OpenAI切断模型训练时联网

RT @TrumpweTrust47: Here is George Clooney’s response after Trump accused him of being a "Hollywood elite." "Here’s the thing: I grew up i…

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications. On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments. While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software. After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely. https://t.co/EmEZsFRMcF

RT @mmpadellan: Republicans control the House. Republicans control the Senate. Republicans control the White House. Republicans control the…

RT @spooked75: I'm an American, a U.S. Army veteran, and I live in Europe. And I'll tell you something that might upset a few people back…

RT @TVietor08: Kind of a perfect metaphor for Hegseth's entire tenure at DOD. Stupid, impulsive, driven by vanity and the result is destruc…

During this red-teaming exercise, we also found a vulnerability that allows for unauthenticated remote code execution in Brotli, a widely-deployed compression library included in many pieces of software such as Node.js, Bun, and Envoy. The flaw could let an attacker run their own code on affected systems by sending or replying with malicious compressed data, without any need to create a user account. That is, an unpatched server that accepts compressed input could be taken over by a malicious client, and an unpatched client that accepts compressed input (such as “curl --compressed”) could be taken over by a malicious server. This root cause issue was actually fixed in the Brotli source code in February, but it wasn’t known at the time that it gave rise to a security vulnerability and there was no more recent release of Brotli. Since Brotli is built into other software, any software which includes this library needs to take the update too, followed by the users of that software. We disclosed the vulnerability to Brotli and other downstream projects such as Node.js, which released or are currently releasing security updates, and Brotli cut new release versions and published a CVE. More details on the patch and mitigations below: https://t.co/Y1Y72aHxrW

RT @jon_barron: Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-ent…

RT @ChatGPT: We’re using dots to build better dots. @rennsong, an engineer at OpenAI, uses her dot to gather feedback from social media an…

RT @MotherJones: The world’s first trillionaire is also a disinformation super-spreader. https://t.co/cLkCsaz9bH

RT @ylecun: @TPurity_Alphas Elon says: "we will not be able to boss around AI that are smarter than us".... while having built his fortun…

RT @MikeLevin: We’ve never seen corruption like this. Under Trump, a whole pardon industry has grown up around the White House, with weal…

Stay safe online. Here is what I do: 1. I use a password manager (Passwords, which comes preinstalled on Apple's ecosystem) 2. I have 2FA enabled everywhere. 3. If the site supports hardware keys, that's what I use. I have 3 YubiKeys 5C. I keep them in different locations. These keys are a bit more inconvenient, but they make it 1000x harder for anyone in the world to access your stuff. 4. If there is no hardware key support, I enable Passkeys and store them on my computer or phone. 5. I do not send authentication codes via text messages unless there's no other alternative to enable 2FA. SMS codes are vulnerable to SIM swapping and interception. 6. My iPhone uses an alphanumeric password. This is much better than a simple 6-digit PIN. 7. I also use a long, complex password on my laptop. It's inconvenient, but if someone steals it, they'll have a hard time breaking in. 8. I enabled Number Lock and SIM Protection with my phone carrier so hackers can't swap my SIM or steal my phone number. 9. I enabled Stolen Device Protection on my iPhone. 10. I have full disk encryption on every device that supports it (starting with my laptop and phone). 11. I have automatic updates enabled on my computers and my phone. I never skip an update, and also update the browser every time I get the notification. 12. I check every critical account every 3 months and make sure security is updated. Critical accounts are your bank, email, social media, and investment accounts. I make sure I don't have old emails, 2FA using SMS, etc.

RT @nxthompson: “The U.S. has benefited from the skills that Indians bring without having had to invest much in their education. Many have…

RT @ylecun: @elonmusk Hitler was as much of a far-left socialist as you are a democracy-loving humanist.

Genuine question: what's up with LeetCode, HackerRank, and all of those services? Are people still paying them to learn how to sort a B-Tree for a coding interview?

