4,679 karma · joined July 29, 2010
But today's AI might do better than the average tech writer. AI might be able to generate reasonably usable, if mediocre, technical documentation based on a halfheartedly updated wiki and the README files and comments scattered in the developers' code base. A lot of projects don't just have poor technical documentation, they have no technical documentation.
America has always been at its best when it lives up to its ideals, and at its worst when it discards those ideals. America has often been in the wrong, but on balance, the world has been better off for having a great champion of liberal democracy.
With Trump, it's not a question of believing things that are right or wrong. For the post-truth mindset, right and wrong don't matter. Democracy doesn't matter. There is only power. The second Trump presidency is the first time in modern history that America is no longer a great nation.
Proponents of what we now call "open source" wanted to distinguish between two senses of the word "free". One sense is not having to pay for something, as in "Come over to my party, the beer is free." Anther sense is "I can criticize the government, because the country I live in is free." People in the free software and open source movement began to phrase the dichotomy in these terms to illustrate how one sense of the word "free" is much more important than the other. The fact that you don't have to pay for some piece of software is nice, but what's more important is that you aren't beholden to the company that developed it.
In the specific case of these chips, the offshoring was to Taiwan, a democracy, and one of our geopolitically closest allies. This business went to TSMC because they did a better job producing these chips than either Intel or AMD (or AMD's successor, Global Foundries). The Twinscan machines that fab the chips are from ASML in the Netherlands, another democracy and geopolitical ally. The parts for those machines come from a truly global supply chain. Modern chip manufacturing is an absolute marvel of global integration.
American consumers and businesses benefited from this arrangement in the form of very competitively priced hardware. Most of the manufacturing was in Asia, but most of the profits were to American companies. This is why America is today the richest country on Earth. It simply would not have happened like it did without free trade.
China's low wages and integrated supply chain can beat American manufacturing. China cannot beat the combined talent of the free world. Along with immigration, from the likes of Taiwanese-American Jensen Huang, free trade and an open economy was what once made America great.
Insofar as there is any plan, the current officeholder's priorities are to project the appearance of personal power on television. If you're wondering what's going on strategically, don't go thinking that there's some grand plan, or even an intention to benefit the United States in the long term. There are some people in the cabinet who are thinking long term, but that's not universal, and that's not what they're selected for. Every action that is taken is to satisfy the president's narcissism and ego in the present moment. You have to understand the "US plan" in this light for anything coming out of the executive office to make sense.
As far as trusting AI, I presume your mother was asking ChatGPT, not Llama 7B or something. The LLM backed up your reasoning rather than telling her that dog feces in bushes is harmless isn't just happenstance, it's because the big frontier commercial models really do know a lot.
That isn't to say the LLMs know everything, or that they're right all the time, but they tend to be more right than wrong. I wouldn't trust an LLM for medical advice over, say, a doctor, or for electrical advice over an electrician. But I'd absolutely trust ChatGPT or Claude for medical advice over an electrician, or for electrical advice over a medical doctor.
But to bring the point back to the article, we might currently be living in a brief period where these big corporate AIs can be reasonably trusted. Google's Gemeni is absolutely going to become ad driven, and OpenAI seems on the path to following the same direction. Xai's Grok is already practicing Elon-thought. Not only will the models show ads, but they'll be trained to tell their users what they want to hear because humans love confirmation bias. Future models may well tell your mother that dog feces can safely be thrown in bushes, if that's the answer that will make her likelier to come back and see some ads next time.
Netflix's revenue is subscription based long form videos. That should be a viable business model. Instead, they seem to be willfully heading in the direction of serving slop to an audience that's not fully engaged. This road leads them into direct competition with YouTube and TikTok.
When my son moved out of the house recently, he went on the additional household plan. What struck me was that the user interface steered heavily toward the option that includes ads. We had to search for the small print to let him pay Netflix a few dollars more for the non-ads version.
