搜索“RAG”

t. Alex Atallah [00:29:09]: And it’s like building something that people want and going and telling them about it. Like, you can do that on a one basis, but there’s way higher leverage to do that in a community where everyone can talk to you at the same time. So we spent a lot of time, like, building things that the community really wanted. We did the same thing for OpenRouter. And, like, the Axie community was one of, like, a zillion communities we did that with. And Anj, like, saw us doing it

匹配字段正文

didn’t just attack Australia’s government servers. And Australia apparently wasn’t the only country attacked by OpenAI’s software. Worse, in disclosing what happened, OpenAI has dragged their feet every along the way. § About the closest thing to an honest account of what happened – and I seriously doubt even this is completely candid – came out about an hour ago. Even this was still a mix of euphemism (“attacks” were called “interactions”) and partial disclosure. “Most cases identified” were l

匹配字段正文

f the SaaS-pocalypse developed this week as whole categories of companies — dubbed the “consumer inertia” sector — saw their shares tank on the perceived agent threat. Banks, brokerages, and mobile phone carriers were down on the prospect of consumers switching effortlessly for better terms. Third-party services like OpenTable and DoorDash, which could be bypassed, got hit. So did subscription-based media businesses that could be vulnerable to agents rationalizing purchases; New York Times Co. s

匹配字段正文

, vague language to make it seem like it’s been far more productive. Nobody on God’s green Earth is sitting here wondering if Microsoft added gigawatts of CPU capacity or cloud storage, nor are analysts desperate to hear about aggregate numbers — they’re asking where is all the money you’re spending on AI going, and the answer, it appears, is “into a warehouse” or “into a data center without power.” To bring home the point, I analyzed Microsoft’s earnings calls between Fiscal Year 2010 and Fisca

匹配字段正文

己的东西做那件事，而且它越可预测，他们就越不需要测试和摆弄它。Jeb 还远没到那个程度。它可以做到，但就像，我们想要多得多的可靠性九数，才能让它好到像数据库查询一样，你甚至都不用去想它。它就在那里，当你需要智能的时候。但安全对齐，就像，是指令遵循的反面。它是当你想要遵循别人的指令时，比如 OpenAI 和 Anthropic Swyx [00:14:13]：RAG 的价值栈。 Diogo Almeida [00:14:14]：正是。而这对一个产品来说非常合理。再说一次，就像，ChatGPT 应该这么做。你——你应——就像，如果他们不想，比如，跟 ChatGPT 搞什么不适合工作场合的角色扮演，那是他们的事，因为，就像，也许那正是他们那些有父母和孩子的用户想要的。就像，那没问题。但在 API 里，那太疯狂了，对吧？就像，那完全不可接受，因为，就像，人们需要，就像，围绕这个来编程，而那太反用户了，以至于它。它。我。哈。我可能是个容易生气的人，所以我应该试着冷静下来。 Swyx [00:14:52]：这，人们能感受到你的热情，我觉得这非常好。我要给你的一个反驳是，就像，如果我们用它来杀人呢，

匹配字段正文

odels to all eighteen adds another 6.4 percentage points. The useful object is not a catalog of hundreds of nearly interchangeable models. It is a portfolio with complementary coverage. The value is capability coverage, not model count. LLMRouterBench evaluates routing across 33 models and more than 400,000 instances. It finds that a handful of models covers most of what the full set can do, and that bigger pools add little without careful curation. The hard part is predicting which model to use

匹配字段正文

odels to all eighteen adds another 6.4 percentage points. The useful object is not a catalog of hundreds of nearly interchangeable models. It is a portfolio with complementary coverage. The value is capability coverage, not model count. LLMRouterBench evaluates routing across 33 models and more than 400,000 instances. It finds that a handful of models covers most of what the full set can do, and that bigger pools add little without careful curation. The hard part is predicting which model to use

匹配字段正文

n examples of that: I had GPT-6 Astra turn a 1977 text adventure game called Zork into a full 3D action-adventure game you can play . Zork has no graphics and each location is a paragraph of prose, so the AI had to decide what the white house looks like, what a grue looks like (the original only tells you that you are likely to be eaten by one in the dark), and how to turn “fight the troll” into an action sequence. I also had Fable 5.1 try to reconstruct Italian author Umberto Eco’s library in 3

匹配字段正文

ng an IMO gold in 2025… The idea that it could be done with a general-purpose language model with no tools and no access to the internet, even people at OpenAI thought this was outrageous. They thought it was almost impossible. Then you get to 2026. Literally two weeks before we got Navier-Stokes, I was talking with a researcher at a frontier lab about how long it would take to get a Millennium Prize, and he was willing to bet me $1,000 that it would take past 2027. He thought it would take unti

