became much more pronounced every time I dug deeply into proposed ML systems: autonomous vehicles, robotic assistants, chatbots, and LLMs.
As I've had time to reflect on our challenges, I think that neural networks very quickly tend to overfit, and deep neural networks are incomparably overfitted. That condition makes them sensitive to hidden attractors that cause the system to break down when it is near these areas - catastrophically.
How do we define "near"? That would have to be determined using some topological method. But these systems are so complicated that we can't analyze their networks' topology or even brute-force probe their activations. Further, the larger, deeper, and more highly connected the network, the more challenging these hidden attractors are to find.
I was bothered by this topic a decade ago, and nothing I have seen today has alleviated my concern. We are building larger, deeper, and more connected networks on the premise that we'll eventually get to a state so unimaginably overfitted that it becomes stable again. I am unnerved by this idea and by the amount of money flowing in that direction with reckless abandon.
I’m not aware of any specific films we were in. We filmed a lot of our robotics trials for various collaborations, but no documentaries while I was there. Shortly after I left, my team got some acclaim for their Lunar Ark project (which is really cool). But, I had been out for a couple of years by that point. If they filmed a documentary, it likely would have been for that project.
What are you talking about? Where was America getting their opioids before The Sacklers? Go ahead and pull a chart of heroin deaths vs. years and draw a vertical bar on where Purdue launched their widespread marketing misinformation and disinformation.
This is the largest and most deadly episode of addictiveness since the tobacco companies were marketing cigarettes as healthy. If this were a fire, you’d be arguing:
“Look, America has always had fires. Sure Purdue started fires in all 50 states that spread into the forest, but they were small when Purdue started them. Now, there is a blaze that has spread uncontrolled across the Continental US. So, obviously there were other parallel organizations starting fires.”
No. That doesn’t follow. Purdue stared the fires. They added tinder with marketing misinformation, then they inhibited any attempt at controlling it with active disinformation campaigns to confuse doctors and regulators about the root cause. Now, it’s a wild uncontrolled inferno, but they started the first fires.
Didn't we also invade the religious nutsos stopping Afghanistan from producing poppies and then prop up a fake regime that let it flourish, massively increasing the supply of natural opiates? If I recall, the timing works out pretty good with illegal opiate use ramping up.
We sure did. About 60-70% of all of the world's opium was sourced from Afghanistan until 2022. That supply was used to make actual prescriptions in non-OECD countries - like India & China - as well as sold to make heroin in the US, EU, Asia, and Africa.
But be careful about confusing cause & effect. The cause of the Afghani farmers growing opium was the insatiable demand for opium that was developed through legal means. Then, when countries like the US started to restrict access to opiates, addicts and patients alike started to seek alternatives to the scarce prescriptions.
More to the point, why didn't patients use something else? Because there is nothing. Purdue's misinformation actively discouraged the discovery of new non-addictive compounds for decades. No one was looking, and the R&D pipeline ran dry.
By the time enough doctors sounded enough alarms to cause a change in the late 2010s, research in non-addictive pain management solutions was decades behind. No one engaged in it because there was no need for it. Now that we realize it was all a lie and these drugs have killed millions, there are no alternatives. Very few potential compounds are even in Phase 2 clinical trials right now, let alone the half dozen that would be needed in Phase 3 to ensure we have a single alternate choice for pain management in the next five years.
So, the outlook looks bleak. Today, in 2024, we don't have good ways to manage pain that is non-addictive, and certainly no good way to reach the millions who are hopelessly addicted to opioids for pain management. But, very little of this is the patient's fault - and a lot of this is directly related to the monumental efforts of Purdue to misinform for profit.
That’s gotta be the most ignorant opinion I’ve heard on the subject. This clearly is coming from someone who has had the benefit of good health, or hasn’t had the misfortune of a critical ER visit.
When I got addicted to opioids, the doctors didn’t ask me what I wanted in the ER. I got Dilauded as they prepped the OR. I wasn’t lucid enough or cognizant enough to give informed consent. During my recovery, there wasn’t an option of alternate anything - I got a lot of morphine and then oxy. I was discharged 4 weeks later from the hospital with a 3 month prescription for 120mg of opioids per day. That would kill most people.
My story is not unique. It is so common it could be a troupe. The doctors saved my life in one way, then discharged me into a hell of addiction that it took 9 months of serious determined effort to overcome.
Then, some uninformed keyboard warrior like yourself comes along and blames people in my situation? Maybe you should have read a little less Opioid Wars and a little more Current Events. You clearly know nothing about this problem.
I've worked in the ER, held a license to work there, and spend everyday with licensed providers in addiction medicine. I assure you I have more familiarity here than even most addicts.
But you win the victim totempole height contest I guess.
