That's what they said about cars at first. Or credit cards. The question to ask is: will the world we make in the wake of this invention afford us to live without it? And if the answer is no, then it's all the more important to have access to truly free and uncensored AIs. How did we learn things before AI? We googled them. How's that working out in the age of AI? AI both poisons our search results and gets integrated with them. There's large interests in making sure everything we see hear and think is prevetted by some approved AI. That's not a future I want to live in, but the signs are there.
I've often found it helpful in search. Specifically, when the topic is well-documented, you can provide a clear description, but you're lacking the right words or terminology. Then it can help in finding the right question to ask, if not answering it. Recall when we used to laugh at people typing in literal questions into the Google search bar? Those are the exact types of queries that the LLM is equipped to answer. As for the "improvements" in GPT 5.1, seems to me like another case of pushing Clippy on people who want Anton.
https://www.latent.space/p/clippy-v-anton
The cynic in me thinks this is just a means to eventually make more money by offering paid unrestricted versions to medical and legal professionals. I'm well-aware that it's not a truth machine, and any output it provides should be verified, checked for references, and treated with due diligence. Yet the same goes for just about any internet search. I don't think some people not knowing how to use it warrants restricting its functionality for the rest of us.
Nah, this is to avoid litigation. Who needs lawsuits when you are seeking profit? 1 loss of a major lawsuit is horrible, there's the case of folks suing them because their loved ones committed suicide after chatting with ChatGPT. They are doing everything to avoid getting dragged to court.
> I'm well-aware that it's not a truth machine, and any output it provides should be verified, checked for references, and treated with due diligence.
You are, but that's not how AI is being marketed by OpenAI, Google, etc. They never mention, in their ads, how much the output needs to be double and triple checked. They say "AI can do what you want! It knows all! It's smarter than PhDs!". Search engines don't say "And this is the truth" in their results, which is not what LLM hypers do.
I appreciate how the newer versions provide more links and references. It makes the task of verifying it (or at least where it got its results from) that much easier. What you're describing seems more like a advertisement problem, not a product problem. No matter how many locks and restrictions they put on it, someone, somwhere, will still find a way to get hurt from its advice. A hammer that's hard enough to beat nails is hard enough to bruise your fingers.
If they do that, they will be subject of regulations on medical devices. As they should be and means the end result will be less likely to promote complete crap then it is now.
I think one of the challenges is attribution. For example if you use Google search to create a fraudulent legal filing there aren't any of Google's fingerprints on the document. It gets reported as malpractice. Whereas reporting on these things is OpenAI or whatever AI is responsible. So even from the perspective of protecting a brand it's unavoidable. Suppose (no idea if true) the Louvre robbers wore Nike shoes and the reporting were that Nike shoes were used to rob the Louvre and all anyone talks about is Nike and how people need to be careful about what they do wearing Nike shoes.
It's like newsrooms took the advice that passive voice is bad form so they inject OpenAI as the subject instead.
That's why I don't install updates, unless and until they've been proven not to break things. I miss the old days when software was expected to work out of the box and updates, on the rare occasions when they appeared, were actually useful.
I hope you are speaking with tongue in cheek. Security is the main reason to keep current with updates. They address various “CVE” reports and go beyond to patch things not reported by CVEs.
I think users wouldn't be so resistant to security updates of they were just that and not bundled with feature removal, unwanted new features, and other things.
Or if they were properly done. Example: Intel and the plundervolt vulnerability. To fix that they removed the ability for undervolting in ny laptop. If I don't use SGX there's no reason for the block. They could've restricted undervolting only when SGX is enabled but no, they had to "fix" it in the worst way possible.
CVE inflation is real. Most CVEs are of very low quality.
Anyway, security updates should be decoupled from feature updates, so that people aren't hesitant to update. Otherwise, you get people who hold out because they're worried the new release is going to break all their settings and "opt-in" into all kinds of new telemetry.
> Security is the main reason to keep current with updates.
It shouldn't be that way though. Especially the billion dollar corporations should not be excused for shipping insecure software - the sad reality though is that Microsoft seems to have lost most of its QA team and what remains of its dev team gets shifted to developing adware for that sweet sweet "recurring revenue" nectar. Apple doesn't have that problem at least, but their management also has massive problems, prioritizing shiny new gadgets over fixing the tons of bugs people have.
Fair point! A single complex multiplication `(a+bi)(c+di)` indeed requires at least 3 real multiplications to be implemented.
However, when researchers (and systems like AlphaEvolve in this context) analyze fast matrix multiplication algorithms like Strassen's, the primary goal is usually to improve the asymptotic complexity (and understand the space of these algorithms better). This complexity is determined by the number of multiplications in the field over which the matrices are defined.
* For real matrices, we count real scalar multiplications.
* For complex-valued matrices (as in the 4x4 example where AlphaEvolve found a solution with 48 scalar multiplications), "scalar multiplication" refers to a complex scalar multiplication.
The key is that these are the operations you recurse on. Additions, or the constant factor cost of implementing one field's multiplication, don't change the exponent in the `N^(log_base(multiplications))` complexity. They are constant factors.
Of course, for practical performance on a specific 4x4 matrix, one would absolutely dive into the real operations, additions, memory layout, etc., but making 4x4 matrix multiplication practically faster on some particular hardware was not the focus in this section. (We do improve practical implementation of large matrix multiplications on the target hardware in the “Enhancing AI training and inference” section of the blog post.)
I think what gives Tao the title is how multidisciplinary he is. He can wander in to a new subfield of mathematics and start making SOTA contributions after not very much time, which is a rare thing.
I'm more interested in the technical details than the publicity. Pretty much anyone these days can learn what a diffusion model is, how they're implemented, what the control flow is. What about this new multimodal LLM? They have no problems with text, they generate images using tokens, but how exactly? There's no open-source implementations that I know of, and I'm struggling to find details.
One thing i'd add is that generating the tokens at the target resolution from the start is no longer the only approach to autoregressive image generation.
Rather than predicting each patch at the target resolution right away, it starts with the image (as patches) at a very small resolution and increasingly scales up.
Paper here - https://arxiv.org/abs/2404.02905
EU has almost no AI, yet is the world leader in AI .... regulation. No upstart social network would have the funds to ensure GDPR compliance and all the other rules that get added along the way. If someone has a business idea for one, they move to the US cause it's easier to start. Mistral AI is expanding to California. The solution would be less bureaucracy and regulation.
DeepMind was EU based, for example, before Google bought them. When I worked three, they were still London based (and in EU). Are they American now? American capital controls most of the AI, but that doesn't mean Europeans are incapable of AI research.
Also I think the push for AI alignment is American. At least three AI alignment people somehow always think that forcing I to adopt American values of the only correct and moral thing to do. Isn't that similar?
A human is not a blank slate. There's millennia of evolutionary history that goes into making a brain adapted and capable of learning from its environment.