RT @judegomila: I'd like to share a little about what I've been working on at my new company @ultrametricai with my cofounder @tylerlastovi…

RT @EeroSimoncelli: Looking for a PhD program where you can study computational neuroscience? NYU has a fantastic array of researchers cov…

RT @mosh_levy: 1/ JEPA selects which features its world model keeps. But how? Will it keep a feature it can't predict, like a traffic light…

Let's say your app is calling Opus 5.5. Go to your code, change the URL you are calling, and replace it with Mubit's URL. Here is what happens internally: • Mubit will learn about the tasks you are solving. • It will use your requests to train a small model. • This small model will take over those calls. • Opus 5.5 will still handle most of the work. This is a pretty cool idea, and if it works well, it should give you faster responses at a much lower cost.

Jev is really useful, actually. I have a script that receives a notification every time I get a new email. I use Jev to decide whether I care about that email or not (some sort of spam filtering). This is "personalized spam" detection. A layer on top of my email provider's spam filter. My goal is to catch emails that aren't necessarily spam but that I still don't care about. I classify new emails with Jev to decide whether they're spam. If Jev's confidence is above 90%, I delete the email. I've been running this for about 3-4 weeks (pretty much since Jev came out), and so far, I've deleted hundreds of messages. During the first week or so, I manually checked most of the emails the script discarded to make sure I wasn't missing anything important. I wasn't. The script is good.

This new open-weight model is pretty interesting: EdgeRock-4B works with text and images, and it's designed to handle long reasoning tasks, but there's a twist: The model keeps its history at a fixed size, making memory usage more predictable as reasoning continues. As conversations get longer, regular models use more memory on your hardware. This makes it hard for people building AI applications to predict how much memory they will need. EdgeRock-4B does something different: • It keeps recent information available • It maintains a compressed summary • It learns which information to keep around Basically, the model tries to keep the most recent details together with the most important information in context. This is good when you want predictable memory. Basically, you can handle more requests on the same hardware. But there are trade-offs. 1. The model decides what to remember, so it can lose important information. 2. Accuracy will vary across different tasks. This model will work well for some problems, but not for others. 3. You'll need compatible software to deploy the model because of its specialized memory architecture. You can see the model's architecture attached to this post.

Your agent sucks at searching the web. I don't care what prompt you are using; you are just getting mediocre results. Here is how you can fix it. https://t.co/qi50rcyjLC

People told me I wasn't writing code anymore as an insult. I was spending most of my time in meetings. My focus shifted from writing code to architecture and client management. I went from spending 100% of my time in a code editor to spending 80% of my time on a whiteboard. Some people laughed and mocked me for it. Not writing code anymore clearly felt like it was going to hurt me. But who's laughing now, huh?

Something else that keeps me safe: I use Privacy for all of my credit cards. Not sponsored, and I've been using them for years now. Basically, instead of putting my credit cards out there, I create single-use cards with Privacy. As soon as I pay, the card is deactivated. I also set up merchant-tied cards for recurring payments. So, for example, I pay YouTube with a card that only accepts charges from YouTube.

RT @KenRoth: Thanks to Trump, the United States sticks out as the only major economy where planet-warming pollution increased dramatically…

This is pretty smart: MemT 1.0 is a new open model that can use information from earlier interactions in later tasks. For example, it can use the feedback you gave it yesterday on a similar task you are completing today. The model uses a Mixture-of-Storage architecture (MoS) with multiple memory experts. Each expert maintains its own memory, and two routers control how it writes and retrieves information. This is different from a traditional Mixture-of-Experts architecture. Every single memory expert in MemT's architecture participates every time, but with different weights. The bottom line is that the model accumulates information from previous interactions and reuses it later without accessing the original conversation.

Nothing is safe from AI — and this is a good thing, I think. This is a new platform for freelancers and people doing creative work, and at the core of it, there's an agent doing a ton of work we usually have to do manually: 1. It drafts proposals for clients 2. It generates tasks from a project description 3. It schedules those tasks around commitments 4. It keeps clients updated on progress 5. It prepares invoices 6. It prepares time logs based on tasks and activity