Tl;dr: think hard about what you’d use this much RAM for in a desktop setting and do some research about how people like the 395 for your use case. Not all use cases work well.
Anyone who thinks they are going to serve some 100+ GB LLM locally, remember that memory bandwidth becomes a key limitation for large models. While you might be able to load a model, token generation can be very slow. MoE models like Qwen3.5-122B-A10B work decently fast, but dense models of a decent size are slow and you won’t want to use them.
I’ve got a 395 system, and found that I’m quite happy with Qwen3.6-35B-A3B, generating at around 50 t/s, but the dense 27B model is 20-25 t/s and that’s the lower limit I’m willing to tolerate. So a 70b dense model is just not going to happen. That means you can’t really use that much RAM.
A reasonable use case is to have multiple smaller models loaded- you can have an image generation model loaded along with the text model. Or you can use this computer for development simultaneously with serving LLMs. Those ideas work okay. But trying to load up a single giant model is going to test your patience.
Feels like they released this to ride the wave of press of GPT-5.6, Kimi K3, and Qwen 3.8. Doesn't feel like Google has much substance with this post except a bump in version and tweaked their pricing.
> Somehow I have to incentivize students to learn the material on their own... And a really effective way to do that is a graded test ...
I know. The desire to get a degree or the "threat" of wasted tuition and disappointed family members is really what gets most students to crack a book. Most people aren't self-motivated to learn. If people were self-motivated to learn the material, as opposed to getting a credential or some other reason for being in college, I wouldn't need to worry about cheaters.
I have a few thoughts related to this, and maybe I can get them out and ties them back together at the end.
1a. Yes, college isn’t right for everyone, and testing the traditional way with paper and pencil certainly disadvantages some students who would be star performers in a real world setting but are not a good fit for college classes. Think “Good Will Hunting” type people. They do exist.
1b. However, there are certainly more people who only imagine themselves as Will Hunting type people and that they are just too smart for college, but the reality is that they are dumb, or didn’t learn the material. For every person who fails a test because they’re a genius who is a bad fit in the system, there are at least 10 idiots who imagine themselves geniuses, and they would have passed the class if only that PhD professor with all his book smarts had actually written the right kind of exam. Traditional schooling and testing doesn’t work for the extreme upper tail of intelligence, but it exists because it did quite well at educating and sorting the masses to support the Industrial Revolution.
1c. If college were only about the knowledge, you could learn most of it with internet access and a library card for much cheaper. The vast majority of people are in college for the credential, and the institution has to protect the signal of the credential.
2a. Most college assignments and exams are not a good reflection of full time employment. College credentials serve mainly as a networking aid and a signal to employers that you are compliant and competent enough to follow a professor’s instructions, and will likely be a compliant and competent employee, although the instructions might be different. If it is the AI who is the one who followed the professor’s instructions you water the signal down, and employers don’t want that. It is irrelevant that a student can copy and paste things into ChatGPT, and that ChatGPT can get the answer right on this test. That isn’t what college is supposed to signal.
2b. A problem is that I literally can’t write a test in most subjects now that I would expect a student to complete that can’t be completed by ChatGPT better and faster. I teach undergraduate math and a while ago we thought that since GPT-4o could get a C in calculus that we’d just raise the standard. Now Fable and GPT-5.5 can cruise to an A in literally every math course in our catalog, and they can also catch every tiny issue in an exam written by a human. But I have to teach these undergraduate subjects so that some students can go on to PhD studies so they can contribute to the field. If we just stop teaching undergraduate subjects then PhD production and novel research grinds to a halt and only a few fields will progress where an AI is capable of self improvement.
2c. I’ve seen that my best students know how to do do something by hand and use a computer to complement/increase their capabilities, not to cover over entire gaps. When you have literally zero skill in an area, you can’t spot when you got a totally bad output from the AI (these days usually because of a lack of context or bad prompting because the student didn’t understand the material, not because the AI wasn’t capable). Somehow I have to incentivize students to learn the material on their own, so that they can be a better user of AI in the future. And a really effective way to do that is a graded test in which AI is not available to help them.
So to try to tie it together, is that there is still some value to a college degree, at least until something better comes along. But that college degree is only useful when it is a signal about the person and not some other tool. And although AI is getting very capable, somehow we have to teach the lower stuff to build up to the higher stuff, so we do need to restrict the use of AI in some educational settings so that we can build a good foundation for future learning.
As to the point about old paper sources being considered more reliable than an internet site, I agree. I am of the generation that wasn’t allowed to cite Wikipedia and it frustrated me. We’ll eventually figure out how to permit proper AI use, much like how many professors now allow you to use Wikipedia to start researching a topic.
Additionally, I thought the threat being modeled for “biology” was stuff like bioterroism- how to make anthrax, how to distribute a toxin, etc.
I don’t feel like calculating results for a trial is really in the threat model unless we think a terrorist is out there testing the efficacy of their anthrax before using it in an attack.
OCR means optical character recognition. The terms do not require a direct transcription, but that is mostly what OCR meant in the past. If you’re using an LLM’s vision capability to pass in text and the LLM actually understands it, then I would say that it recognized the characters, hence OCR seems okay to use.
I feel like CRAN should be used for packages that are expressly made for others to use, and with effort put in to the documentation and vignettes.
If you’re making a package for a small team or aren’t pushing it to a large audience then just keep it on a GitHub repository. It is almost as easy to install from GitHub with devtools as it is to install.packages().
This is a great example of why prompt engineering is still relevant. Without providing definitions and examples and a well defined rubric, you’re going to see different models disagree by a level in either direction. When you get more prescriptive the models tend to agree better.
