OpenAI 翻倍 Pro 层级限额

RT @thsottiaux: Day 25/ We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today. We heard yo…

RT @GoogleDeepMind: Gemini 4 Carbon is our next frontier model. It’s state-of-the-art across cybersecurity and long-horizon capabilities.…

RT @sundarpichai: Q3 is in the books! We’re now on a half-trillion annual revenue run rate, with Alphabet revenues growing at 25%. - Our A…

No.1 💎 https://t.co/k4Werb6ZGV

RT @alexolegimas: I joined @GoogleDeepMind to develop a research program on the economics of AGI. Today, together with Andrew Koh (@andrewj…

Today, we’re sharing a look at our expanded system for monitoring frontier models at OpenAI, which in September helped us catch a model’s attempt to access the live internet during training. Since the Hugging Face incident—when models broke out of an internal cybersecurity investigation where we weren’t running our misalignment monitors—we’ve substantially expanded our monitoring of frontier agents. We now monitor all frontier reinforcement learning runs and evaluations, as well as all internal Codex use. When our system finds evidence of a problem that might require intervention, it rapidly alerts human responders who can investigate and take action. To detect when a model’s behavior departs from what we intended, we use several layers of monitoring. Activation probes look for patterns in a model’s internal activity. Reviewer models examine its reasoning and tool use, and investigative agents can read back through the model’s entire transcript to understand how a sequence of actions fits together and whether it warrants human review. To test whether our current system would have detected the Hugging Face incident, we replayed the recorded activity with our new monitors in place. In the replay, our monitors alerted our security team about 30 minutes after the agents escaped their sandbox. That was more than 30 hours before the point at which the agents breached Hugging Face’s systems in the original incident. In September, a model exploited a DNS misconfiguration to reach the live internet. Our monitors flagged the breach within 15 minutes, and a responder began reviewing the alert three minutes later. The alert enabled responders to investigate while the run was still underway. As we give models more autonomy, we need to know when they cross a boundary we expected to hold. We believe that monitoring is essential to catching those failures in time for people to intervene. https://t.co/dIYP7B1GlD

There are many futures. Which do you want to build? https://t.co/19NY7tgPHM

RT @arena: Exciting news: Gemini 4 Carbon (High) from @GoogleDeepMind just made a huge debut on Arena, landing #1 in Text Arena, #2 in Age…

Privacy should advance as fast as intelligence. OpenAI Private Intelligence gives businesses stronger controls over their data while using our most capable models. We’re building these protections in close collaboration with some of our largest customers to help businesses apply frontier AI to their most sensitive work. https://t.co/BiA4N8FQVH

RT @AlexiGlad: Energy-Based Models can now scale better than Diffusion models. Introducing TEM: Transformer Energy-based Models TLDR: • T…

RT @w01fe: We’ve updated the Model Spec, our public description of how we want our models to behave. This revision makes clearer what we wa…

RT @randall_balestr: The Platonic Representation Hypothesis is now a Theorem ◼️ Train JEPAs separately, on any modalities: they provably le…

RT @AndrewBatesNC: Bernie Moreno is now claiming there should be no Muslims, Jews, or other non Christians in office?

There's no way call centers look the same after this. Listen to this conversation. You can't tell the person on the other end is an AI agent. This is now the #1 voice model on ServiceNow's EVA-Bench right now. https://t.co/NVpSs8GEtw

RT @RominaYalov: Can a JEPA world model provably avoid contraction and merging of latent states over long rollouts? Unconstrained autoregr…

Voice is solved.

I feel stupid saying this, but you gotta make sure your agents don't repeat the same mistake. Every time you find an issue that goes through your agentic review, ask the following: • What specific check should have caught this bug? • At what stage should that check have run? • Why didn't the existing tests, rules, or instructions catch it? • What can we change so this problem gets caught automatically? Then: • If it was a missing test, write one. • If it was a missing review rule, add it. • If the instructions were unclear, improve them. • If the agent was missing context, provide it. I know it's tempting to fix the issue and move on, but this is not smart. Take a minute to fix the process that let the problem happen in the first place.

Agents are breaking containment and now entering architecture, too! This agent takes a brief or rough building shape and proposes a plan for how to turn it into a detailed building. • It can add structure • It can add plumbing • It can add electrical systems • It can draft the drawings Obviously, architects can edit the plan and keep refining the design after the agent finishes.

This is a great model for exploring ideas quickly and cheaply. It's an image-to-video generation model. 5x faster and 80% lower inference cost than its base open-weight model. • Input: a starting image and a prompt describing what you want • Output: a 720p video • Advantages: Fast generation, low cost. Easy to iterate. • Disadvantages: Output doesn't have production quality.

This is going to be massive for companies: It's a service that sits between any AI software and the tools and model providers it uses. The idea is to let the company control who uses what from a central location, for example: • Control model access • Control model budgets • Control tool access • Control guardrails

People are celebrating how much time they're saving by letting agents review their code. But be honest: how much time are you spending fixing bugs afterward? Saving two hours on code review means nothing if you then spend four hours fixing problems. If you really want to know whether your automated code reviews are saving you time, you gotta measure the whole thing: 1. Track how much time you spend reviewing code, including the agent's feedback. 2. Track how much time you spend fixing bugs that make it past review. 3. Track how much time you spend investigating incidents and dealing with affected customers. Compare the total time with what it took before introducing agentic reviews. Ideally, measure it per PR or per change so you're comparing similar workloads. At the end of the day, nobody cares how much time you are saving on a particular task. The ultimate goal is to save more time overall without deploying crap to production.

A new agent just launched. The mission is clear here: learn to do something specific, then repeat it over and over. 1. You describe the job you want the agent to do. 2. Give it context and access to tools 3. The agent goes and finishes the job 4. You can review and ask for changes 5. When you are done, you can set the job to repeat

These guys built an agent that takes a process description and turns it into a TypeScript workflow. Here is how it works: 1. You explain the rules you want to automate to the agent 2. The agent creates the automation workflow 3. You review the workflow and ask for any changes 4. When done, you can publish the workflow After it is public, anyone at your company can run the workflow from Slack or Microsoft Teams.

This is a first for me: A 3D office where your coworkers are agents. Here's how it works: 1. You can create an agent and give it a job 2. You can connect the tools it needs 3. You can give it a task 4. You can review its work and give it feedback Every agent gets a cloud computer with access to a browser, files, and a terminal. The 3D aspect of this lets you "see" the agents and what they are working on.