In the late 90's and early 2000's there was a sense that ads were a reasonable tradeoff for free services on the internet. If the last 20 years have taught us anything, it's that the perverse incentives are a catastrophe.
In unconstrained capitalism, a small group of powerful individuals use corporate control of government for their personal benefit.
Clearly, these outcomes are completely different.
For instance, I like the bleakness of Alien 3 opening with Newt and Hicks both dead. That doesn't spoil my enjoyment of Aliens, which ends on a triumphant note. These are different stories, and they can be treated on completely different planes. If you want, you can imagine the movies as representing alternate branching universes, where one branch led to Newt and Hicks dying in hibernation, and in some other branch that's too uninteresting to be put to film, they live happily ever after.
I also liked Blade Runner 2049, but I don't need to retroactively reevaluate the original Blade Runner in light of any of the questions that are settled in the sequel. In Ridley Scott's original film, Deckard's humanity is still open to question, regardless of what's presented in Villeneuve's version.
Of course when the sequel is complete trash, it's easy to ignore entirely. Terminator 3 being the obvious example.
Porn, obviously, though if you look at what's popular on civitai.com, a lot of it isn't photo-realistic. That might change as photo-realistic models are fully out of the uncanny valley.
Presumably personalized advertising, but this isn't something we've seen much of yet. Maybe this is about to explode into the mainstream.
Perhaps stock-photo type images for generic non-personalized advertising? This seems like a market with a lot of reach, but not much depth.
There might be demand for photos of family vacations that didn't actually happen, or removing erstwhile in-laws from family photos after a divorce. That all seems a bit creepy.
I could see some useful applications in education, like "Draw a picture to help me understand the role of RNA." But those don't need to be photo-realistic.
I'm sure people will come up with more and better uses for AI-generated images, but it's not obvious to me there will be more demand for images that are photo-realistic, rather than images that look like illustrations.
Everyone knows that these LLMs were trained on copyrighted material, and as a next-token prediction model, LLMs are strongly inclined to reproduce text they were trained on.
I have no idea that "patriarchal standards" are, but we absolutely should value people based on what they perform and do. That's not to say that people don't have an inherent value, or that we shouldn't care for people who need help. We value people who help those around them to the best of their abilities, and we denigrate those who use their abilities to hurt others. Regardless of gender.
> We must reclaim masculinity and not allow it to be synonymous with domination or the will to do violence.
There is an extremist element that roughly hews to this notion of masculinity, and this is something that Galloway warns against. I don't think that definition of masculinity that was ever mainstream. It's always been on the fringe, and rightly so.
For actual real work, I use Claude.
If you want to use an open weights model to get real work done, the sensible thing would be to rent a GPU in the cloud. I'd be inclined to run llama.cpp because I know it well enough, but vLLM would make more sense for models that runs entirely on the GPU.
This would be a terrifyingly dystopian outcome. Whoever owns this super intelligence is not going to use it for the good of humanity, they're going to use it for personal enrichment. Sam Altman says OpenAI will cure cancer, but in practice they're rolling out porn. There's more immediate profit to be made from preying on loneliness and delusion than there is from empowering everyone. If you doubt the other CEOs would do the same, just look at them kissing the ass of America's wannabe dictator in the White House.
Another possible outcome is that no single model or company wins the AI race. Consumers will choose the AI models that best suit their varying needs, and suppliers will compete on pricing and capability in a competitive free market. In this future, the winners will be companies and individuals who make best use of AI to provide value. This wouldn't justify the valuations of the largest AI companies, and it's absolutely not the future that they want.
In my experience, the teams at AWS are pretty diverse, reflecting the diversity in the area. Even if a lot of the Indian employees are taking the day off, there should be plenty of other employees to back them up. A culturally diverse employee base should mitigate against this sort of problem.
If it does turn out that the outage was prolonged due to one or two key engineers being unreachable for the holiday, that's an indictment of AWS for allowing these single points of failure to occur, not for hiring Indians.