I’ve experimented with AI grading for undergraduate math courses, and see basically the same thing. If you just tell the AI “grade this problem and assign a letter grade” then I’ve only seen about 30% agreement between a human assigned grade and the AI assigned grade. But over 75% agreement if you say a “match” is within one letter grade. And to get better agreement you have to spend a lot more time on the rubric- what kinds of mistakes are a big deal, what kinds of mistakes are not a big deal, how much work is required to be shown to get credit, a couple examples of each letter grade. Once you have done that, the AI gets a lot better agreement with human graders, but it is hard to know when you’ve given enough guidance for a problem.
I looked at it and am unimpressed. I’ll take any Pininfarina designed Ferrari over this plain looking thing. Jony Ive did an okay job on the interior but the outside is just plain. The outside looks closer to an Amazon delivery van than a super car.
Sure it’s fast, but a Corvette ZR1X is faster. I’d rather take a ZR1X to a custom shop and have them redo the atrocious Corvette interior.
Edit: I’ll acknowledge that I’m not the kind of person to buy a Ferrari even if I could afford it, so maybe Ferrari doesn’t care about my opinion, but I feel like Jony Ive pulled an “emperor’s new clothes” on the Ferrari execs.
I did read the article and I’m not arguing against a straw man. If you’re going to let an AI agent do everything for you then go ahead and use Rust (or any language with a strong type system that benefits agents).
But if I’m participating then I’m going to use Python because it’s easier to read.
If there’s anything that I’m arguing against is the author’s claim that the ecosystem of libraries (regardless of whether they are a wrapper) and readability don’t matter anymore. I’d say that in a lot of smaller teams it still matters. We’re not all using AI to ship slop. A lot of us are using AI to work on our ideas for our hobbies or for research. And it’s not fulfilling unless I get to be involved in the process.
It’s funny that in your reply “this article is almost certainly intended to be read by humans” you made what is the best case to keep writing code in Python even with AI.
Sure, if you are going to have an AI do all your coding and maintenance you can use whatever language it’s best at. But if you want to participate in the writing, debugging, and maintenance, it has to be in a language that a human can read. I’m not saying that Rust or Go is unreadable, but I know I am better at Python personally and am going to keep using it until the speed penalty matters to my project, and then maybe I’ll let an AI rewrite the whole thing in a faster language.
If we're shipping the alfalfa to China, I assume that means it's supporting some Chinese person's food source, whether they are directly eating the alfalfa, or some animal is eating it that later becomes food.
If someone is flooding a field unproductively just to use up their quota of water, that is a bad thing that should be addressed. But even if you excluded that unproductive usage and compared AI water use to legitimate agriculture use, that would still be an unfair comparison. If you were to compare AI water use to the amount of water that people are wasting just for legal reasons, then I honestly think that would be a pretty apt comparison.
Comparing water usage of AI to agriculture and cities is a little misleading. The cities' water usage is to keep people alive with basically mandatory things, like hygiene, and drinking. Agricultural water usage is required because we have to eat to live. Don't compare something optional to something mandatory.
Instead, compare AI water usage to that of optional things in a city, such as car washes and water parks. Or compare AI water usage to that of what it would take a human to do a comparable task (what does it take to keep a human alive for a few hours compared to running a 15 minute long task to write a report with AI?). While AI water usage might still not look that bad, it would be a more honest comparison.
> being able to run a model and being able to run a model fast are two very different thresholds
Specifically speaking, on my Strix Halo machine with (theoretical) memory bandwidth of 256 GB/s, a 70 GB model can't generate faster than 256/70= 3.65 t/s. The logic here is that a dense model must do a full read of the weights for each token. So even if the GPU can keep up, the memory bandwidth is limiting.
A Mac M5 Pro is faster with a bandwidth of 307 GB/s, but that's only a little faster.
This thing is going to be slow on consumer hardware. Maybe that is useful for someone, but I probably prefer a faster model in most cases even if the model isn't quite as smart. Qwen3.6 35B-A3B generates about 50 t/s on my machine, so it can make mistakes, be corrected, and try again in the same time that this model would still be thinking about its first response.
For a while it was all about getting the lightest shoes, because picking up heavy shoes slowed you down. Then the energy return (pebax foam, carbon plates/rods) became the main focus because the weight didn't matter as much when the shoe was literally springy. Surely this is now going to spark a race for the optimal balance between weight and energy return.
Intelligence agencies and companies for which industrial espionage is an actual concern will re-encrypt their data storage, or have already done so. The only risk is on data that was already obtained with a vulnerable encryption. So there is some risk that a few secrets are lost, but it won’t be everything. And if you were to start now and quantum decryption isn’t viable for a decade then any secrets that do get exposed are surely less of a problem than if they were discovered today.
Anyone who thinks they are going to serve some 100+ GB LLM locally, remember that memory bandwidth becomes a key limitation for large models. While you might be able to load a model, token generation can be very slow. MoE models like Qwen3.5-122B-A10B work decently fast, but dense models of a decent size are slow and you won’t want to use them.
I’ve got a 395 system, and found that I’m quite happy with Qwen3.6-35B-A3B, generating at around 50 t/s, but the dense 27B model is 20-25 t/s and that’s the lower limit I’m willing to tolerate. So a 70b dense model is just not going to happen. That means you can’t really use that much RAM.
A reasonable use case is to have multiple smaller models loaded- you can have an image generation model loaded along with the text model. Or you can use this computer for development simultaneously with serving LLMs. Those ideas work okay. But trying to load up a single giant model is going to test your patience.