匹配字段正文

sks , but may not materially improve medium/low-complexity coding. Notably, access increased total coding spend by 60% , so Databricks created a dedicated Astra sub-budget to encourage selective use. Benchmarks are converging on a similar picture : @EpochAIResearch said Astra now leads their overall Epoch Capabilities Index , with a new Math-ECI record, while Claude Fable 5.1 remains strongest on software engineering. @arena showed Astra and Fable as top-tier but expensive, with Astra Max at +$1

匹配字段正文

, but the strain of years of rapid datacenter growth. Much of this can be traced to the electricity-bill shock of mid-2025, when datacenter-driven PJM capacity costs pushed the average Ohio residential bill up 25.7% year over year, and to hyperscale proposals spreading into townships with little or no zoning. In Sunbury, the city council enacted a moratorium on April 15, 2026, effective through January 2027, for the city to review zoning, building, and business regulations. Moratorium indicators

匹配字段正文

potentially viral stories for Instagram, YouTube, and TikTok without recycling topics we’ve covered. First, the workflow reviewed Rowan’s public short-form archive to create a coverage map of past topics and editorial taste. It uses that map as a filter to distinguish a new angle from another version of the same AI launch, robot demo, or research story. Each potential story is checked for four things: a clear visual hook, a technical idea that can be explained in 45–90 seconds, broad audience in

匹配字段正文

rkloads that have very high context lengths that place pressure on HBM capacity but also have long pauses between turns, making it worthwhile to expel KVCache into DRAM or even storage rather than recomputing it altogether. In our recent AgentX benchmarking, many workloads’ pareto optimal results were obtained with KVCache offloading above certain concurrencies. AgentX - InferenceXv3: Does CUDA Moat Hold up in Agentic Inferencing? To test the impact of reduced HBM in a real workload, we ran Infe

匹配字段正文

of something bad happening seems to be between 10-25%. And well, Dario then wrote about pacing the frontier . And Sam read it and wants to pace too. And well, so does Musk. I encourage you strongly to read the post, because I think it’s a good one. And yet, when I read the post I could not help but feel in strong opposition to it, despite the fact that I think I’m on the same page with regard to all observations and, to a large degree, the concerns. I thought it might be interesting to write dow

匹配字段正文

ened by the vibe shift in AI safety. Last month, in the wake of Mark Zuckerberg’s glad-handing treatise on “personal superintelligence,” I argued that true superintelligence is a dragon. And while I remain hopeful that it can be tamed, I believe the researchers who say we are nowhere close to being sure of it — and are quitting, in protest, jobs that would make them rich. The right people are now listening to them, and here’s hoping they take action. Elsewhere in AI safety: I can't decide whethe

匹配字段正文

ot be a co-author. Buckmaster was furious and went public with the story. Predictably, the drama overshadowed the mathematics. People debated whether Buckmaster was right to be outraged — or whether Bubeck’s response exonerated OpenAI. But I think that focusing on the details of the drama risks missing the larger point. Mathematics is as much about cultivating a community of experts as it is about solving individual problems. Some of OpenAI’s behavior might have been reasonable in the context of

匹配字段正文

ld. According to its prospectus, Unitree spent 145 million yuan on R&D in 2025, or 8.53% of revenue — a middling level among tech companies, given that STAR Market-listed firms average roughly 11%. UBTech spent about 500 million yuan on R&D in 2025, about 25.4% of revenue. Unitree’s R&D-to-revenue ratio was 31.39% in 2023 and 17.83% in 2024. The decline is mainly because revenue grew explosively over the same period, diluting R&D as a share of the total. … Unitree’s offices in the Zhongguo Shugu

匹配字段正文

and: What we do have is a serious housing shortage and housing affordability crisis. Despite robust construction, unsold inventory stands at four months, well below its 25-year average. Private builders complain they can't get land permitted to meet demand. Low-income housing advocates complain housing prices are out of reach for many Americans, and that government subsidies have been slashed. Hey, this kind of reminds me of something that NVIDIA CFO Colette Kress said on its latest earnings cal

匹配字段正文

s performance on ARC-AGI-3 is quite interesting. Dropped into an abstract game it had never seen, it used fewer actions than the human median on 96% of the levels it completed, averaging 51.7% fewer actions per level. This goes against researchers’ original expectation that even when an AI solves an environment, it might fumble around and be less efficient than humans. But Astra invented a symbolic model to hold an entire environment in a compact notation system. In a way, it replaced trial-and-

匹配字段正文

Our guide to the up-and-coming names in AI shopping and payments infrastructure Venture capitalists are betting that a barrage of AI agents will soon arrive and upend the marketplaces for e-commerce and payments. Most of the action today is in building the infrastructure and tools for the day when agents can be trusted with a customer’s credit card and shopping choices. How close we are to that day is one of several very big questions facing the nascent AI commerce industry. Our upcoming Machine

匹配字段正文