There is the ChatGPT product, operated by OpenAI, Inc, which you can access via their web site or their API. OpenAI does publish gpt-oss as an open-weights model. I suppose you could argue that gpt-oss is "a ChatGPT," though I'd normally think of it as "a large language model." Much like Claude, DeepSeek, Qwen and so on are other large language models.
But clearly we also see some major downsides. We already have an epidemic of social media rotting people's minds, and everything about this capability is set to supercharge these trends. OpenAI addresses some of these concerns, but there's absolutely no reason to think that OpenAI will do anything other than what they perceive as whatever makes them the most money.
An analogy would be a company coming up with a way to synthesize and distribute infinite high-fructose corn syrup. There are positive aspects to cheaply making sweet tasting food, but we can also expect some very adverse effects on nutritional health. Sora looks like the equivalent for the mind.
There's an optimistic take on this fantastic new technology making the world a better place for all of us in the long run, after society and culture have adapted to it. It's going to be a bumpy ride before we get there.
In the medium to longer term, we might be in a situation where only the most powerful next-generation AI models are able to make sense of giant vibe-coded balls of spaghetti and mud we're about to saddle ourselves with.
When you run llama.cpp on your home computer, it holds onto the key-value cache from previous runs in memory. Presumably Claude does something analogous, though on a much larger scale. Maybe Claude holds onto that key-value cache indefinitely, but my naive expectation would be that it only holds onto it for however long it expects you to keep the context going. If you walk away from your computer and resume the context the next day, I'd expect Claude to re-read your entire context all over again.
At best, you're getting some performance benefit keeping this context going, but you are subjecting yourself to context rot.
Someone familiar with running Claude or industrial-strength SOTA models might have more insight.
You probably don't need fp16. Most models can be quantized down to q8 with minimal loss of quality. Models can usually be quantized to q4 or even below and run reasonably well, depending on what you expect out of them.
Even at q8, you'll need around 235GB of memory. An Nvidia RTX 5090 has 32GB of VRAM and has an official price of about $2000, but usually retails for more. If you can find them at that price, you'd need eight of them to run a 235GB model entirely in VRAM, and that doesn't include a motherboard and CPU that can handle eight GPUs. You could look for old mining rigs built from RTX 3090s or P40s. Otherwise, I don't see much prospect for fitting this much data into VRAM on consumer GPUs for under $10k.
Without NVLink, you're going to take a massive performance hit running a model distributed over several computers. It can be done, and there's research into optimizing distributed models, but the throughput is a significant bottleneck. For now, you really want to run on a single machine.
You can get pretty good performance out of a CPU. The key is memory bandwidth. Look at server or workstation class CPUs with a lot of DDR5 memory channels that support a high MT/s rate. For example, an AMD Ryzen Threadripper 7965WX has eight DDR5 memory channels at up to 5200 MT/s and retails for about $2500. Depending on your needs, this might give you acceptable performance.
Lastly, I'd question whether you really need to run this at home. Obviously, this depends on your situation and what you need it for. Any investment you put into hardware is going to depreciate significantly in just a few years. $10k of credits in the cloud will take you a long way.
One of the observations is how much difference memory speed and bandwidth makes, even for CPU inference. Obviously a CPU isn't going to match a GPU for inference speed, but it's an affordable way to run much larger models than you can fit in 24GB or even 48GB of VRAM. If you do run inference on a CPU, you might benefit from some of the same memory optimizations made by gamers: favoring low-latency overclocked RAM.
If you're passing data between processes, and you still want the data to be human readable, then JSON is a good choice.
If you're writing a configuration file that's going to be edited by a human, then YAML is easier to look at and understand.
Surely, some of these people blocked development and did everything they could to boost their personal net worth at the expense of younger generations. Other people did the reverse. Most people just went with the flow and didn't bother to vote either way.