That doesn’t clarify anything about your knowledge. The specialized hospital cleaning crews are in the ER every day and dozens of hours per week in the OR, but that doesn’t allow them to practice medicine or perform surgery. Proximity is not the same as practice.
The medical practice (until a few years ago) was profoundly and intentionally misinformed about the addictiveness of Oxycodone and Oxycontin. You brought up a question about why people didn't learn from the Opioid Wars, but your flippant question could just as easily be applied to the MDs who wrote the prescriptions.
Medical doctors are smart; they MUST know about the Opioid Wars and the horror the opium wrecked on Chinese society for centuries. Why would they prescribe such a dangerously addictive compound to injured and vulnerable people having the worst day of their lives?
Instead, you blame patients with little choice (or capacity to evaluate choices) in a terrible situation. And, if you talked to any addicts, you'd sure find that the top of the "victim totempole" is actually really crowded with people who got there the exact same way. That is no coincidence.
The fact of the matter is that Purdue intentionally misinformed everyone - doctors, patients, caregivers, and family. They used marketing, propaganda, bribes, recognition, and myriad other tools to convince everyone that their product was different. Doctors, knowing full well about the Opium Wars, wholeheartedly believed the propaganda. I wasn't the first patient to be discharged from the hospital with a multi-month supply of opioids. This was a common practice.
The worst part is for someone who claims to be around addicts, you think I am somehow unique in my story. My story is so mundanely typical that it should be nauseating. Medical treatment is, by far, the most common way addicts start with opioids. 75% of all heroin users whose addictions started in the 2000s reported that it began with prescriptions from their doctors.
I genuinely suggest you spend some time talking to the addicts around you rather than judging them. You'd quickly come to regret the flippancy of your statement and probably have some empathy for their situation. These were regular people with jobs, families, lives, hopes, and dreams. In a crisis, they entrusted medical doctors to make the best medical decision for them, and the price they paid was their future.
That’s a hilarious directive since their homegrown CPU is 900 MHz. They are on something like a 30nm die. There is just no way they can transition to those chips in modern devices. This sounds like exactly the sort of order that would come from Xi. He’s living in his information bubble and is excited to go to war.
No, they are not that far behind. You're thinking of Longsoon I guess, but they have a 7nm process at SMIC and Hisilicon designing ARM-based server processors on it[1]. Sure, ARM is western, but guess what, they have the IP anyway, as they required ARM to do business through a subsidiary in China.
China is serious about semiconductors. Hisilicon is competitive in many markets.
It's also a "so what" situation. If China's economy is severed from the rest of the world's, competitors can't take its market share. Being five or even ten years behind on chips is irrelevant when your consumers can't go elsewhere. What did people use their smartphones for ten years ago? Recording themselves and watching cat videos. What do they do now? You don't need 2nm chips to do anything except take market share from the company using 3nm chips. The F-22 Raptor uses an i960 processor from 1988.
Do checkout Dxomark which smartphone has best camera. Do checkout how well Kunlun vs Gorrilla best. Do check the number of Mate 60 (hint iPhone sales dropped 20%). And maybe also check the cpu size. They have already rumours seeing actual 2-3nm in used in Huawei engineering sample unit. Even AI GPUs ES sample now floating if you have connections to get it (definitely better than Intel Arcs). I think you truly live in Joe info bubble. I have friends directly in Chinese semicon. You will see 4-5 years down the road, TSMC become irrelevant. Did you know now 90% worldwide mid tier and low tier chip dominated by Chinese made? Used to be TI and EU foundry during Obama time. You can do the research. Helps a lot if you are strong in read Chinese simplified characters sources.
If you read Sun Tzu, theft of info is warranted and morally right to do so. Read the spycraft section of Sun Tzu. US did the same with Sputnik. They path and legitimize the act decades ago.
Nope. Loongson is building on 14/12nm nodes and competing on performance with first-gen Ryzen and Coffee Lake. Zhaoxin is still on 16nm AFAIK, but they're matching modern architectures on IPC. CXMT is making LPDDR5 dies in meaningful quantities. Underestimate Chinese industry at your peril, particularly when they've got a blank cheque from the Party.
I run the ML team at a larger unicorn. I fully agree with this. We are multi & hybrid cloud with global SD Wan and everything on CI/CD. There is still a ton of maintenance, patches, deployments, and updates that require us to manually intervene in our otherwise automated systems.
If OP were to learn sys admin skills, CI/CD, and Terraform, he would be a key candidate for any ML team globally. They’re all going in that direction. We work closely with Google and they coached us in that direction. Sys Admin skills will always be in demand.
You’d have to update the WG configuration each time a new IPv6 address connected. So, you would probably need to connect through something like a client that could push a config update and restart the service.
Not impossible, but that’s another layer of complexity to maintain.