ARCHIVE
正文、图片和 Markdown 均已本地保存，不包含回复区。
Beautiful work from @perplexity_ai @AravSrinivas with weights on @huggingface of course! huggingface.co/perplexity-ai/… huggingface.co/perplexity-ai/… Perplexity Developers @perplexitydevs Our next generation of open weights multimodal decision models comes in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2. Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass. Flash is the top 4B on the Decision Index. 🔗 View Quoted Tweet
Surprised that so few people are talking about this. Persistence has become a core capability of recent models (Opus 5.5, GPT 6 Astra, Gemini 4 Argon,...). I was just reading through the Gemini 4 Argon Minecraft paper details and noticed that no roles were assigned to the agents. Somehow the four agents completed the goal after 27 hours, 2,783 programs, and 605 deaths. They recovered items, revised strategies, and shared lessons. I think this is a good example of persistence helping them eventually succeed. Some more thoughts for discussion: So this is not just about hooking an agent up to an external memory system. It's also not about loop-prompting it with /goal, although that is one of the most effective techniques right now. This is about the model's native ability to choose what to persist, and how and when to use it. It unlocks long-horizon capabilities, as agents often execute more complex tasks more effectively through persistence. So it's less that the models got more intelligent and more that they got better at coordinating and planning for more complex tasks. We are in the early days of this. But this is an important capability to optimize to really unlock AI for scientific discovery. This is also important for building more personalized agents with identities and roles that can track and follow human preferences well and learn continually over time. We also don't entirely know how to quantify and measure it as a field. Reward hacking combined with persistence can also enable weird, unsafe agent behaviors. Just a few quick thoughts, and I hope we can start deeper conversations about this. I am tracking this space closely and will be sharing research on the topic as it surfaces.
🧭In this week's letter from Andrew Ng: AI improves fastest on tasks you can verify. Judgment tasks are harder, and it's easy to overgeneralize progress in math and code to the rest of human activity. Also inside: 🤖Claude Haiku 5.5 and Sonnet 5.5 🇪🇺Mistral Large 4 💻Computers for local AI 📊Causal deeplearning.ai/the-batch/issu… qo #DeepLearningAI r #AIEngineering n #MachineLearning earning
GPT-6 Astra Pro in ChatGPT got an upgrade with faster responses and improved use of plugins and tools. And now, multiple agents can tackle different parts of a task in parallel. You can follow each agent’s progress and open completed tasks to review their work. Now available to Pro users in the Chat tab.
GPT-6 Pro in ChatGPT got an upgrade with faster responses and improved use of plugins and tools. And now, multiple agents can tackle different parts of a task in parallel. You can follow each agent’s progress and open completed tasks to review their work. Now available to Pro users in the Chat tab.
Span-02 from @RespanAI is now on OpenRouter An open-weight decision model: give it text and images, define your questions, and get back a choice, a yes/no/unknown, or a score, with probabilities. $0.01/M input tokens, free output, 131K context. Respan @RespanAI Span-02 is the cheapest decision model: $0.01/M input tokens. With thinking off, it delivers frontier-level decision-making at lowest latency. Need more? Give it a reasoning budget. Span-02 thinks until the deadline, then returns its best decision so far. Stop reasoning anytime. Or let Span-02 decide. In auto mode, an adaptive learned gate picks the reasoning budget for every decision. Fully open weights with a rust inference engine. 7× faster than vLLM. 🔗 View Quoted Tweet
This Week in the Arena: - Gemini 4 Argon (High) by @GoogleDeepMind landed the #5 spot in Agent Arena, with +13.2% net improvement. - MiMo V2.6 Pro by @XiaomiMiMo now ranks #1 among open models in Agent Arena ( #13 overall), with +3.9% net improvement. - Muse Wonder (Max) by @AIatMeta debuted at #2 in Text Arena (1513 pts), #15 in Code Arena: WebDev (1651 pts), and #12 in Agent Arena with +6.4% net improvement. - Muse Spark Open is live in the Agent Arena, with scores coming soon. Stay tuned! - We introduced our new agentic version of Max: Arena’s model router that achieves frontier-level performance at 37% lower cost. More info in thread. See leaderboard results and Arena product updates from this last week in 60 seconds. youtube.com/watch?v=yDSaPJ…
x.com/arena/status/2… Arena.ai @arena Introducing our new agentic version of Max: Arena’s model router that achieves frontier-level performance at 37% lower cost. Built on 2.5M real-world agentic sessions from Agent Arena, Max learns to dynamically select the best model for each task, optimizing for both performance and cost. Max achieves a +14.7% net improvement score in Agent Arena at a median cost of $3.27 per task, matching the performance of Claude Fable 5.1 (Max) at 37% lower cost. Try Max today for a limited time, and dive into more technical details in the thread. 🔗 View Quoted Tweet
New content workflow. Step 1: Paste your script into ChatGPT Step 2: HeyGen MCP makes your A-roll with the most realistic avatar and voice you’ve ever heard Step 3: @HyperFrames_ Studio edits it, captions it, and color matches it Scaling your channels just got a lot easier. 🔗 View on Twitter
We just open-sourced another pair of state-of-the-art Decision Models: pplx-decider-flash-v1.2: 4B params, 1 cent per million tokens (top 4B model on HF Decision Index) pplx-decider-v1.2: 27B params, 2 cents per million tokens ( #2 on HF Decision Index overall - for now, will improve rapidly) Both are available on the Perplexity Decisions API as well as in the form of open weights on @huggingface Enjoy! Perplexity Developers @perplexitydevs Our next generation of open weights multimodal decision models comes in two sizes, pplx-decider-v1.2 and pplx-decider-flash-v1.2. Text and vision scores improve. Inference is up to 5x faster: all questions in a request run in one pass. Flash is the top 4B on the Decision Index. 🔗 View Quoted Tweet
Will O'Brien of @UlyssesInc on why today's young founders are more Steve Jobs than Jeff Bezos: "The founders look very different, right? It's like the kind of rah-rah, smoking, go to church on Sunday, maybe a bit louder, working typically on hardware rather than software." "It's like a different motivation. There is seemingly more of a kind of a spiritual reason to why people start their businesses." "I often think of there being head and heart type founders, where Jeff Bezos is the canonical head-based founder... 'What would be the best thing to sell on the internet? Oh, books are good because you can move them easily.'" "Steve Jobs being the canonical heart-based founder, the bicycle for the mind, the hippie on acid... And I think this era of founders are way more oriented on the heart-based." "A lot of us were very, I think, impacted by something like COVID. A lot of us are typically in our later 20s now, but were in college or something like that when it happened. It kind of made me feel a little bit cheated." "We were forced to be in the digital world, so we feel cause to impact the physical world... Or maybe we're just more of a product of the times that we're in." @Willob on Cheeky Pint 🔗 View on Twitter
It’s time to 🇺🇸-max. Beam is an American-made open-weight model by Reflection. And it’s now free on @vercel AI Gateway. Plug it into deployable @eve agents, start coding with @fx , or build your own harness. Freedom. Never bet against America. Vercel Developers @vercel_dev Beam from @reflection_ai is free on AI Gateway for a limited time. Until today it ran in stealth as 𝚜𝚝𝚎𝚊𝚕𝚝𝚑/𝚌𝚊𝚝-𝚓𝚊𝚖. It's a low-latency US open-weight model. 𝚛𝚎𝚏𝚕� vercel.com/changelog/refl… 𝚛𝚎𝚎 https://t.co/aKZZNLsZVv 🔗 View on Twitter 🔗 View Quoted Tweet